Alberto Angulo
ES

Companies and markets · Labs and startups

Why Labs Build Products

Model labs sell models to companies whose products compete with the labs' own. The product is where a lab's model gets paid for, where it sees what happened next, and where the customer sits. That puts anything that is mostly scaffolding around one model in the lab's path, and tells you where to build instead.

Alberto Angulo9 min read · 19 October 2026

On June 3, 2025, the coding tool Windsurf announced that Anthropic had given it under five days' notice before cutting most of its direct access to two Claude models. Windsurf was a customer. It paid Anthropic for every token its users consumed. It was also, by then, reported to be in talks to be bought by OpenAI for about $3 billion.

Two days later Anthropic's co-founder Jared Kaplan explained: "I think it would be odd for us to be selling Claude to OpenAI." A month after that, the OpenAI deal fell apart. Google hired Windsurf's chief executive and researchers, and another company, Cognition, bought the rest.

That episode is the clearest picture of a relationship anyone building on AI has to understand. The labs that make the models sell them to companies like Windsurf, and they also build products that compete with those same companies. In February 2026, Anthropic said its own coding agent, Claude Code, had run-rate revenue above $2.5 billion, meaning a recent month's revenue multiplied up to a year. That same month Cursor, one of the most prominent independent coding tools, was reported at $2 billion, running on models from several labs, Anthropic's among them. Supplier and customer, growing in the same category, at the same time.

Here's why that's happening, and what follows from it. A lab builds products because the product is where its model gets paid for: the lab can tune model and product together, it sees what happened after each answer, it pays less for a token than anyone reselling it, and it owns the customer. So anything that sits close to the model and mostly patches today's weaknesses is in the lab's path. What protects you is distance: things a model release can't ship.

What moved, layer by layer

Picture an AI product as a stack. At the bottom is the model. Above it sits the harness, the ordinary software that decides what the model sees, which tools it can use and when it stops. Above that, the app people actually open. At the top, the customer's own workflow: their data, their permissions, their budget.

WHERE MODEL LABS MOVED UP THE STACK, AND WHEN conceptual layers; launch dates from company posts and reporting (see Sources) Workflow and customer admin controls, connectors, marketplaces, deployment teams; the customer's permissions, data and process owners stay here App and product chat, documents, design, browser, mobile the lab sells to people and teams directly Harness and agent tools, loop, context, checks coding agents that act on files and systems Model (API) sold per token; list price set by the lab 2 1 3 ChatGPT Enterprise, Aug 2023 Claude Enterprise, Sep 10, 2024 DeployCo, May 2026 (reported) Gemini business agent, Oct 8, 2026 ChatGPT, Nov 30, 2022 Claude iOS app + Team, May 1, 2024 Claude Design, Apr 17, 2026 ChatGPT Work, Jul 2026 OpenAI Operator, Jan 2025 Claude Code preview, Feb 24, 2025 Codex CLI, Apr 16, 2025 Gemini CLI, Jun 25, 2025 Opus 4, May 2025: $15 in, $75 out Opus 4.5, Nov 2025: $5 in, $25 out Opus 5.5, Sep 2026: $4 in, $20 out (USD per million tokens) Numbers give the order of each layer's first lab-built product: 1 app (2022), 2 enterprise plans (2023), 3 agents (2025). The top layer kept filling in during 2026. Dates mark launches, not usage or revenue. The accent outline marks the layer where the customer's own assets sit.
Labs reached each layer above the model at different times, and the customer's own assets sit in the top layer. Conceptual four-layer stack (model, harness and agents, apps, workflow and customer), with dated examples of the labs' own products through October 2026; a selection, not a census.Conceptual illustration

The labs have climbed that stack one layer at a time. Chat apps came first: ChatGPT in November 2022, enterprise plans in 2023 and 2024. Coding agents came in 2025: Claude Code, OpenAI's Codex and Google's Gemini CLI within a few months of each other. In 2026 the labs moved into everyone else's work: Anthropic's agent for non-coding tasks and its design tool, OpenAI's job-specific plug-ins and a shared workspace with documents and slides, Google's business agent with connectors to Microsoft 365, Slack, Jira and Snowflake.

