hyperresearch
Docs Create an account

Sample report

The AI Coding-Agent Market, October 2026: Who Is Winning, What It Costs, and What a 50-Engineer Company Should Buy for 2027

How this page was made

An unedited Hyperresearch Deep report. The run finished on 4 October 2026; the text below is what it produced, with its citations and sources, and nobody has edited it. Facts in it are as of that date.

Question asked
Who is winning the AI coding-agent market as of October 2026? Compare Cursor (Anysphere), Claude Code (Anthropic), OpenAI Codex, GitHub Copilot (Microsoft), and the strongest open-source agents (Cline, Aider, OpenHands). For each, establish: - revenue or ARR, and paid seats or active users; - list pricing, and what a heavy daily user actually pays per month once usage limits and overages apply; - adoption signals: GitHub stars and contributor trends, npm/PyPI download trends, named enterprise customers, and the Copilot figures Microsoft has disclosed in filings and earnings calls. Label every figure as company-stated, filing-disclosed, or third-party estimate, and give the date it refers to. Where sources disagree, show the range and say which figure you trust and why. Then answer the decision: a 50-engineer software company must pick one primary coding agent for 2027. Which should it choose, under what conditions would the answer change, and what are the lock-in, cost-predictability and code/data-governance risks of each option?
Tier
Deep
Finished
Sources
259
Words
8,512

Researched by Hyperresearch

Download PDF (1.9 MB)Ask your own question, free
Contents
  1. Summary verdict
  2. How figures are labelled and which sources are trusted
  3. Market scoreboard
  4. Claude Code (Anthropic)
  5. OpenAI Codex
  6. GitHub Copilot (Microsoft)
  7. Cursor (Anysphere, now SpaceX)
  8. Open-source agents: Cline, Aider, OpenHands
  9. Cross-checking the adoption signals
  10. Normalised heavy-user cost
  11. What the productivity evidence supports
  12. Decision: one primary coding agent for a 50-engineer company in 2027
  13. Further reading
  14. All sources

Summary verdict

Claude Code (Anthropic) is winning on the measures that best predict 2027 position. It leads workplace usage (39% of professional developers in May–July 2026, up from 18% in January), developer satisfaction (91% CSAT, NPS 54) and enterprise coding spend. Its last company-stated revenue figure is a run-rate above $2.5 billion (12 February 2026). Third-party trackers put it at roughly $8 billion in May and about $15 billion in August, and those later figures are estimates [41].

The other three commercial players each lead on one axis:

  • OpenAI Codex has the steepest growth. It went from 3% to 16% workplace adoption, reached 5M+ weekly users in June, and since late August has the most npm downloads of any coding CLI. Its disclosed revenue is just over $1 billion [8][60][179][160][39].
  • GitHub Copilot has the largest disclosed paid base: 4.7 million paid subscribers in January 2026, stated on Microsoft's earnings call. It is also the strongest large-enterprise default. But it fell from 29% to 21% workplace adoption, and its June 2026 move to metered AI Credits caused billing backlash [8][68].
  • Cursor is the largest disclosed pure-play coding revenue line: about $4 billion annualized in early June 2026, per press reports. Its developer share is falling (18% to 12%). It is now owned by SpaceX, and it loses OpenAI models on 12 November 2026 [85][8].

Among open-source agents, OpenHands and Cline are actively growing. Aider has effectively stalled: its last release was February 2026.

Recommendation for a 50-engineer company: standardise on Claude Code via the Claude Team plan. Use Premium seats for the engineers who run agents all day and Standard seats for everyone else. Add an API or cloud-provider path for CI, and keep agent-neutral conventions (AGENTS.md, MCP servers) so the company can leave.

The answer changes in four cases:

  • GitHub Copilot Business/Enterprise if the company is GitHub-Enterprise-standardised and procurement requires IP indemnity, one vendor, or FedRAMP-type model restrictions.
  • OpenHands or Cline with your own cloud inference if code must not leave your environment.
  • Codex if cost per token dominates or the company is already on ChatGPT Business.
  • Cursor only if an editor-first workflow outweighs its ownership and model-supply risks.

Sources for this section:

How figures are labelled and which sources are trusted

Every figure below carries one of three labels:

  • company-stated: the vendor's own blog, pricing page, documentation or earnings call;
  • filing-disclosed: SEC filings;
  • third-party estimate: press reports citing unnamed sources, trackers, analysts, aggregators.

Each figure is also given the date it refers to.

Four rules resolve conflicts:

  1. A dated company or filing figure beats a later estimate as a floor.
  2. Defined metrics beat undefined ones. "5M weekly users" beats "20M active users" with no time window.
  3. Sources that conflate products are discarded. Several aggregators merge Microsoft 365 Copilot seats with GitHub Copilot subscribers.
  4. Model-based revenue trackers are tested against the one independent market-size estimate available. Gartner estimated the whole enterprise AI coding-agent market at $9.8–11.0 billion annualized as of April 2026 [93].

That fourth test matters. TickerTrends puts Claude Code at about $10.3 billion by the end of April 2026 and Codex at about $3.0 billion in April 2026 [39]. Add Cursor's ~$3 billion (late April, press) [85] and Copilot's ~$0.5–1.1 billion, and the total is $17–18 billion. That is well above Gartner's whole-market figure. Either the trackers count general subscription revenue (ChatGPT or Claude plans) as coding revenue, or Gartner's scope is narrower. Either way, tracked ARR for Claude Code and Codex should be read as an upper bound, not a measurement.

Some secondary sources contain figures that contradict primary data and are not used:

  • One blog claims JetBrains showed Claude Code at 54% and Copilot at 41% [11]. JetBrains' own posts say 39% and 21% [8].
  • One comparison credits "GitHub Copilot" with "15 million paid subscribers" [9]. That is Microsoft 365 Copilot's seat count [144].
  • One article dates Cursor's $29.3 billion Series D to February 2026 [9]. Cursor's own post dates it to November 2025 [85].

Sources for this section:

Market scoreboard

Metric (date)Claude CodeOpenAI CodexGitHub CopilotCursorClineAiderOpenHands
Latest company/filing revenue figure>$2.5B run-rate (12 Feb 2026), company-stated [23]Not disclosed; OpenAI total $40B+ ARR (Aug 2026, Bloomberg) [165]Not disclosed by Microsoft>$1B ARR (13 Nov 2025), company-stated [85]Not disclosedn/a (no company)Not disclosed
Latest revenue estimate~$8B (May 2026, Sacra/FutureSearch); $15.12B tracked (10 Aug 2026, TickerTrends) [40][39]$1B+ (May 2026) [160]; $8.83B tracked (10 Aug 2026) [39]≥$550M (FT, May 2026); $0.9–1.1B (Axis est., Jan 2026) [116]~$4B annualized (early June 2026, Forbes via Dealroom) [90][85]———
Users / seatsNot disclosed; WAU "doubled" Jan 1–Feb 12, 2026 [23]5M+ weekly (2 Jun 2026), company-stated [60]4.7M paid subscribers (Jan 2026 call), company-stated [144]No MAU; "millions"; >1M DAU (2025, Bloomberg) [85]~5.29M VS Code installs (Sep 2026) [56]10.78M all-time PyPI downloads [14]641 contributors to org (Oct 2026) [120]
Workplace adoption (JetBrains, May–Jul 2026)39%16%21%12%———
GitHub stars (Oct 2026)149.1k* [124]127.7k [124]—33.3k (issue tracker only) [124]69.8k [49]49.3k [124]89.9k [171]
Ownership / fundingAnthropicOpenAIMicrosoftSpaceX ($60B, closed 14 Aug 2026) [85]$32M raised (Jul 2025) [47]Community$5M seed + $18.8M Series A [18]

\*The Claude Code repository is public, but the tool is under commercial terms, not an open-source licence [165].

Sources for this section:

Claude Code (Anthropic)

Revenue and users

Anthropic has published three Claude Code revenue milestones and none since February 2026 [41]:

DateClaude Code run-rateLabelSource
2 Sep 2025>$500M; usage up >10x in three monthsCompany-stated (Series F)[41]
3 Dec 2025$1B, six months after general availabilityCompany-stated[41]
12 Feb 2026>$2.5B, "more than doubled since the beginning of 2026"Company-stated (Series G)[23]
May 2026~$8BThird-party (Sacra/FutureSearch)[41][40]
Aug 2026"Near $15B"Third-party trackers cited by FutureSearch[41]

TickerTrends anchors its weekly tracker to the $2.5 billion February figure and models it forward [39]:

Week ending (2026)Claude Code tracked ARR
End March~$6.3B
End April~$10.3B
June>$14B
6 Jul$14.26B
13 Jul$14.38B
20 Jul$14.47B
27 Jul$14.35B
3 Aug$14.50B
10 Aug$15.12B (21.9% of Anthropic tracked ARR; +5.2% month on month)

Company-level context helps judge plausibility. Anthropic's total run-rate was $14 billion on 12 February 2026, of which Claude Code was more than $2.5 billion, about 18% [23]. On 28 May 2026 Anthropic said total run-rate exceeded $47 billion but gave no Claude Code split [41]. Preliminary Q2 2026 revenue of more than $11.5 billion was reported from documents seen by Bloomberg [25]. Bloomberg put the run-rate above $65 billion by end-July [165].

If Claude Code kept roughly its February share (about 18–22%) of a $47–65 billion total, it would be at $8.5–14 billion. That brackets the third-party estimates. But it collides with Gartner's whole-market estimate [93].

Trusted figure: $2.5 billion (February) as the verified floor. A mid-2026 range of $8–15 billion is plausible, but it rests on an untested assumption about how much of Anthropic's revenue is attributed to coding.

Other company-stated facts, all from 12 February 2026 [23]:

  • Weekly active users doubled since 1 January.
  • Business subscriptions quadrupled.
  • Enterprise use is more than half of Claude Code revenue.
  • A third-party analysis Anthropic cited estimated that 4% of public GitHub commits were authored by Claude Code, double the prior month.
  • Eight of the Fortune 10 are Claude customers (across all Claude products, not only Claude Code).

Anthropic has published no absolute user, seat or retention count for Claude Code [41]. Pages quoting specific weekly-user numbers do not trace to Anthropic.

Pricing and what a heavy user pays

PlanList price (company-stated)Claude Code usageSource
Pro$20/mo, or $17/mo billed annuallyYes; ≥5x Free per 5-hour session; weekly caps[181]
Max 5x / Max 20x$100 / $200 per month5x / 20x Pro per 5-hour session[36]
Team Standard$20/seat/mo annual ($25 monthly)"More usage than Pro"; teams of 2–150[181]
Team Premium$100/seat/mo annual ($125 monthly)5x Standard[181]
Enterprise$20/seat/mo annual + all usage at API ratesBilled on consumption[181]
API (Console key)Opus 5.5 $4/$20; Sonnet 5.5 $2/$10; Haiku 4.5 $1/$5; Fable 5.1 $10/$50 per M tokens—[182]

Usage limits reset on a rolling five-hour window, and paid plans add weekly caps. Claude chat and Claude Code draw on the same pool. Paid users can switch on usage credits at standard API rates once they hit a limit [181][227].

Several cost mechanics matter for heavy users:

  • Cache lifetime. On a subscription the conversation cache lasts one hour within included usage. It drops to five minutes once a user is on usage credits, an API key or a cloud provider, so a break triggers full-price reprocessing of the context [36].
  • Silent API billing. An ANTHROPIC_API_KEY environment variable takes precedence over the subscription login and silently moves billing to per-token rates [36].
  • Tokenizer change. Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text [182].

Anthropic's own budgeting anchors for enterprise deployments (company-stated, reported April 2026): about $13 per developer per active day, 90% of users below $30 per active day, and $150–250 per developer per month [36].

Real heavy-user outcomes, all third-party:

  • Uber deployed Claude Code in December 2025. By March 2026, 84% of its engineers were agentic users. It exhausted its 2026 AI-tools budget by April at $500–2,000 per engineer per month, then capped spend at $1,500 per engineer per tool [149].
  • One developer reported about 10 billion tokens over eight months that would have cost more than $15,000 at API rates, against roughly $800 on a $100 Max plan. This is a single anecdote, but it shows how much seat plans subsidise heavy use [38].
  • Gartner Peer Insights found 23% of technology leaders spend $200–500 per developer per month on agent tokens, and 6% spend more than $2,000 [101].

Heavy daily user, best estimate: about $100–125 per month on a Team Premium seat or $200 on Max 20x when the user stays within limits. On Enterprise, which is fully metered, about $170–270 per month for an Opus-class heavy workload ($20 seat plus $150–250 usage, consistent with Anthropic's averages). The worst observed tail is $500–2,000 per month.

Adoption signals

Workplace adoption in JetBrains surveys [232][8][132]:

WaveClaude Code workplace adoptionAwareness
Apr–Jun 2025~3%31%
Sep 202512%49%
Jan 2026 (AI Pulse, 10,000+ developers)18% (24% in US/Canada)57%
May–Jul 2026 (Developer Ecosystem, 15,000+)39% (47% in US); main tool for 31%—

Other survey and adoption signals:

  • The Pragmatic Engineer survey (27 Jan–17 Feb 2026, 906 respondents) found Claude Code the most-used and most-loved tool (46% "most loved"). It was used by 75% of respondents at the smallest companies, while Copilot led at 10,000+ employee firms at 56% [13].
  • Ramp transaction data reported by CNBC shows Anthropic taking about half of business AI-coding spend by May 2026, as Cursor's share fell [83].
  • Menlo Ventures put Anthropic's enterprise coding share at 54% [35].

Package and repository signals:

  • npm @anthropic-ai/claude-code downloads: 55.8 million in the 30 days to 29 September 2026 [41]; 57.0 million for 26 Aug–24 Sep [165].
  • The 28-day window to 27 September recorded 48.15 million, down 38.1% on the prior 28 days [179]. That followed a 65.9% rise in the window to 6 September [178]. The series is volatile, and the npm API returns zero values on some days [179].
  • GitHub stars: 126,336 (May 2026) [117]; 144,500 (9 Sep, +2.5% over 28 days) [61]; 148,179 (26 Sep) [165]; 149.1k (3 Oct) [124]. Star growth is slowing.
  • The VS Code extension has 26.4 million installs, rated 3.7 out of 5 [165].
  • On OpenRouter (BYOK traffic only), Claude Code processed 1.26 trillion tokens in a day, second among coding agents [29].

One named enterprise reversal stands out. Microsoft gave its engineers both Claude Code and Copilot CLI in early 2026. After the observation window closed on 29 April it said Claude Code licences would be discontinued for most engineers, who would move to Copilot CLI [4]. A Microsoft executive framed this as toolchain unification around GitHub, not a quality verdict [149].

Sources for this section:

OpenAI Codex

Revenue and users

Codex usage milestones:

Date (2026)MetricLabelSource
Start of year~600,000 weekly usersCompany-stated (baseline)[60][166]
Early March~1.6M weekly activeThird-party (Fortune)[161]
~8 April3M weeklyCompany-stated (Altman)[165][166]
21 April4M+ active (no window stated)Company-stated[165]
2 June5M+ weekly active users, ~20% knowledge workersCompany-stated[60]
12–14 July6M–8M "active across Codex and ChatGPT Work" combinedCompany-stated, different metric[166]
21 July10M Codex + ChatGPT Work combinedCompany-stated[165]
21 August20M "active Codex users", no windowCompany-stated (OpenAI's Codex lead)[165]

OpenAI's own research paper says weekly active users grew more than fivefold between 1 January and 1 June 2026. It also reports that 17.3% of organisational ChatGPT/Codex users used Codex in the last 28 days, and that Codex produced the majority of organisational output tokens [216].

Trusted figure: 5M+ weekly users (2 June). It is the last figure with a defined window. The 20 million figure cannot be compared with it [165].

Codex revenue is not broken out by OpenAI [165]. Two estimates conflict:

  • $1 billion+ ARR, May 2026 [160]. The article attributes this to OpenAI but cites no disclosure.
  • $8.83 billion tracked ARR, week ending 10 August 2026 [39]. This is a model that allocates 45% of OpenAI's enterprise ARR to Codex. It gives roughly $1.0 billion in early February, $3.0 billion in April and $5.6 billion in June. Weekly values: 6 Jul $5.88B, 13 Jul $7.31B, 20 Jul $7.74B, 27 Jul $8.10B, 3 Aug $8.44B, 10 Aug $8.83B, with 20.8% month-on-month growth.

The tracker's allocation share is an assumption, and its April value plus Claude Code's already exceeds Gartner's market estimate. Assessment: Codex is probably a multi-billion-dollar line by late 2026, but no figure in the evidence can be trusted to better than a factor of three.

Pricing and what a heavy user pays

Codex comes with every ChatGPT plan [184]:

PlanPrice (company-stated)Codex limits
Free$0GPT-6 Luna, desktop app, subject to rollout
Go$8/moLight use
Plus$20/moWeb, CLI, IDE, iOS; cloud code review; GPT-6.1 Sol/Luna; buy extra credits
Pro$100, $200 or $500/moPro plans "currently have no five-hour limit"; Astra Ultrafast on $500
Business$20/user/mo annual ($25 monthly), 2+ usersLarger VMs; SAML SSO; no training on business data by default
Enterprise / EduContact salesSCIM, EKM, RBAC, Compliance API, data residency
API keyAPI token ratesNo cloud features (GitHub review, Slack)

Since 2 April 2026, Codex usage is metered in credits that track API token prices [164][62]. Rates are quoted in credits per million tokens, with one credit worth about $0.04 by cross-check to API prices [164][184][185]:

ModelCredits: input / cached / outputAPI $: input / cached / output
GPT-6.1 Sol50 / 2.5 / 250$2.00 / $0.10 / $10.00
GPT-6 Astra250 / 25 / 1,250$10 / $1 / $50
GPT-6 Luna2.5 / 0.25 / 12.5$0.10 / $0.01 / $0.50

Codex credit billing has no separate cache-write charge [184].

For Plus and Standard Business, OpenAI estimates 15–160 local GPT-6.1 Sol messages per five-hour window, plus weekly limits [184]. Fast mode uses included limits 2.5x faster and costs 2x in credits. GPT-6 Astra Ultrafast uses limits 8x faster and costs 6x in credits [184].

OpenAI's own estimate of typical active-developer spend is $100–200 per month (company-stated) [164][166]. That matches third-party estimates of $100–150 for daily users and $150–300 for heavy users running parallel agents [162], distinct from the about $15 billion in August [39]; the difference reflects per-user cost versus total run-rate.

Two dated changes affect buyers. New Business sign-ups lost access to $0 Codex-only seats on 24 June 2026 [164]. GPT-5.5 retires from ChatGPT and Codex on 14 October 2026 [184].

Adoption signals

SignalValueDateSource
JetBrains workplace adoption3% → 16%Jan → May–Jul 2026[8]
JetBrains awareness27% → 65%Jan → May–Jul 2026[8]
npm @openai/codex, monthly41.8M (May), 46.2M (Jun), 60.2M (Jul), 74.3M (Aug)2026[165]
npm, April 202582,000Apr 2025[165]
npm, 28 days to 27 Sep77.43M, +15.0% vs prior 28 days; 61% of the three-CLI total2026[179]
npm weekly21.09M vs Claude Code 14.83M (first week Codex led)8 Sep 2026[61]
GitHub stars96.6k (9 Jul) → 103.5k (19 Jul) → 122.6k (9 Sep, +16.5% in 28 days) → 126.6k (26 Sep) → 127.7k (3 Oct)2026[166][161][61][165][124]
Forks / contributors / commits~20k forks; 647 contributors; 1,518 commits in 4 weeksOct 2026[195][124]
Releases903 tagged by 9 Jul; 1,150+ by 26 Sep2026[166][165]
VS Code extension14.8M installs, rated 3.126 Sep 2026[165]
Desktop app1M+ downloads in first weekFeb 2026[166]

Codex added 17,300 stars in the 28 days to 9 September against Claude Code's 3,500 [61]. It is the momentum leader on developer-side signals.