The newest layer is the budget. OpenAI and Anthropic both now run marketplaces where a business can spend part of its commitment to the lab on approved partner software. Anthropic's lists products such as Cursor, Harvey and Lovable. That makes the lab the place where the customer decides what else to buy.

Four reasons the product pays

The model and the harness are trained together

A model doesn't learn to use tools in the abstract. When a lab trains a model to act as an agent, it trains it inside a particular environment: a particular way of editing files, a particular set of tools, a particular loop. OpenAI's developer changelog lists at least six models made specifically for its Codex tool in under a year. A lab that owns both can ship them as one product and tune each against the other. A company calling the API gets the general model and has to build its harness around it.

That advantage is real, and it's easy to overstate. In April 2026 tests by the security firm Endor Labs, OpenAI's GPT-5.5 did better inside Cursor than inside OpenAI's own Codex, and Anthropic's Opus 4.7 did slightly better inside Cursor than inside Claude Code (Same Model, Different Results). Co-training gives a lab a head start in its own harness. It doesn't hand it the market.

Usage shows what happened next

The common assumption is that labs build products to harvest training data. The real picture is narrower and more interesting.

Privacy defaults limit what a lab can learn. Anthropic's consumer plans can feed training only if the user allows it, and business and API use are excluded. OpenAI's API data isn't used for training unless the customer opts in.

And the most valuable signal is the outcome, which the prompt never contains. What improves a coding model most is knowing whether the suggestion was accepted, or whether the change passed review. Cursor described in 2025 how its autocomplete model learns from exactly that: whether a user accepts or rejects each suggestion, across more than 400 million requests a day, with a new version every couple of hours. The company reported 21 percent fewer suggestions and a 28 percent higher accept rate.

So the data argument is really an argument for owning the surface where outcomes appear. An independent tool can own it as well as a lab can.

A lab that sells only through an API sees what was asked. A lab with a product sees what happened next.

A reseller pays the lab's price; the lab pays its cost

This is the mechanism that matters most for anyone building on a model, so it's worth doing the arithmetic.

Guess first

A user pays $20 a month and uses tokens that would cost $40 at the lab's list prices. A company reselling the model and the lab itself both serve that user at $20. Do both lose money, or only one?

Take a $20-a-month plan, which is where Claude Pro, Cursor Pro and ChatGPT Work all start. Suppose one user's monthly usage would cost $40 at the lab's list prices. A company reselling that model access pays the lab's price, list or negotiated, so it loses $20 on that user. Now suppose the lab's own cost to serve is half of list price; nobody publishes that number, so treat it as an assumption to show the shape. The lab serving the same user at the same $20 breaks even.

Same user, same price, one company loses money and the other doesn't. And the gap grows with usage, because the difference is a fixed share of every token. That's exactly what agents do: they use more. A TechCrunch writer's casual use of ChatGPT Work over four days consumed tokens the tool itself priced at about $65, more than three times the subscription.

A reseller pays the lab's price for a token. The lab pays something closer to its cost.

The resellers have felt it. In the summer of 2025 Cursor changed its pricing to pass the cost of Anthropic's latest model on to its heaviest users. By April 2026 it was reported to be making money on large enterprise customers and losing money on individual developers.

Meanwhile the price of the token itself keeps falling. Anthropic's top Opus tier listed at $75 per million output tokens in May 2025, $25 in November 2025 and $20 in September 2026: down about three quarters in sixteen months, for models that kept getting better. A company that only sells tokens has to keep launching a more expensive tier above the old one just to hold its revenue. A company that sells a product sets its own price. With the labs committing hundreds of billions to compute (Anthropic's prospectus was reported to plan $518 billion of spending on cloud, compute and infrastructure), they need revenue that grows with usage at a margin they control. That's an inference, but a strong one.

The product holds the customer

A model sold through an API is a line item an engineer picks. A product sold to the business is a relationship with whoever owns the budget. The marketplaces are the clearest evidence, and so is OpenAI's Sign in with ChatGPT, which lets users spend their ChatGPT allowance inside other companies' apps. When OpenAI announced it, it had announced no billing or revenue-sharing plan for those apps. A lab that is both your channel and your competitor gets the benefits of a platform before it decides what to charge you.