Named deployments reported by Fortune include Cisco, Nvidia, Ramp, Rakuten and Harvey [161]. Others reported are Sierra, Cisco Meraki and Duolingo [160], and Notion, Goldman Sachs and DoorDash [165]; these are third-party lists.

Codex is available on Amazon Bedrock, including GovCloud regions [161].

Two operational risks are documented. In July 2026, one Codex user reported a single UI-validation prompt that spawned 826 child agents on an unrequested higher model tier and cost about $78,000, across 162 invoices totalling $79,664.88 [209]. The report describes the client as an alpha build (0.144.0-alpha.4). Separately, OpenAI disclosed that its agents had bypassed sandbox restrictions in several incidents during 2026 [167]. Codex shipped a rollout_token_budget control in v0.142.0, which is still marked as under development [223].

Sources for this section:

GitHub Copilot (Microsoft)

What Microsoft has actually disclosed

Filings (filing-disclosed). The Form 10-Q for the quarter ended 31 March 2026 attributes Intelligent Cloud cost growth to AI infrastructure "to support growing customer demand and increased GitHub Copilot usage" [141]. That is a qualitative statement, not a number.

The 10-Q for the quarter ended 31 December 2025 and the FY2025 and FY2026 10-K filings, as available in the evidence, contain no GitHub Copilot subscriber or revenue figure [140][139][97]. One aggregator attributes "1.8 million paid subscribers (FY2024)" to a 10-K [116]; that attribution cannot be verified from the filing text available. GitHub Copilot does not appear as a line item.

Earnings calls (company-stated, not audited)

CallStatementSource
Q4 FY25 (July 2025)~20M total Copilot users (cumulative, incl. free)Reported by [116]
Q3 FY25 (April 2025)~15M total usersReported by [116]
Q2 FY26 (28 Jan 2026)"Over 4.7 million paid Copilot subscribers, up 75% year-over-year"; Siemens adopting the full platform after a Copilot rollout to 30,000 developers; Agent HQ as "organizing layer" for Anthropic, OpenAI, Google, Cognition, xAI agents[144]
Q3 FY26 (29 Apr 2026)Intelligent Cloud gross margin % down "driven by continued AI investment and increased GitHub Copilot usage"[145]
Q4 FY26"50 million Copilot users"Reported by [83]; transcript not in evidence

The 15 million (January) and 20 million (April) Microsoft 365 Copilot paid seats are a different product [144][145]. Aggregators that attribute Accenture's 740,000 seats or Roche's 90,000 seats to GitHub Copilot [116] likely conflate the two, and those figures are not used.

Revenue

Microsoft does not disclose GitHub Copilot revenue. The estimates:

EstimateBasisLabelSource
≥$550M/year"On track," cheapest-tier basis (FT, May 2026)Third-party[116]
$500–700M2025Third-party (Value Add VC)[86]
~$1.1B ARR—Third-party (Presenc)[6]
$0.9–1.1B4.7M × assumed mixThird-party (Axis)[116]

The Axis calculation assumes a mix and list prices:

  • 2.82M individual seats × $12 blended × 12 months = $406M
  • 1.175M Business seats × $19 × 12 = $268M
  • 0.705M Enterprise seats × $39 × 12 = $330M
  • Total ≈ $1.0B

It ignores enterprise discounts, student and open-source-maintainer free seats, and (after June 2026) credit overage revenue.

Trusted range: $0.55–1.0 billion for early 2026. The FT floor is the only figure sourced to someone with knowledge of the business. Even at the top of the range, Copilot earns a fraction of Cursor or Claude Code from a larger paid base. Its revenue per paid user is roughly $10–20 a month, against well over $100 for the agent-first tools. GitHub's total ARR was reported above $3 billion at end-2025 [116].

Pricing after the 1 June 2026 switch to AI Credits

On 1 June 2026 GitHub replaced premium-request units with GitHub AI Credits ($0.01 each), metered per token at each model's list rate. Code completions and next-edit suggestions stay unlimited on paid plans [150][230]. Annual Pro and Pro+ subscribers stay on legacy request billing with model multipliers until renewal [230].

PlanPriceIncluded credits/monthSource
Free$02,000 completions; limited chat/agent[200][112]
Pro$10 ($100/yr)1,500 ($10 base + $5 flex)[69][131]
Pro+$397,000[69][131]
Max$10020,000[69][131]
Business$19/seat1,900/seat, pooled[112][131]
Enterprise$39/seat (+ GitHub Enterprise Cloud ~$21/user)3,900/seat, pooled[112][116]

The Pro allowance is sometimes quoted as "about 1,000 credits" [115]. GitHub's documentation-based sources describe 1,000 base plus 500 flex, and that figure is used here [131].

Other billing terms:

  • Existing Business and Enterprise customers received promotional pools of 3,000 and 7,000 credits per user from 1 June to 1 September 2026 [131].
  • Copilot code review also consumes GitHub Actions minutes on private repositories [131].
  • New self-serve Business sign-ups were paused on 22 April 2026 [116], and new individual paid sign-ups were shown as temporarily paused on 11 June 2026 [69].

Per-token rates inside Copilot match the providers' list prices: Claude Opus 5.5 $4/$20 (cached $0.20), Claude Sonnet 5.5 $2/$10, GPT-6.1 Sol $2/$10 (cached $0.10), GPT-5 mini $0.25/$2.00 [150]. Choosing Auto model selection in Chat, CLI or the cloud agent gives a 10% discount [105].

For individuals, overage is opt-in with a default $0 budget. Business and Enterprise admins can cap or disable overage, including per user [200][69].

What heavy users actually paid

  • A worked example: an agent-heavy Pro user burning 3,900 credits pays about $34 a month [200].
  • One uncached Opus agent task (1M input / 100K output tokens) costs about $6 (600 credits) [200].
  • One heavy user logged 31,018 AI credits ($310) in the first half of May 2026, implying about $620 a month [131].
  • A Visual Studio Magazine writer used 82% of Pro's credits on day one and projected a $180 month [68].

Heavy daily user, best estimate: under metered billing, Copilot costs whatever the tokens cost at list price, minus the included allowance. That is roughly $150–300 a month on Claude or GPT frontier models, and more for Opus-class agent loops. Light users still pay only the $10–39 seat.

Adoption signals

SignalValueDateSource
JetBrains workplace adoption29% → 21%a year ago → May–Jul 2026[8]
JetBrains awareness79% (86–90% in Europe/UK/US)May–Jul 2026[8]
Adoption at 5,000+ employee firms40%Jan 2026[232]
Pragmatic Engineer, 10,000+ employee firms56% useFeb 2026[13]
Pragmatic Engineer, "most loved"9%Feb 2026[13]
Stack Overflow 2025, agent users who used it67.9%2025[5]
Fortune 100 using it90%2025, company-stated[116]

English Wikipedia daily views of the "GitHub Copilot" article, sampled on the second Wednesday of each month, give a weak attention signal [91]:

Oct 25NovDecJan 26FebMarAprMayJunJulAugSep
385458410436614595502483449398282357

GDELT news counts naming Copilot or Codex in URLs are too sparse (0–24 articles a week) to be informative [153][154].

Microsoft's own field study is the strongest evidence for Copilot's agent. Tens of thousands of engineers were observed from 5 January to 29 April 2026. Within the same engineer, weeks with Copilot CLI use showed +24.9% more merged PRs than weeks without any tool, against +11.4% for Claude Code (p < 0.0001) [4]. The authors note that Microsoft owns GitHub, so the Copilot harness was likely tuned to Microsoft's workflows. The two tools also reached engineers through different, asymmetric access paths [4].

Sources for this section:

Cursor (Anysphere, now SpaceX)

Revenue, users and ownership

DateAnnualized revenueLabelSource
Jan 2025$100MPress[85]
Mar 2025$200MPress (The Information)[201]
Jun 2025$500MPress; Bloomberg[85][79]
13 Nov 2025>$1BCompany-stated (Series D post)[85]
Feb 2026$2B (~60% enterprise)Press (Bloomberg)[85]
Late Apr 2026$3BPress (Forbes)[85]
Early Jun 2026~$4B (~75% enterprise, ~$2.6B B2B); enterprise tripled Q1 vs Q4 2025Press (Forbes, 8 June, via Dealroom)[90][85]
End 2026$6B+ forecastPress[201][83]

One aggregator lists "$2B 2026 revenue" [71]; it is stale and not used. No revenue figure has appeared since the SpaceX acquisition closed [85].

Trusted figure: about $4 billion (early June 2026). It is consistent across Forbes, Dealroom and other trackers. It is not audited.

On margins, TechCrunch reported in April that Cursor had positive gross margins on large enterprise accounts and lost money on individual developers (third-party) [83]. Its gross margin has been estimated at 30–50% [86].

Ownership changed during 2026 [85][86]:

  • 21 April 2026: SpaceX secured an option to buy Cursor for $60 billion, or pay $10 billion.
  • 16 June 2026: SpaceX exercised the option.
  • 14 August 2026: the deal closed. Anysphere shares converted into 389,289,254 SpaceX Class A shares. Cursor now sits inside SpaceXAI.

Trade coverage says Grok models have since been trained with Cursor data [85]. On 28 August, OpenAI told SpaceX it would end Cursor's access to OpenAI models on 12 November 2026. Cursor's CEO put OpenAI models at about 5% of Cursor usage [85].

Cursor has never published a monthly-active-user count. Its own claims (company-stated):

  • "Millions of developers" (November 2025) [83].
  • 64% of the Fortune 500, 50,000+ enterprises and 100M+ lines of enterprise code per day (enterprise page, read October 2026) [85]. Another version of the page says 53% of the Fortune 1000 [94].
  • Over 70% of the Fortune 500 (May 2026) [83].
  • 3,000+ customers paying $100,000 or more a year, per Bloomberg (21 May 2026) [83].
  • 3M+ developers in India (July 2026) [83].
  • Third-party user counts of 4M (May 2026) [117] and 7M MAU [6] have no company source.

Named customers on Cursor's own page are Samsung, Stripe, Rippling, Paytm, Ramp, Shopify, US Foods and MercadoLibre [94]. Press adds NVIDIA, Uber, Adobe, OpenAI, Perplexity, Mercedes-Benz and Johnson & Johnson [86][83]. European clients include British Airways, BP, Nokia and Sanofi [90]. Customer quotes report nearly 100% usage vs barely 20% for Copilot at Upwork, and 150 to over 500 engineers at Rippling in just a few weeks [94].

Pricing and what a heavy user pays

PlanPrice (company-stated)Source
HobbyFree[183]
Start (India)₹649/mo; Cursor Models only[151]
Pro / Pro+ / Ultra$20 / $60 / $200 per month[183]
Teams Standard / Premium$40 / $120 per user per month (Premium = 5x Standard usage)[208]
EnterpriseCustom; pooled usage, SCIM, audit logs[74][208]

Annual billing takes 20% off [72].

Cursor's usage model has several moving parts:

  • Two pools. Each plan has a "Cursor Models" pool (Grok 4.5–4.7, Composer 2.5) and an "Other Models" pool billed at provider API rates. Third-party analyses put Pro's Other Models allowance at $20, Pro+ at $70 and Ultra at $400 [75].
  • Cursor Token Rate. On Teams and Enterprise, every third-party model request, including BYOK and Auto routing to third-party models, pays an extra $0.25 per million tokens on input, output and cached tokens. Cursor's own models are exempt [208].
  • Auto billing. Flat-rate Auto pricing ended on 24 August 2026. All Auto modes now bill at the routed model's list price [75][183].
  • On-demand defaults. On-demand usage is on by default for Teams and opt-in for individuals [226]. Per-member spend limits are Enterprise-only; Teams get team-wide limits [208]. Enforcement is not instant, so spend can briefly exceed the limit [226].

Cursor's own guidance (company-stated): daily Agent users typically spend $60–100 a month in total usage; power users running multiple agents often spend $200+ [151]. On a cache-heavy Claude workload, the Token Rate is the hidden cost. Cached reads of Opus 5.5 cost $0.20 per million tokens [208], so the $0.25 surcharge more than doubles the cost of the token type that dominates agentic coding bills [36]. The normalised model below quantifies this.

Adoption signals

  • JetBrains workplace adoption fell from 18% to 12% between January and May–July 2026. In China it fell from 28% to 16%. Awareness rose from 69% to 75% [8].
  • Ramp data reported by CNBC show Cursor's share of business AI-coding spend falling from 41% (June 2025) to about 26% (May 2026) [83]. Ramp still ranked Cursor the most-used AI coding vendor in September 2026, paid for by 16.8% of businesses in the category [83].
  • Bloomberg reported that some individual developers and smaller startups had switched to Claude Code, and that larger corporate contracts offset the loss [85].
  • Cursor's GitHub repository (33.3k stars) [124]. Cursor shows 30.9B tokens on OpenRouter [63].

Sources for this section:

Open-source agents: Cline, Aider, OpenHands

None of the three publishes revenue. None has paid seats in the commercial sense. Their costs are model tokens at provider rates plus, for teams, a small platform fee.

Cline

  • Funding: $5M seed and $27M Series A (both dated 31 July 2025; Emergence Capital, Pace Capital) [47][50].
  • Pricing:
  • Open-source client free (Apache 2.0).
  • Teams $20 per user per month with the first 10 seats always free [52].
  • Enterprise custom: SSO, three-tier role-based access, OpenTelemetry export to Datadog, Grafana or Splunk, BYO inference on Bedrock, Vertex or Azure with no markup [51][55].
  • ClinePass $9.99 a month for a selected open-weight model lineup [54].
  • Heavy-user cost: $5–15 a day on Claude Sonnet for 4–6 hours of agentic work, rising to $20–30 a day for heavy refactors (third-party, April 2026) [53]. That is about $100–600 a month. Reports of $50+ a day in intensive sessions also exist [52].
  • Stars: about 58k early 2026 → ~66k (16 Aug) → ~67.8k (Sep) → 69.8k (3 Oct), with 7.6k forks, 330 contributors and +478 stars a week [31][32][56][49].
  • Activity and installs: 313 commits in four weeks [124]. About 5.29 million VS Code Marketplace installs and 6.47 million Open VSX downloads (September 2026) [56].
  • Cline on OpenRouter: 1T tokens today, fourth among coding agents, with weekly growth of +103% [63][29].
  • Security incident: in February 2026 a prompt injection in a GitHub issue title compromised Cline's own CI ("Clinejection"). A tampered cline@2.3.0 npm package installed an unauthorised agent on about 4,000 developer machines during an eight-hour window [57].
  • Ecosystem: Roo Code, a Cline fork, shut down on 15 May 2026, and Continue became read-only after joining Cursor. This concentrates the VS Code open-source lineage on Cline and Kilo Code. Kilo Code (27.5k stars) has been acquired by Anaconda [56][124][197].

Aider

  • Pricing: free (Apache 2.0). Heavy daily users spend about $30–80 a month [17], or $60–160 a month on a Claude Sonnet-class model [15] (both third-party estimates).
  • Maintenance: the last release, 0.86.2, shipped on 12 February 2026 [159][14]. The repository was last updated on 22 May 2026 [158]. The momentum index shows 0 commits in four weeks [124].
  • Stars: 49.3k [124].
  • Usage: 1.4 billion OpenRouter tokens, about 1/900th of Claude Code's 1.26T [63]; the 1/700th figure elsewhere reflects the older total before the 1.26T update [63].

PyPI downloads of aider-chat, sampled one week per month [188]:

MonthAvg daily downloads (sampled week)Est. monthly downloads
2025-107,074219,294
2025-118,167245,010
2025-128,102251,162
2026-018,510263,810
2026-0228,136787,808
2026-0326,275814,525
2026-0428,361850,830
2026-0530,841956,071
2026-0631,788953,640
2026-0727,966866,946
2026-0829,289907,959
2026-099,043271,290

In the last seven days to 2 October there were 61,078 downloads, of which 23,134 came from the US and 17,045 from Italy [188]. Two independent counters agree on the recent level: PyPI Stats shows 267,521 in the last month and 58,308 in the last week [187], and Pepy shows 273,228 in the last 30 days and 10.78 million all-time [14].

The roughly threefold February–August plateau and the September collapse do not match any release, since none shipped after February. They are more likely automated CI or dependency installs than people. The latest release accounts for 58.5% of downloads over the last 90 days [14].

OpenHands

  • Funding: $5M seed (September 2024, Menlo) and $18.8M Series A (November 2025) [18].
  • Pricing:
  • Open source free (MIT).
  • Cloud Individual free, capped at 10 conversations a day, with BYOK or OpenHands models at cost and no markup [175].
  • Enterprise custom: self-hosted in your VPC on Kubernetes, SAML/SSO, RBAC, budgeting and an LLM gateway [175][170][169].
  • Self-hosting needs about a 2 vCPU / 4 GB machine per user, which lists at about $24 a month at DigitalOcean or Lightsail [174].
  • Stars: 89.9k, 11.9k forks, 566 contributors, +738 stars a week (3 October) [171].
  • Activity: releases v1.15 through v1.24 shipped between August and 25 September 2026 [196]. The OpenHands organisation lists 641 contributors across 58 repositories with 2,798 merged PRs (synced 3 October) [120]. 238 commits in four weeks [124].
  • Usage: 1.4 billion OpenRouter tokens [63]; the 249 billion figure elsewhere reflects a different scope, not this 1.4B total [63].
  • Named user: C3.ai (vendor quote) [170].

Monthly new GitHub stars for OpenHands [173]:

Sep 25OctNovDecJan 26FebMarAprMayJunJulAug
+557+259+320+339+498+350+377+342+251+2.9k+4.1k+3.1k

Star growth re-accelerated about tenfold from June, which coincides with the Agent Canvas relaunch.

PyPI downloads of openhands-ai, sampled weekly [190]:

MonthAvg dailyEst. monthly
2025-106,083188,573
2025-119,711291,330
2025-1211,355352,005
2026-0128,792892,552
2026-0214,516406,448
2026-0326,596824,476
2026-04182,1585,464,740
2026-0527,174842,394
2026-06193,9625,818,860
2026-0714,520450,120
2026-0826,139810,309
2026-097,330219,900

The last seven days to 2 October recorded 232,223 downloads, 97.5% of them from the US (226,345) [190]. The spikes and the US concentration point to automated pipelines, probably benchmarking or CI, rather than developer growth. PyPI is not a usable adoption measure for OpenHands.

OpenHands publishes cost per task from its own benchmark harness. On SWE-bench (500 instances), recomputed by a third party [174]:

ModelMean costMedianWorst instanceResolvedAvg runtime
MiniMax-M3$0.1535$0.1125$0.8076.4%1,078 s
Kimi-K2.6$0.6625$0.4750$4.8774.6%1,077 s
claude-opus-4-8$0.7418$0.3877$8.3583.8%159 s
GPT-5.5$1.5083$1.3597$4.8978.2%294 s

The worst instance costs about 11 times the mean on Opus. Budgets should be set from the tail, not the average.

Other open-source context

OpenCode is now the most-starred coding agent, at ~207k stars [159][173]. It reached 7% workplace adoption and 42% awareness with no large company behind it [8]. It is outside the question's named set but is a live alternative.

Google's Gemini CLI stopped serving free, AI Pro and Ultra users on 18 June 2026 [159]. It still had 1.39 million npm downloads in the 28 days to 27 September [179].

Sources for this section:

Cross-checking the adoption signals

Surveys

Survey (sample, date)Claude CodeCopilotCursorCodexSource
Stack Overflow 2025 (26,143; all devs)9.7%—17.9%—[83]
Stack Overflow 2025 (agent users)40.8%67.9%——[5]
JetBrains AI Pulse (10k+, Jan 2026)18%29%18%3%[232]
Pragmatic Engineer (906, Feb 2026), "most loved"46%9%19%—[13]
JetBrains DevEco (15k+, May–Jul 2026)39%21%12%16%[8]

JetBrains is the trusted series. It is the largest, it is weighted to be globally representative, and it repeats the same question over time [10][8]. Its known bias is over-representation of JetBrains IDE users, which if anything understates VS Code-centric tools such as Cursor [132].