There's a competing explanation, and it deserves a hearing: maybe users simply want the lab's own tool because it works best with the lab's model, so products are pulled by demand rather than pushed by strategy. Both can be true at once.

What holds the labs back

None of this means the labs win everything. The limits are as visible as the advances.

They retreat. OpenAI launched Agent Builder, a visual tool for building multi-agent workflows, in October 2025 and announced it would be shut down about eight months later. A lab's attention follows whatever compounds with its models.

They can't own every profession. Companies such as Harvey in law and Clay in sales compete with model-agnostic products built around one profession's work.

Enterprises want more than one model. Google's business agent runs Claude. Anthropic's newest model is available on AWS, Google Cloud and Microsoft Azure. And integration work is slow, human and bounded by what a customer's systems expose; OpenAI's move to buy deployment capacity, reportedly including consulting-style firms, is a lab admitting where the hard part moved.

They need the builders they compete with. In October 2026 Anthropic launched a program giving startups a free year of its Team plan and API credits, saying the benefits of AI will reach most people through companies that build on top of models. OpenAI courts the same builders with apps inside ChatGPT. My reading is that labs will compete where the product is general and partner where it is specific, and that a builder should assume both at once.

Where that leaves you

The useful question is the one A Product That Improves With the Model asks: does the next model make your product stronger, or absorb it? Two dimensions make it concrete.

The first is how much of your product is scaffolding: code and prompts that exist because today's model is weak at something, such as remembering a long task, formatting output, picking the right tool or recovering from errors. Scaffolding expires with each model release. That's why a lab's next release so often ships your feature.

The second is distance from the model: how much your product owns that a model vendor can't ship. Customer data you have the right to use. Permissions inside a customer's systems. Integrations into the systems where their records live. Accountability for a signed-off result. A lab can ship a better chat interface in a quarter. It can't ship another company's access rights.

WHAT SITS IN A LAB'S PATH: A JUDGMENT MAP conceptual; positions are the author's judgment; markers show lab products in the lab's path Scaffolding share: how much exists to cover today's model weaknesses low high Distance from the model far close Prompt-library chat wrapper Generic file chat and Q&A Multi-agent frameworks General research agent General coding agent Meeting-notes summarizer Slide and design generator Vertical front end (health, finance) Deployment services for one company Evals, observability Data-access and permission layer Vertical system of record with sign-off a lab has shipped a first-party product here (as of Oct 11, 2026) no current lab product found (OpenAI is retiring its Evals platform)
Products close to the model and mostly scaffolding are where labs have already shipped; far from the model, they mostly have not. Judgment map of twelve product categories by scaffolding share and by what the product owns that a model vendor cannot ship; filled markers show a lab product as of October 11, 2026, and positions are the author's judgment, not measurements.Conceptual illustration

The categories close to the model and heavy on scaffolding all have a lab product already. Labs also buy their way across distance, with money and people, when the prize is big enough. But the further you sit from the model, the more each release works for you.

A release ships scaffolding. It does not ship another company's permissions.

What to do

Four questions turn this into a decision.

If the model were perfect tomorrow, which parts of your product would still be needed? The parts that wouldn't are scaffolding and in the lab's path. The parts that would, such as approval rules, permissions and the systems that hold the truth, are what Same Model, Different Results calls walls, and they aren't.

What do you own that effort can't buy? Time, rights and other people's agreement. A year of being trusted with a customer's data counts. So does a record of how they decide (Where the Moat Moves).

Does each model release help you or replace you? If a better model lets your product do more for the same customer, you're riding the curve. If it makes your product unnecessary, you're standing in its path.

Could you move your traffic to a second model faster than a supplier could change terms on you? Windsurf got under five days. Run that test against the notice you'd actually be given (An API Is a Strategic Dependency).

And read the labs' terms pages as carefully as their launch posts. In every case above, the changes that hurt a business were policy and capacity decisions, often with days of notice or less. The launches only told you where the lab was heading. The terms told you when it arrived.

My advice is to treat a lab as what it is: an excellent supplier with its own product plans. Use its models, use its channels when they bring you customers, and spend your engineering on the part of your product no release will ship. Where it's useful to start, What a Startup Can Still Win looks at where the room is.