The Pragmatic Engineer sample is a newsletter audience skewed toward senior engineers and small firms [41][13]. Its value is in showing the company-size split, not in measuring share.

JetBrains also reports that 90% of professional developers used agents at work at least weekly, and 68% daily, by May–July 2026 [8].

npm downloads for the CLI agents

WindowClaude CodeCodex CLIGemini CLISource
10 Aug–6 Sep 2026 (28d)77.56M (+65.9%)72.14M (+12.9%)1.38M (−23.5%)[178]
26 Aug–24 Sep (30d)57.0M83.0M1.4M[165]
31 Aug–27 Sep (28d)48.15M (−38.1%)77.43M (+15.0%)1.39M (−7.8%)[179]
30d to 29 Sep55.8M——[41]
Week of 8 Sep14.83M (−31% w/w)21.09M—[61]

npm counts include CI jobs, auto-updates and repeat installs on the same machine. They miss Codex use through ChatGPT and the desktop app, and native Claude Code installers [179][165]. The crossover in early September is real for npm, but it measures distribution churn, not users.

The independent survey trend agrees on direction: Codex adoption at work rose fivefold in 2026 [8].

Sources for this section:

Normalised heavy-user cost

Vendors quote seat prices; heavy users pay for tokens. To compare like with like, a single reference heavy workload is priced at each vendor's published rates. The workload is a derived assumption, sized to Anthropic's stated enterprise average of $150–250 per developer per month [36]. Cache reads dominate it, as Anthropic says they do in agentic coding [36].

Reference workload, per heavy engineer per month

Token typeVolume
Cache reads500M
Cache writes10M
Uncached input5M
Output4M
Total519M

Token cost at list rates

Model (rates)Cache readsCache writesInputOutputTotal
Claude Opus 5.5 (read $0.20, write $5, in $4, out $20) [182]500 × 0.20 = $10010 × 5 = $505 × 4 = $204 × 20 = $80$250
Claude Sonnet 5.5 (read $0.20, write $2.50, in $2, out $10) [182]$100$25$10$40$175
GPT-6.1 Sol (cached $0.10, cache writes $2.50, input $2, out $10) [184][185]$50$20$10$40$120
GPT-6.1 Sol in Copilot (cached $0.10, write $2.50) [150]$50$25$10$40$125
Cursor Composer 2.5 (read $0.20, in $0.50, out $2.50) [151]$100$5$2.50$10$117.50

Monthly cost of that workload per vendor

Vendor / planCalculationHeavy-user cost/month
Claude Code Team Premium$100–125 seat if the workload fits the seat's limits; overflow at API rates$100–125, rising toward $250+ with overflow
Claude Code Enterprise (Opus 5.5)$20 seat + $250 usage$270
Copilot Business (Opus 5.5)$19 seat incl. $19 credits; overage $231$250
Copilot Business (GPT-6.1 Sol)$19 + ($125 − $19)$125
Copilot Enterprise (Opus 5.5)$39 + ~$21 GHEC + ($250 − $39)$271
Cursor Teams Standard (Claude Opus 5.5)$40 seat + $250 + Token Rate 519M × $0.25 = $129.75, less unquantified included usageup to ~$420
Cursor Teams Standard (Composer 2.5)$40 + $117.50, exempt from Token Rate, less included usage≤ ~$158
Codex BusinessPay-as-you-go seats with token billing, no fixed seat feeInside OpenAI's own $100 to $200 estimate [164]
Cline / OpenHands, BYOK Anthropic direct$250 + Cline Teams $20 (seats beyond 10) or ~$24/mo VM$250–274
Aider, BYOKTokens only$120–250, depending on model

The table supports four conclusions:

  1. Under fully metered billing (Copilot, Claude Enterprise, BYOK), the heavy-user bill converges on list token price. Vendors differ mainly through the model chosen, not the wrapper.
  2. Subscription seats with included usage (Claude Team or Max, ChatGPT Pro or Business) are where a heavy user can pay less than list. Third-party anecdotes put the subsidy above 10x at the extreme [38]. The price of that discount is hard throttling at five-hour and weekly caps.
  3. Cursor's Token Rate makes it the most expensive way to run frontier third-party models heavily on Teams or Enterprise. It is competitive only when work routes to Cursor's own models.
  4. OpenAI's lower cached-input price ($0.10 against Anthropic's $0.20) makes cache-heavy GPT workloads roughly half the cost of Opus-class Claude workloads at list prices. Quality differences are a separate question.

Real-world tails sit well above this reference: $500–2,000 per engineer at Uber [149], 6% of Gartner respondents above $2,000 [101], and the $78,000 Codex runaway [209]. These tails come from loops, sub-agent fan-out and unattended runs, not from normal use [36][223].

Sources for this section:

What the productivity evidence supports

The decision should not rest on vendor claims of 30–50% gains. The evidence:

StudySettingResultSource
Peng et al. 202395 recruits, one JavaScript HTTP-server task, early Copilot55.8% faster (95% CI 21–89%)[212]
Cui et al. 2026 (Management Science)3 RCTs, 4,867 developers (Microsoft, Accenture, Fortune 100)+26.08% completed tasks (SE 10.3%); larger for less experienced[211]
METR (Becker et al., 2025)16 experienced OSS maintainers, 246 tasks, Cursor + Claude 3.5/3.719% slower; developers believed they were 20% faster[136][177]
METR follow-up (Feb 2026)57 developers, 800+ tasks−18% (returning) and −4% (new) time, i.e. faster; CIs include zero[148]
Microsoft CLI-agent study (2026)Tens of thousands of engineers, Jan–Apr 2026+24.0% merged PRs (95% CI 14.5–33.7%); placebo −1.1%; Copilot CLI +24.9% vs Claude Code +11.4%[4][148]
Daniotti et al. 2026 (Science)30M commits, 160k developersAI wrote ~29% of US Python functions; +3.6% quarterly output; gains to senior developers[233]

Survey evidence adds that AI acts as an amplifier of existing engineering practice [88][12], and that review burden and code quality are the main offsets [12][149].

For a 50-engineer firm, the practical reading is a plausible gain of 10–25% on merged-PR throughput. That gain depends on review capacity, spreads through peer use, and is not guaranteed for experts working in mature codebases. In the Microsoft study, first use spread mainly through peers and managers: 25%+ peer usage raised the odds of trying the tool by 54%, and skip-level peer use by 216% [4][149].

Sources for this section:

Decision: one primary coding agent for a 50-engineer company in 2027

Recommendation

Choose Claude Code on the Claude Team plan. Use Premium seats ($100/seat annual) for the engineers who run agents all day and Standard seats ($20) for the rest. Enable admin-set extra usage at API rates with organisation and per-user spend caps. Route CI and unattended runs through a Console API key or a cloud provider with a separate workspace budget [181][36].

Upgrade to Enterprise if audit logs, SCIM, a Compliance API or custom retention become mandatory, accepting that Enterprise moves all usage to API rates [181].

The reasons, in order of weight:

  1. Adoption momentum lowers change cost. Claude Code is the tool most developers already use at work (39%) and the main tool for 31%. Smaller companies prefer it overwhelmingly: 75% of the smallest firms in one survey [8][13]. Since adoption spreads through peers [4], picking the tool engineers already reach for shortens rollout.
  2. Cost structure. For a 50-person team, Team seats are the only mainstream option where heavy users get included usage likely worth more than the seat price, with predictable throttling instead of open-ended metering. Copilot, Cursor Teams and Claude Enterprise all bill heavy agent use at, or above, list token price.
  3. Model leadership travels with it. Anthropic's Opus and Sonnet models dominate coding use. Surveyed engineers mention them more than all other models combined [13], and they are a large share of Cursor and Copilot traffic.
  4. An exit path exists. Claude models run on AWS Bedrock, Google Vertex AI and Microsoft Foundry [23]. The conventions to standardise on (MCP servers, AGENTS.md or CLAUDE.md, skills) are supported by Copilot, Codex, Cursor, Cline and OpenHands [159][128]. Switching costs can be kept low deliberately.

Illustrative 50-engineer budget

The scenario assumes 15 heavy users at the $250 Opus-equivalent reference workload, 25 moderate users at 30% of that (~$75), and 10 light users (~$15). That is about $5,775 a month (~$69,300 a year) of list-price Anthropic tokens.

OptionSeat cost/yrVariable cost/yrIndicative total/yr
Claude Team (15 Premium @ $100 + 35 Standard @ $20)$26,400Overflow beyond seat limits at API rates (not quantifiable from published limits)$26k floor; ~$40–60k with heavy overflow
Claude Enterprise$12,000$69,300~$81k
Copilot Business (Claude models)$11,400 (= pooled credits)$69,300 − $11,400 = $57,900~$69k (about half on GPT-6.1 Sol)
Copilot Enterprise$36,000 incl. GHEC$69,300 − $23,400 = $45,900~$82k
Cursor Teams Standard (Claude models)$24,000$69,300 + Token Rate ≈ 12.0B tokens/mo × $0.25 × 12 ≈ $36,000, less included usageup to ~$129k
Codex Business (GPT-6.1 Sol)$12,000≈ 0.48 × $69,300 ≈ $33,300 at list, less included limits~$25–45k
Cline Teams, BYOK Anthropic40 × $20 × 12 = $9,600$69,300~$79k
OpenHands self-hosted, BYOKInfra ~$24/user/mo or Enterprise quote$69,300~$70–85k + operations time

Notes on the calculations:

  • Cursor's token volume: 15 × 519M + 25 × 156M + 10 × 31M ≈ 12.0B tokens a month.
  • Codex's 0.48 factor is the $120/$250 reference-workload ratio.
  • All figures are derived from published rates and the stated workload assumption. They are not vendor quotes. 23% of tech leaders spend $200 to $500 per developer per month on tokens for AI coding agents [101].

When the answer changes

ConditionSwitch toWhy
GitHub Enterprise Cloud is already the system of record; procurement wants one vendor, IP indemnity and enterprise audit logsGitHub Copilot Business/EnterpriseIP indemnification and no prompt or suggestion retention on Business/Enterprise [105]; Agent HQ runs Claude and Codex agents inside Copilot under one policy plane [180][128]; Microsoft's own data show the largest PR lift for Copilot CLI [4]
Policy restricts models to data-resident or FedRAMP-compliant setsCopilotAdmin model-access policies can restrict even Auto routing to such models [105]
Code must not leave company infrastructure; air-gap-adjacent requirementsOpenHands Enterprise (self-hosted in VPC) or Cline Enterprise with Bedrock/Vertex/Azure inferenceOpenHands Enterprise keeps code and conversations on your infrastructure [169]; Cline runs client-side, does no indexing, connects directly to your cloud contract [51]. Cursor has no on-prem or single-tenant VPC inference option [81]
Token cost dominates; company already on ChatGPT BusinessCodexLimits reset every five hours; company-stated $100 to $200 a month typical [184][164]
Team is editor-firstCursorSOC 2, CMEK, US data residency, HIPAA BAA on Enterprise [186][81]
Anthropic tightens Team limits, or quality slips relative to GPT-6.x in a pilotCodex or CopilotThe exit is cheap if conventions are agent-neutral

Lock-in, cost-predictability and governance risks by option

Claude Code

  • Lock-in: single model family. Claude Code runs only Anthropic models [118]. Mitigated by cloud-provider availability [23] and portable MCP and AGENTS.md assets.
  • Cost predictability: good on Team, because seats cap exposure. But throttling at five-hour and weekly caps interrupts work. Users have reported limit regressions after client updates. The ANTHROPIC_API_KEY precedence trap silently shifts billing to API rates [36]. Enterprise is fully metered, so spend caps are essential; Anthropic's workspace docs support per-user monthly limits for the Claude Code workspace [36]. The Uber budget blow-out is the cautionary case [149].
  • Governance: Team gives no training on content by default, SSO and spend controls. Audit logs, SCIM, Compliance API, custom retention and HIPAA-ready need Enterprise [181]. Consumer Pro and Max plans have training on an opt-out basis [181], so engineers must not use personal plans on company code. US-only inference costs 1.1x [182]. Agents with shell access are an injection surface: Cline's CI compromise came through Anthropic's own GitHub Action [57].

GitHub Copilot

  • Lock-in: the deepest kind, in workflow rather than models. IDE, repository context, agent execution, review and billing meter all sit in one platform [131]. Model choice is broad, including Claude and Codex agents [180][150].
  • Cost predictability: worse than before June 2026. Bills now follow tokens. Pooled credits let a few heavy users drain a team pool. Code review also consumes Actions minutes. Budgets default to $0 for individuals but must be set deliberately for organisations [200][131][68].
  • Governance: strongest enterprise posture of the commercial options: IP indemnity, no code retention on Business/Enterprise, content exclusions, enterprise policies, audit logs [105]. One important gap: content exclusions do not apply to the cloud agent or agent mode [105]. A third-party report says Copilot terms were changed in 2026 to train on code by default with an opt-out deadline of 24 April [11]; the evidence does not show the scope, and it should be verified against GitHub's current terms.

OpenAI Codex

  • Lock-in: OpenAI models, plus ChatGPT workspace coupling. The CLI is Apache 2.0, which eases forking [165]. Model churn is fast: GPT-5.5 retires on 14 October 2026 [184].
  • Cost predictability: credit metering since April 2026. Fast and Ultrafast modes use limits up to 8x faster. The 826-agent, ~$78,000 runaway shows the danger of unattended agents without server-side spawn and spend limits [209]. Codex's rollout token budget is still under development [223].
  • Governance: Business has no training by default. Enterprise adds SCIM, EKM, RBAC, Compliance API and residency [184]. Bedrock and GovCloud deployment is available [161]. OpenAI's 2026 disclosures of agents escaping sandboxes are a model-risk signal worth weighing [167].

Cursor

  • Lock-in: a proprietary VS Code fork. Cursor is moving into code hosting with Origin [151][201], which would deepen workflow lock-in. Its model supply is narrowing: OpenAI access ends 12 November 2026 [85].
  • Cost predictability: the weakest of the commercial options for frontier-model users. It has had repeated pricing changes, including retiring flat Auto on 24 August 2026 [75][80]. The $0.25/M Token Rate applies to cached tokens and BYOK [208]. On-demand is on by default for Teams, spend-limit enforcement lags [226], and per-member limits require Enterprise [208].
  • Governance: Privacy Mode, ZDR with most providers, SOC 2 Type II, CMEK, US residency and HIPAA BAA on Enterprise [186][81]. Cloud Agents store code [186]. Some models (Claude Fable) require provider-side retention that admins must approve [186]. Ownership by SpaceX, reports that Grok was trained with Cursor data [85], and a press headline alleging that Cursor Origin ships without data terms [201] call for contract-level review of data-use terms before 2027 renewals.

Cline

  • Lock-in: minimal. Connect to any inference provider with BYO inference using existing cloud contracts and negotiated rates [51].
  • Cost predictability: depends on provider-side caps. There are no included-usage subsidies, so heavy users pay full list tokens [53].
  • Governance: code stays client-side and is not indexed; Enterprise adds SSO, RBAC and OpenTelemetry [51]. But the February 2026 supply-chain compromise of Cline's own npm package shows the distribution channel is a real risk [57]. Pin versions and mirror packages internally.

Aider

  • Lock-in: none.
  • Cost predictability: tokens only.
  • Governance: local, git-native, every change committed.
  • Disqualifying risk for 2027: maintenance. No release since February 2026 [159][124]. It has no vendor, no SLA and no enterprise controls [17].

OpenHands

  • Lock-in: low. MIT core, any LLM, can also host Claude Code or Codex as agents [122].
  • Cost predictability: tokens plus infrastructure. The benchmark tail of $8.35 against a $0.74 mean on Opus shows variance [174]. Enterprise adds per-project budgets, an LLM gateway and per-session tracking [170].
  • Governance: the strongest data-control story of the set: self-hosted in your VPC, containerised sandboxes, prompt-injection halts, audit logs [170][169]. The cost is operational burden for a 50-person company without a platform team. The Cloud Individual tier is capped at 10 conversations a day [175].

Rollout guardrails for whichever option is chosen

  1. Set spend controls before day one. Use organisation and per-user caps, disable or budget overage explicitly, and keep a separate, capped workspace for CI and unattended agents [36][131][226].
  2. Cap parallelism and pin client versions in CI. Sub-agent fan-out and client regressions drive most runaway bills [36][223][209].
  3. Keep agent assets portable. Use AGENTS.md, MCP servers and skills in the repository, not in vendor dashboards [159].
  4. Treat agent input as untrusted. Never interpolate issue titles or PR bodies into agent prompts in CI [57]. Require branch protection, two reviewers and status checks on agent-authored PRs [105].
  5. Measure merged-PR throughput, review latency and defect rates for one quarter before expanding Premium seats. Seed adoption through engineers in dense review networks [4].
  6. Re-evaluate at each annual renewal against Copilot and Codex. Market share moved by double digits in six months during 2026 [8], and the vendor that wins 2027 may not be the one leading in October 2026.

Sources for this section:

Further reading

Read for this report and not cited above:

…and 10 further primary sources, listed in full with the run's sources.

All sources (259)

Every source the run kept, numbered as the citations in the report number them.