Check yourself

Why can a reseller lose money on a $20 plan where the lab itself breaks even?

The reseller pays the lab's price for each token, while the lab pays closer to its cost. The gap is a fixed share of every token, so it grows as heavy users and agents use more.

Why is scaffolding in the lab's path while permissions are not?

Scaffolding patches a weakness of today's model, so the next release tends to ship the feature. A lab can ship a better chat interface in a quarter but can't ship another company's access rights.

Why does the essay say the data argument is really about owning the surface where outcomes appear?

The most valuable signal is what happened after the answer, such as whether a suggestion was accepted. An independent tool can own that surface as well as a lab can.

Notes, sources and method

For readers who want to check the numbers behind Why Labs Build Products. Every date and figure is as of 11 October 2026, when the sources were checked.

Run-rate revenue is a recent month or week multiplied up to a year. It is not audited revenue, and companies count it differently. Where a figure comes from press reports of company statements, it is marked reported.

Key numbers

Number in the articleExact value and contextSource and dateNotes
Windsurf: under five days' notice, June 2025Anthropic cut most (Windsurf: "nearly all") of its direct first-party capacity for two Claude modelsTechCrunch, 3 Jun 2025 [24]
OpenAI–Windsurf talks, about $3 billionReported April 2025; collapsed 11 Jul 2025TechCrunch [27][26]An investor at SignalFire told TechCrunch labs would be acquisitive at the app layer.
Kaplan: "I think it would be odd for us to be selling Claude to OpenAI"Also said Anthropic was short of compute and wanted to reserve it for lasting partnershipsTechCrunch, 5 Jun 2025 [25]
Google hired Windsurf's CEO and researchers; Cognition bought the restNonexclusive license, reported at $2.4 billion (Bloomberg); Cognition three days laterTechCrunch, 11 Jul 2025 [26]TechCrunch's August 2025 analysis read the sale as locking in a return before the labs could undermine it [28].
Claude Code run-rate above $2.5 billion, February 2026About 18% of Anthropic's $14 billion total on 12 Feb 2026 (2.5 / 14 = 17.9%, using the floor); $1 billion in November 2025, about six months after general availabilityAnthropic [1][2]A multiple of at least 2.5 in about three months.
Cursor $2 billion annualized, February 2026Bloomberg, via TechCrunch; passed $1 billion in November 2025; forecast above $6 billion by end of 2026TechCrunch, 17 Apr 2026 [30]
ChatGPT November 2022; enterprise plans 2023 and 2024ChatGPT 30 Nov 2022; ChatGPT Enterprise August 2023; Anthropic Team plan 1 May 2024; Claude Enterprise 10 Sep 2024[45][5][4]ChatGPT Enterprise launch month is background knowledge; its existence before Claude Enterprise is verified via a secondary source. ChatGPT was reported at about 900 million weekly users in February 2026.
Coding agents 2025Claude Code research preview 24 Feb 2025; Codex CLI 16 Apr 2025 (changelog); Gemini CLI 25 Jun 2025; OpenAI Agents SDK 11 Mar 2025[3][13][15]
2026 moves into everyone's workCowork research preview (January); Claude Design (17 Apr); six Codex plug-ins (2 Jun: analytics, creative production, sales, product design, equity investing, investment banking) and more than five million weekly Codex users, about 20% knowledge workers; ChatGPT Work (July, $20 tier); shared workspace with documents and slides (29 Sep); Google business agent with connectors to Microsoft 365, Slack, Jira and Snowflake, running Claude among its models (8 Oct)[45][40][20][21][19][17]