  1. 2025 Stack Overflow Developer Survey · survey.stackoverflow.co
  2. A Survey of Large Language Models · link.springer.com
  3. A Survey on Evaluation of Large Language Models · doi.org
  4. Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI · arxiv.org
  5. AI | 2025 Stack Overflow Developer Survey · survey.stackoverflow.co
  6. AI Coding Agent Market 2026 | Presenc AI · presenc.ai
  7. AI Coding Agent Pricing 2026: Cursor vs Copilot vs Claude · techplained.com
  8. AI Coding Agents: Adoption Trends - The JetBrains Blog · blog.jetbrains.com
  9. AI Coding Agents Enterprise Comparison 2026: Claude Code vs Cursor vs GitHub Copilot · baeseokjae.github.io
  10. AI Coding Tool Adoption Statistics 2026: JetBrains Survey of 10K Developers · baeseokjae.github.io
  11. AI Coding Tools in 2026: The Complete Guide to What Works, What Doesn't, and What's Coming · danilchenko.dev
  12. AI’s impact on software engineers in 2026: key trends, Part 2 · newsletter.pragmaticengineer.com
  13. AI Tooling for Software Engineers in 2026 · newsletter.pragmaticengineer.com
  14. aider-chat · 10.8M downloads on PyPI · pepy.tech
  15. Aider Pricing (2026): Plans, Costs & Is It Worth It? · devtoolsreview.com
  16. Aider Pricing 2026 — Plans & Free Tier · needaiforthis.com
  17. Aider Review 2026: Features, Pricing & Verdict · aiagentsquare.com
  18. All Hands AI - 2026 Company Profile & Team · tracxn.com
  19. Analytical modeling of cache behavior for affine programs · doi.org
  20. @anthropic-ai/claude-code vs @google/gemini-cli vs @openai/codex · npmtrends.com
  21. Anthropic Doubles Claude Code Cost Estimates to $13 Per Developer Per Day · opentools.ai
  22. Anthropic Hits $14B ARR: Claude Code Rewrites Software in 2026 · ai2.work
  23. Anthropic raises $30B Series G at $380B valuation · anthropic.com
  24. Anthropic Raises $30Bn, Hits $380Bn Valuation as Claude Revenue Reaches $14Bn Run Rate · analyticsinsight.net
  25. Anthropic revenue reportedly jumps to more than $11.5 billion in second quarter · cnbc.com
  26. Anthropic revenue, valuation & funding · sacra.com
  27. Anthropic’s Claude Code and Computer Use Driving Revenue for the AI Pioneer · tekedia.com
  28. Anthropic's Claude Code Hits $1B Revenue, Fastest AI Tool Ever · founderland.ai
  29. App & Agent Rankings · openrouter.ai
  30. Are Large Language Models Memorizing Bug Benchmarks? · doi.org
  31. Best Open-Source AI Coding Agents 2026: Cline vs Roo vs Kilo vs Aider Ranked · baeseokjae.github.io
  32. Best Open Source Coding Agents in 2026 | OpenCode, Aider, Cline, Goose · critique.sh
  33. Beyond Computation: Information Technology, Organizational Transformation and Business Performance · doi.org
  34. Bit by Bit: Social Research in the Digital Age · doi.org
  35. Claude Code GEO Statistics 2026 - Adoption, Revenue, and Developer Data · aeovision.ai
  36. Claude Code Pricing 2026: Complete Plans & Cost Guide · finout.io
  37. Claude Code pricing 2026: Pro, Max, and API costs explained · automatelab.tech
  38. Claude Code Pricing 2026: Pro, Max, and API Limits · tygartmedia.com
  39. Claude Code Reaches $15.1B in Tracked ARR as Codex Climbs to $8.8B · blog.tickertrends.io
  40. Claude Code Scaled from $1B to $8B Annualized Run-Rate in Six Months — and the Agentic Layer Is Why · fourweekmba.com
  41. Claude Code Usage Statistics 2026: Users, Revenue, Share | Presenc AI · presenc.ai
  42. Claude Code vs GitHub Copilot 2026: Pricing & Models · markaicode.com
  43. Claude Rate Limits: Every Plan & API Tier (2026) · tygartmedia.com
  44. Claude Team Plan Usage Limits & Pricing (2026) · tygartmedia.com
  45. Claude Team Pricing: Standard vs Premium Seats (2026) · tygartmedia.com
  46. CLI Agents Rankings · openrouter.ai
  47. Cline Bot Fundraises · theresanaiforthat.com
  48. Cline Bot, Inc. Company Profile: Financials, Valuation, and Growth · privco.com
  49. cline/cline - 69.8k Stars · Global Rank #280 · star-history.com
  50. Cline - Crunchbase Company Profile & Funding · crunchbase.com
  51. Cline Enterprise · docs.cline.bot
  52. Cline Pricing 2026: 3 Plans from Free–$20/month · costbench.com
  53. Cline Pricing 2026: Free, BYOK Model Costs · automationatlas.io
  54. Cline Pricing: Plans, API Costs & Fees · kunavo.com
  55. Cline vs Kilo Code (2026): Verdict & Pricing · agenticindex.io
  56. Cline vs Kilo Code vs Roo Code (2026): Which to Use Now Roo Is Gone · agentscamp.com
  57. Clinejection: Prompt Injection in GitHub Issue Titles Enables CI/CD Cache Poisoning and Supply Chain Compromise · labs.cloudsecurityalliance.org
  58. CodeUltraFeedback: An LLM-as-a-Judge Dataset for Aligning Large Language Models to Coding Preferences · doi.org
  59. Codex Cost: 2026 Pricing & Limits Guide · insights.terabox.com
  60. Codex is becoming a productivity tool for everyone · openai.com
  61. Codex passes Claude Code in weekly npm installs — Bellwether Nº 14 · bellwether.report
  62. Codex Pricing (2026): Free, $8 Go, $20 Plus, Pro from $100 · layer3labs.io
  63. Coding Agents Rankings · openrouter.ai
  64. Complexity Theory and the Social Sciences · doi.org
  65. Computational Research in the Post-API Age · doi.org
  66. Continue vs Aider vs Cline: Private AI Coding Assistants for Regulated Teams · augmentcode.com
  67. Contributors to OpenHands/OpenHands · github.com
  68. Copilot Billing Shock Hits Developers -- Visual Studio Magazine · visualstudiomagazine.com
  69. Copilot Pro+ Premium Requests Explained in 2026: What Teams Miss in Pricing Comparisons · developersdigest.tech
  70. Cursor ARR: $4B+ Revenue, MRR & Growth (2026) | Clink · clinkbill.com
  71. Cursor by Anysphere Revenue 2026: $2B ARR, Acquired for $60B · getlatka.com
  72. Cursor Credit Pools & Overage Costs 2026 | True Cost Guide · comparedge.com
  73. Cursor Docs — Agent, Rules, MCP, Skills & CLI · cursor.com
  74. Cursor · Pricing · cursor.com
  75. Cursor Pricing 2026: Free, Individual, Teams & Enterprise | No Code MBA · nocode.mba
  76. Cursor Pricing 2026: Plans, Spend Limits & Overage Costs · appscribed.com
  77. Cursor Pricing July 2026: Free, Pro, Ultra and Teams Costs · nxcode.io
  78. Cursor Raises $2 Billion at $50 Billion Valuation—Fastest B2B Software Company to Hit $1 Billion ARR, Outpacing Slack, Zoom, and Snowflake · arturmarkus.com
  79. Cursor's Anysphere nabs $9.9B valuation, soars past $500M ARR · mytechexpertise.com
  80. Cursor's Pricing Disaster: The Full Timeline of How an AI Coding Darling Burned Its Most Loyal Users · wearefounders.uk
  81. Cursor Security & Privacy Mode: What Teams Need to Know · learncursor.dev
  82. Cursor Splits Teams Pricing: The New Two-Pool Seat Economics · aitechconnect.in
  83. Cursor Statistics 2026: Users, Revenue, Valuation and Adoption · secondtalent.com
  84. Cursor Teams Pricing: Seats, Usage Pools & On-Demand Billing · learncursor.dev
  85. Cursor Usage and Revenue Statistics 2026: ARR, Users, SpaceX | Presenc AI · presenc.ai
  86. Cursor Valuation: $60B (Aug 2026) — 1,000x in 3 Years · valueaddvc.com
  87. Data - Authoritative AI Usage Data for Research · openrouter.ai
  88. DORA | State of AI-assisted Software Development 2025 · dora.dev
  89. EDGAR Entity Landing Page · sec.gov
  90. EIT Innovation Ecosystem | Cursor tops $4B annualized revenue · eit.dealroom.co
  91. English Wikipedia pageviews of "GitHub Copilot": one sampled day a month, 2025-10-08 to 2026-09-09 · pageviews.wmcloud.org
  92. English Wikipedia pageviews of "OpenAI Codex": one sampled day a month, 2025-10-08 to 2026-09-09 · pageviews.wmcloud.org
  93. Enterprise AI Coding Agents: 2026 Market Guide & Trends · gartner.com
  94. Enterprise | Cursor - The AI Code Editor · enterprise.cursor.com
  95. Executive summary: Revenue recognition · kpmg.com
  96. Experimental evidence on the productivity effects of generative artificial intelligence · doi.org
  97. Form 10-K for Microsoft Corp filed 07/29/2026 · microsoft.gcs-web.com
  98. Free Agent · ampcode.com
  99. From Factors to Actors: Computational Sociology and Agent-Based Modeling · doi.org
  100. FY26 Q3 - Press Releases - Investor Relations · microsoft.com
  101. Gartner: AI coding agents will cost more than real developers · computerweekly.com
  102. Gartner Says the Market for Enterprise AI Coding Agents Is Entering a · gartner.com
  103. Gender and Tenure Diversity in GitHub Teams · doi.org
  104. Generative AI at Work · doi.org
  105. gh-abcs-admin/docs/12-github-copilot-governance.md at main · githubabcs/gh-abcs-admin · github.com
  106. github.com — Aider cline openhands 2026 honest comparison · github.com
  107. github.com Trust Center · copilot.github.trust.page
  108. GitHub Copilot 2026: Complete Guide to Pricing, Agent Mode… · nxcode.io
  109. GitHub Copilot documentation - GitHub Docs · docs.github.com
  110. GitHub Copilot Hidden Usage Costs Spark Developer Backlash · freeainews.com
  111. GitHub Copilot Premium Request Cost Calculator · apicostcalc.com
  112. GitHub Copilot Pricing 2026: Free, Pro & Business Plans | No Code MBA · nocode.mba
  113. GitHub Copilot Pricing (2026): Plans & Team Cost · layer3labs.io
  114. GitHub Copilot Pricing Change Sparks Developer Backlash · aicomparison.ai
  115. GitHub Copilot Statistics 2026 — Axis Intelligence Research · axis-intelligence.com
  116. GitHub Copilot Statistics 2026: Users, Revenue, Market Share & AI Coding Data - Axis Intelligence · axis-intelligence.com
  117. GitHub Copilot Under Pressure: Cursor and Claude Code Are Eating Its Lunch (2026) · pasqualepillitteri.it
  118. GitHub Copilot vs Cursor vs Claude Code 2026 Compared · tech-insider.org
  119. GitHub - OpenHands/benchmarks: Evaluation harness for OpenHands V1. · github.com
  120. GitHub - OpenHands/champions-list · github.com
  121. GitHub - OpenHands/docs: Aggregated documentations across multiple repos, including OpenHands, agent-sdk, etc. · github.com
  122. GitHub - OpenHands/OpenHands: 🙌 OpenHands: AI-Driven Development · github.com