MarketplacesOpenAI enterprise marketplace (29 Sep 2026; spend commitment on approved partner software; more than 30 launch partners); Anthropic Claude Marketplace (limited preview; 16 partner products including Cursor, Harvey, Lovable, CodeRabbit)[18][12]
At least six Codex-specific models in under a yearcodex-mini-latest (15 May 2025), gpt-5-codex, gpt-5.1-codex (with mini), gpt-5.1-codex-max, gpt-5.2-codex, gpt-5.3-codex (24 Feb 2026)OpenAI API changelog [13]Read through a keyword filter, so there may be more.
Endor Labs: GPT-5.5 better in Cursor than Codex; Opus 4.7 slightly better in Cursor than Claude CodeApril 2026Endor Labs [46], as reviewed in A06Not re-opened for this piece.
Privacy defaultsAnthropic consumer terms from 28 Aug 2025: Free, Pro and Max chats and Claude Code sessions may be used for training if the user allows, retention up to five years if allowed and 30 days if not; Team, Enterprise, Government, Education and API use (including via Amazon Bedrock and Google Cloud) excluded. OpenAI API: no training unless the customer opts in; abuse-monitoring logs up to 30 days by default[6][14]OpenAI's consumer settings are a separate document, not covered.
Cursor autocomplete: accept or reject as reward, 400 million requests a day, new version every 1.5 to 2 hours, 21% fewer suggestions, 28% higher accept rateSeptember 2025, company-reportedCursor [35]
$20 plansClaude Pro (monthly billing), Cursor Pro (as of August 2025), ChatGPT Work (as of August 2026)[21][28]
ChatGPT Work: about $65 in four daysOver 80 million tokens, priced by the tool's own estimate; more than three times the subscriptionTechCrunch, 24 Aug 2026 [21]
Cursor passed the cost of Anthropic's latest model to its most active users (summer 2025); positive gross margins on large enterprises, losing money on individual developers (April 2026)TechCrunch [28][30]A source close to Windsurf called its gross margins very negative (August 2025, anonymous).
Opus tier output price: $75 → $25 → $20Opus 4 (May 2025) $15 / $75; Opus 4.5 (24 Nov 2025) $5 / $25; Opus 5.5 (22 Sep 2026) $4 / $20. Output down 66.7% in six months, 73.3% over sixteen ("about three quarters")Anthropic [8][9][10]Different models of rising capability: a pricing slot, not equal quality at a lower price. Claude models from 4.7 on use a tokenizer that produces about 30% more tokens for the same text, so the per-word fall is smaller. Above the slot, Fable 5.1 lists at $10 / $50, the same as OpenAI's GPT-6 Astra.
Anthropic prospectus: $518 billion on cloud, compute and infrastructureOver the coming yearsReuters via TechCrunch, 28 Sep 2026 [23]OpenAI raised $122 billion in a single round in March 2026 [22].
Sign in with ChatGPT; no billing or revenue-sharing plan announced16 launch partners including Notion and VercelTechCrunch, 29 Sep 2026 [18]
Agent Builder: launched 6 Oct 2025, deprecation announced 3 Jun 2026240 days, about eight monthsOpenAI API changelog [13]
Harvey (law) and Clay (sales)Compete with model-agnostic products built around one profession's workTechCrunch, 24 Aug 2026 [21]
Opus 5.5 on AWS, Google Cloud and Microsoft AzureAnthropic [9]
OpenAI deployment companyAnnounced May 2026, reported at more than $4 billion of investment; has reportedly acquired consulting-style firms; OpenAI's chief revenue officer said in June the challenge was now helping companies integrate these systems[45] (Axios and Reuters headlines)Reported.
Anthropic startup program, 6 Oct 2026Free year of Team plan for up to five seats, $1,000 in API credits, marketplace access; stated reason: the benefits of AI will reach most people through companies that build on top of modelsTechCrunch [41]OpenAI courts the same builders with apps inside ChatGPT [42].

The reseller squeeze, worked

Take a $20 per month plan. Let U be what one user's monthly usage would cost at the lab's list API prices. A reseller's margin before any other cost is 20 − U. Suppose, as an assumption and not a reported number, that the lab's own cost to serve is half of list price. The lab's margin is then 20 − 0.5U.

Usage at list API pricesReseller marginLab margin (cost at 50% of list)
$10+$10+$15
$20$0+$10
$40−$20$0
$80−$60−$20

Between $20 and $40 of usage, the lab's product makes money and the reseller's does not, at the same price to the user. The gap is 0.5U, so it widens as usage grows. The 50% figure is invented to show the shape. No lab publishes its cost to serve; labs also pay for reserved capacity and cap usage with limits, and a large reseller can negotiate below list, which narrows the gap without closing it.

Method

  • Claude Code share of Anthropic run-rate: 2.5 / 14 = 17.9%, using the stated floor of "over" $2.5 billion.
  • Claude Code growth: $1 billion (November 2025) to over $2.5 billion (12 Feb 2026), a multiple of at least 2.5.
  • Opus output price: $75 to $25 is a 66.7% fall; $25 to $20 is 20%; $75 to $20 is 73.3%.
  • Agent Builder, 6 Oct 2025 to 3 Jun 2026: 240 days.
  • Anthropic Q2 2026 revenue of $11.5 billion times four: $46 billion.
  • GPT-6.1 Sol against GPT-6 Astra list prices: input $2 against $10, output $10 against $50, each a ratio of 0.2.
  • Teams bundling case cited in A22: Slack's complaint in 2020 to the Commission's acceptance of commitments in September 2025, about five years.

Figures. The stack and the distance map are conceptual; the map's placements are the author's judgment. The distance map's lower-right zone holds seven categories close to the model and heavy on scaffolding, each with a lab product; two categories above the zone also have lab entries (front ends for health and finance, and deployment services for a single company). Hollow markers mean no current lab product in the sources checked (OpenAI announced the retirement of its Evals platform in June 2026), not that a lab won't ship one. A dated timeline of twenty-five events (a76-timeline) was dropped from the article in the 19 Oct 2026 rewrite.

Several primary pages could not be opened (OpenAI's own site returned an access error for automated fetches, and some news sites were not retrievable), so facts about OpenAI's products rely on OpenAI's developer documentation and on press coverage.

Detail left out of the article

  • Acquisitions. Anthropic bought Bun, a JavaScript runtime, on 3 Dec 2025 to speed up Claude Code, and Stainless, which generates SDKs from API specifications, on 18 May 2026 (it will wind down hosted products). OpenAI's reported purchases include Statsig and Astral. SpaceX, which had merged with xAI earlier in 2026, closed its $60 billion purchase of Cursor in August 2026; in April TechCrunch reported Cursor had been about to raise $2 billion at a $50 billion valuation when SpaceX offered an option to buy it, a person familiar with the matter said that round would have fallen short of cash-flow breakeven, and TechCrunch noted competition from Claude Code and Codex could make later fundraising harder [31][32].
  • Cursor's own model. Cursor announced Composer on 29 Oct 2025; the second version was built on an open-weights model from Moonshot AI, as Cursor disclosed in March 2026 [35][34]. Its CEO argued in December 2025 that the labs offered an engine with a concept car around it, against Cursor's finished car [29].
  • Cognition raised at a $48 billion valuation in September 2026 on about $900 million in run-rate revenue, still using third-party models from OpenAI and Anthropic while training its own [33].
  • Clio. In December 2024 Anthropic described a tool that groups conversations into topic clusters and shows analysts only anonymized, aggregated clusters above a minimum size; over 10% of 1 million analyzed conversations involved web and mobile app development [7]. Whether labs use aggregate product signals to choose which categories to enter is a hypothesis; OpenAI says it guides ChatGPT Work with an evaluation of 44 occupations plus user feedback (TechCrunch, August 2026), which is consistent with it [21].
  • Company-wide revenue. A CNBC report in August 2026, linked by TechCrunch, put Anthropic's run-rate at about $65 billion; OpenAI reportedly told investors its annualized revenue was approaching $50 billion after a figure near $70 billion circulated. Anthropic counts sales through cloud partners and OpenAI does not, so the two are not comparable. The Financial Times put Anthropic's Q2 revenue at $11.5 billion ($46 billion annualized), below the reported run-rate [22][23].
  • Concentration. The Financial Times reported that nearly a quarter of Anthropic's 2025 revenue came from two unnamed customers; OpenAI has made no comparable disclosure. A lab with that concentration has a reason to want direct relationships with end customers; this is a hypothesis about incentives, untested while the customers are unnamed [23].
  • Convergence. In September 2026 Anthropic released Opus 5.5 with what TechCrunch called Fable-level performance at lower prices, and OpenAI released GPT-6.1 Sol at $2 and $10, a fifth of GPT-6 Astra's list price, saying it nearly matched Astra [43][47]. If models converge, products create habits and integrations that are harder to copy than a benchmark score.
  • Third-party harnesses. On 4 Apr 2026 Anthropic told Claude Code subscribers they could no longer use subscription limits with third-party harnesses such as OpenClaw and would pay separately, metered, effective the day of the report. Claude Code's head, Boris Cherny, said subscriptions were not built for those usage patterns and full refunds remained available. OpenClaw's creator, reported to be joining OpenAI, disputed that and accused Anthropic of copying popular features before locking out open source [37]. Anthropic's legal page tells developers building products on Claude to use API keys and does not permit routing requests through consumer plan credentials [11].
  • Code review. Anthropic's Code Review (9 Mar 2026), estimated by an Anthropic product lead at an average of $15 to $25 per review, entered a category while the same lab's marketplace lists CodeRabbit, a code-review company [39][12].
  • Regulation. No regulator action aimed at labs competing with their own customers turned up in the sources. The nearest official document is the FTC staff report of 17 Jan 2025 on cloud and AI partnerships, which flagged switching costs, lock-in and cloud partners' access to sensitive information such as customer usage and revenue [44]. Bundling cases against software suites are a closer precedent and took about five years.
  • Six kinds of platform exposure in the cases: access (capacity cut with under five days' notice), terms (a policy change reported the morning it took effect), overlap (the lab ships your feature), retirement (the lab drops what you adopted), neutrality (a vendor you used is bought by a lab or its parent: Stainless, Cursor), channel (the lab sets price anchors and controls discovery). Not all dependence is bad; a lab's channel can bring customers faster than a builder could alone.
  • Reading a lab announcement: which layer it enters; what it left unsaid (for example, billing terms); whether it is a commitment (previews can be retired; commitment shows in billing and admin controls).