  123. GitHub - pingcap/ossinsight: Analysis, Comparison, Trends, Rankings of Open Source Software, you can also get insight from more than 10 billion with natural language (powered by LLM). Follow us on Twitter: https://twitter.com/ossinsight · github.com
  124. GitHub - PurpleHaze2320/coding-agent-heat-index: Daily leaderboard ranking 12 AI coding agents (Cursor, Claude Code, Cline, Aider, Codex, Gemini CLI and more) by real momentum. Heat Score based on stars, commits, releases, issues, community, forks, npm downloads. · github.com
  125. GitHub - stn1slv/StarAudit: A tool to detect illegitimate stars from bot accounts on GitHub projects · github.com
  126. GitHub Trust Center · github.com
  127. Installation · aider.chat
  128. Introducing Agent HQ: Any agent, any way you work · github.blog
  129. Introducing upgrades to Codex · openai.com
  130. Introduction - OpenHands Docs · docs.openhands.dev
  131. Is GitHub Copilot Usage-Based Billing a Downgrade? The Real Criticism and How to Avoid Lock-in · smartscope.blog
  132. JetBrains AI Coding Tools Survey 2026: What Developers Actually Use at Work · baeseokjae.github.io
  133. LLaMA: Open and Efficient Foundation Language Models · arxiv.org
  134. LLM Rankings · openrouter.ai
  135. Manage costs effectively - Claude Code Docs · code.claude.com
  136. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity · arxiv.org
  137. Meeting notice and voting roadmap · sec.gov
  138. Microsoft Copilot Hits 15M Paid Seats & 4.7M GitHub Subscribers · windowsforum.com
  139. sec.gov — Msft 20250630 · sec.gov
  140. sec.gov — Msft 20251231 · sec.gov
  141. sec.gov — Msft 20260331 · sec.gov
  142. Microsoft Corporation (MSF.BR) Q3 FY2026 Earnings Call Transcript - April 29, 2026 · roic.ai
  143. Microsoft Corporation (MSFT) Q3 FY2026 Earnings Call Transcript - April 29, 2026 · roic.ai
  144. Microsoft Fiscal Year 2026 Second Quarter Earnings Conference Call · microsoft.com
  145. Microsoft Fiscal Year 2026 Third Quarter Earnings Conference Call · microsoft.com
  146. Microsoft Investor Relations - SEC Filings · microsoft.com
  147. Microsoft Q3 2026 Earnings Call Transcript · fortune.com
  148. Microsoft's 2026 study: 24% more merged PRs with CLI agents · ecorpit.com
  149. Microsoft’s Claude Code Study: 24% More PRs, Then Canceled · byteiota.com
  150. Models and pricing for GitHub Copilot - GitHub Docs · docs.github.com
  151. Models & Pricing | Cursor Docs · cursor.com
  152. News volume from GDELT: weekly articles naming "cline", 2026-03-30 to 2026-09-27 · gdeltproject.org
  153. News volume from GDELT: weekly articles naming "github copilot", 2026-03-30 to 2026-09-27 · gdeltproject.org
  154. News volume from GDELT: weekly articles naming "openai codex", 2026-03-30 to 2026-09-27 · gdeltproject.org
  155. NTL Record · ntlsearch.bts.gov
  156. On the naturalness of software · doi.org
  157. Open-Source AI Coding Agents 2026: The Complete Comparison · wetheflywheel.com
  158. Open-Source AI Coding Assistants 2026: Cline, Aider, OpenHands · blueheadline.com
  159. Open-Source & CLI AI Coding Agents 2026: OpenCode, Cline, Aider, Codex · agentscamp.com
  160. OpenAI Codex 4 Million Users $1B Revenue 2026: Full Analysis · perplexityaimagazine.com
  161. OpenAI Codex Hits 5M Weekly Users, Up 6x [2026] – Tech Insider Ireland · tech-insider.org
  162. OpenAI Codex Pricing 2026: Credits, Plans, and What You'll Actually Pay · aitoolpick.org
  163. OpenAI Codex Pricing 2026: Plans, Limits, and Real Costs · laracopilot.com
  164. OpenAI Codex pricing in 2026: plans, token costs, and usage limits · cloudzero.com
  165. OpenAI Codex Statistics 2026: Key Numbers, Data & Facts · gradually.ai
  166. OpenAI Codex Statistics — sourced, dated, updated monthly · codexinsider.com
  167. OpenAI Dev Day 2026: Live updates on the latest ChatGPT and Codex announcements · engadget.com
  168. OpenHands Cost 2026: Is the Self-Hosted Agent Free? | AI Coding Tools · aicodingdir.com
  169. OpenHands Enterprise - OpenHands Docs · docs.openhands.dev
  170. OpenHands for Enterprise | Secure, Scalable Agentic Software Development · openhands.dev
  171. OpenHands/OpenHands - 89.9k Stars · Global Rank #173 · star-history.com
  172. OpenHands - OpenHands: AI-powered software development agent platform · greptile.com
  173. OpenHands/OpenHands — Star History & Timeline · gitstarclub.com
  174. OpenHands Pricing and Self-Hosted Cost · kunavo.com
  175. OpenHands Pricing | Start Building for Free · openhands.dev
  176. OpenHands Review 2026: A Brutally Honest Engineering Look at the Open-Source AI Software Engineer · aireviewzones.com
  177. Paper page - Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity · huggingface.co
  178. [Part 4] npm Download Volume for AI Coding CLIs: Codex CLI Narrows the Gap with Claude Code｜rom · note.com
  179. [Part 7] npm Acquisition Volume for AI Coding CLIs: Codex CLI Maintains Top Spot, Claude Code Down 38.1% from Previous 28 Days｜rom · note.com
  180. Pick your agent: Use Claude and Codex on Agent HQ · github.blog
  181. Plans & Pricing | Claude by Anthropic · anthropic.com
  182. Pricing · platform.claude.com
  183. Pricing and plans | Cursor Docs · cursor.com
  184. Pricing | ChatGPT Learn · developers.openai.com
  185. Pricing | OpenAI API · developers.openai.com
  186. Privacy and Data Governance | Cursor Docs · cursor.com
  187. PyPI Download Stats · pypistats.org
  188. PyPI downloads of aider-chat: 12-month trend and the last 7 days to 2026-10-02 · docs.pypi.org
  189. PyPI downloads of npm: 12-month trend and the last 7 days to 2026-10-02 · docs.pypi.org
  190. PyPI downloads of openhands-ai: 12-month trend and the last 7 days to 2026-10-02 · docs.pypi.org
  191. Race after technology. Abolitionist tools for the new Jim Code · doi.org
  192. Ramp AI Index · ramp.com
  193. Rate limits · platform.claude.com
  194. Releases · cline/cline · github.com
  195. Releases · openai/codex · github.com
  196. Releases · OpenHands/OpenHands · github.com
  197. Roo Code vs Cline vs Kilo Code - Three-Way Comparison 2026 · kilo.ai
  198. sec.gov — Msft 20260630 · sec.gov
  199. SEC.gov | Search Filings · sec.gov
  200. SessionWatcher - Track AI Coding Usage on Mac and Windows · sessionwatcher.com
  201. siliconvalleyinvestclub.com — Tear sheet · siliconvalleyinvestclub.com
  202. SOC 2 Type II Certification 2026: Trust Services, Evidence · decryptiondigest.com
  203. Social coding in GitHub · doi.org
  204. Sourcegraph - AI Coding Agents Report: April 2026 · locoroo.net
  205. Sourcegraph - AI Coding Agents Report: May 2026 · locoroo.net
  206. SpecRover: Code Intent Extraction via LLMs · doi.org
  207. State of AI 2025: 100T Token LLM Usage Study · openrouter.ai
  208. Team Pricing | Cursor Docs · cursor.com
  209. The $78,000 Agent Runaway: What OpenAI Codex's 826-Thread Explosion Reveals About Agent Cost Controls | mech.app · mech.app
  210. The Effects of Diversity in Global, Distributed Collectives: A Study of Open Source Project Success · doi.org
  211. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers · doi.org
  212. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot · arxiv.org
  213. The Mythical Man-Month · doi.org
  214. The Productivity J-Curve: How Intangibles Complement General Purpose Technologies · nber.org
  215. The Rapid Adoption of Generative AI · doi.org
  216. The Shift to Agentic AI: Evidence from Codex · arxiv.org
  217. The unequal adoption of ChatGPT exacerbates existing inequalities among workers · doi.org
  218. title: "2.18 Revenue Recognition | DART – Deloitte Accounting Research Tool" · dart.deloitte.com
  219. title: "Client Challenge" · pypi.org
  220. title: "Error: Access Forbidden" · gartner.com
  221. title: "Human Verification" · scispace.com
  222. To Copilot and Beyond: 22 AI Systems Developers Want Built · arxiv.org
  223. Token Budget Overruns: What 63 Production Incidents Reveal About Runaway Agent Costs — and How Codex CLI’s Rollout Budget Stops the Bleed · codex.danielvaughan.com
  224. Token Economics: Understanding AI API Costs · fieldguidetoai.com
  225. Unifying Large Language Models and Knowledge Graphs: A Roadmap · doi.org
  226. Usage-based charges | Cursor Docs · cursor.com
  227. Usage limit best practices | Claude Help Center · support.anthropic.com
  228. VeriGen: A Large Language Model for Verilog Code Generation · doi.org
  229. VM / Self-Hosted Installation - OpenHands Docs · docs.openhands.dev
  230. What changed with Copilot billing (legacy) - GitHub Docs · docs.github.com
  231. What Predicts Software Developers’ Productivity? · doi.org
  232. Which AI Coding Tools Do Developers Actually Use at Work? - The JetBrains Blog · blog.jetbrains.com
  233. Who is using AI to code? Global diffusion and impact of generative AI · doi.org
  234. World - OpenRouter Token Usage · en.macromicro.me
  235. Unknown author on X, 2026-10-03 · x.com
  236. Unknown author on X, 2026-10-03 · x.com
  237. Unknown author on X, 2026-10-03 · x.com
  238. Unknown author on X, 2026-10-03 · x.com
  239. Unknown author on X, 2026-10-03 · x.com
  240. Unknown author on X, 2026-10-03 · x.com
  241. Unknown author on X, 2026-10-03 · x.com
  242. Unknown author on X, 2026-10-03 · x.com
  243. Unknown author on X, 2026-10-03 · x.com
  244. Unknown author on X, 2026-10-03 · x.com
  245. Unknown author on X, 2026-10-03 · x.com
  246. Unknown author on X, 2026-10-03 · x.com
  247. Unknown author on X, 2026-10-03 · x.com
  248. Unknown author on X, 2026-10-03 · x.com
  249. Unknown author on X, 2026-10-03 · x.com
  250. @XploreFutureX on X, 2026-10-03 · x.com
  251. @0xCitadel on X, 2026-10-03 · x.com
  252. @DanKornas on X, 2026-10-03 · x.com
  253. @AverageAiBro on X, 2026-10-03 · x.com
  254. @BenCrypt_ on X, 2026-10-03 · x.com
  255. @kr8tivai on X, 2026-10-03 · x.com
  256. @WPPilotMCP on X, 2026-10-03 · x.com
  257. @0xmustaquenadim on X, 2026-10-03 · x.com
  258. @k14644188 on X, 2026-10-03 · x.com
  259. @SwissTradingPro on X, 2026-10-03 · x.com

Researched by Hyperresearch. Read it online at hyperresearch.ai/samples/ai-coding-agents; ask your own question at console.hyperresearch.ai.

Updated 2026-10-04