Caveats and open questions

  • Co-training. Independent evaluations that hold the model fixed and show lab harnesses winning consistently would change the reading that co-training gives a head start without locking the market.
  • Pull versus push. The strongest competing explanation is demand: users want the lab's own tool because it works best with the lab's model. A falling share of product revenue alongside falling token prices would favor it. Not resolved by data.
  • Capital and pricing. That compute commitments push labs toward products with margins they set is an inference, not something either company says.
  • Unverified or open: ChatGPT Enterprise's exact August 2023 launch date; Anthropic's $65 billion run-rate (CNBC page not opened); a reported ending of Cursor's OpenAI API access after the SpaceX deal (unverified, left out); Codex CLI date versus the cloud Codex agent (May 2025).
  • Related reading: Capability, Product and Business Model, What a New Release Actually Changes, Who Should Build It, Where the AI Money Goes, Data Access Is the Real Integration Project.

Sources

  1. Anthropic. Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation. 12 Feb 2026. https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation
  2. Anthropic. Anthropic acquires Bun as Claude Code reaches $1B milestone. 3 Dec 2025. https://www.anthropic.com/news/anthropic-acquires-bun-as-claude-code-reaches-usd1b-milestone
  3. Anthropic. Claude 3.7 Sonnet and Claude Code. 24 Feb 2025. https://www.anthropic.com/news/claude-3-7-sonnet
  4. Anthropic. Claude for Enterprise. 10 Sep 2024. https://claude.com/blog/claude-for-enterprise
  5. Anthropic. Team plan and iOS app. 1 May 2024. https://claude.com/blog/team-plan-and-ios
  6. Anthropic. Updates to our consumer terms. 28 Aug 2025. https://www.anthropic.com/news/updates-to-our-consumer-terms
  7. Anthropic. Clio: privacy-preserving insights into real-world AI use. 12 Dec 2024. https://www.anthropic.com/research/clio
  8. Anthropic. Introducing Claude Opus 4.5. 24 Nov 2025. https://www.anthropic.com/news/claude-opus-4-5
  9. Anthropic. Introducing Claude Opus 5.5. 22 Sep 2026. https://www.anthropic.com/claude-opus-5-5
  10. Anthropic. Pricing. Fetched 11 Oct 2026. https://platform.claude.com/docs/en/about-claude/pricing
  11. Anthropic. Claude Code legal and compliance. Fetched 11 Oct 2026. https://code.claude.com/docs/en/legal-and-compliance
  12. Anthropic. Claude Marketplace. Fetched 11 Oct 2026. https://claude.com/platform/marketplace
  13. OpenAI. API changelog. Fetched 11 Oct 2026. https://developers.openai.com/api/docs/changelog
  14. OpenAI. Data controls in the API. Fetched 11 Oct 2026. https://developers.openai.com/api/docs/guides/your-data
  15. Google. Introducing Gemini CLI. 25 Jun 2025. https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/
  16. Google Cloud. Introducing Gemini Enterprise. 10 Oct 2025. https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise
  17. TechCrunch. Google brings agentic AI to Gemini, starting with businesses. 8 Oct 2026. https://techcrunch.com/2026/10/08/google-brings-agentic-ai-to-gemini-starting-with-businesses/
  18. TechCrunch. OpenAI's latest features take direct aim at the app store model. 29 Sep 2026. https://techcrunch.com/2026/09/29/openais-latest-features-take-direct-aim-at-the-app-store-model/
  19. TechCrunch. OpenAI takes on Microsoft with the launch of what feels a whole lot like ChatGPT's own office suite. 29 Sep 2026. https://techcrunch.com/2026/09/29/openai-takes-on-microsoft-with-the-launch-of-what-feels-a-whole-lot-like-chatgpts-own-office-suite/
  20. TechCrunch. OpenAI launches new Codex tools for white-collar work. 2 Jun 2026. https://techcrunch.com/2026/06/02/openai-launches-new-codex-tools-for-white-collar-work/
  21. TechCrunch. OpenAI is building an AI agent for everything. Will everyone use them? 24 Aug 2026. https://techcrunch.com/2026/08/24/openai-is-building-an-ai-agent-for-everything-will-everyone-use-them/
  22. TechCrunch. OpenAI's revenue is reportedly $20 billion less than previously projected. 8 Oct 2026. https://techcrunch.com/2026/10/08/openais-revenue-is-reportedly-20-billion-less-than-previously-projected/
  23. TechCrunch. Anthropic's prospectus details losses, growth, and, yes, a warning that its AI could end humanity. 28 Sep 2026. https://techcrunch.com/2026/09/28/anthropics-prospectus-details-losses-growth-and-yes-a-warning-that-its-ai-could-end-humanity/
  24. TechCrunch. Windsurf says Anthropic is limiting its direct access to Claude AI models. 3 Jun 2025. https://techcrunch.com/2025/06/03/windsurf-says-anthropic-is-limiting-its-direct-access-to-claude-ai-models/
  25. TechCrunch. Anthropic co-founder on cutting access to Windsurf. 5 Jun 2025. https://techcrunch.com/2025/06/05/anthropic-co-founder-on-cutting-access-to-windsurf-it-would-be-odd-for-us-to-sell-claude-to-openai/
  26. TechCrunch. Windsurf's CEO goes to Google; OpenAI's acquisition falls apart. 11 Jul 2025. https://techcrunch.com/2025/07/11/windsurfs-ceo-goes-to-google-openais-acquisition-falls-apart/
  27. TechCrunch. Why OpenAI wanted to buy Cursor but opted for the fast-growing Windsurf. 22 Apr 2025. https://techcrunch.com/2025/04/22/why-openai-wanted-to-buy-cursor-but-opted-for-the-fast-growing-windsurf/
  28. TechCrunch. The high costs and thin margins threatening AI coding startups. 7 Aug 2025. https://techcrunch.com/2025/08/07/the-high-costs-and-thin-margins-threatening-ai-coding-startups/
  29. TechCrunch. Why Cursor's CEO believes OpenAI, Anthropic competition won't crush his startup. 9 Dec 2025. https://techcrunch.com/2025/12/09/why-cursors-ceo-believes-openai-anthropic-competition-wont-crush-his-startup/
  30. TechCrunch. Sources: Cursor in talks to raise $2B+ at $50B valuation as enterprise growth surges. 17 Apr 2026. https://techcrunch.com/2026/04/17/sources-cursor-in-talks-to-raise-2b-at-50b-valuation-as-enterprise-growth-surges/
  31. TechCrunch. How SpaceX preempted a $2B fundraise with a $60B buyout offer. 22 Apr 2026. https://techcrunch.com/2026/04/22/how-spacex-preempted-a-2b-fundraise-with-a-60b-buyout-offer/
  32. TechCrunch. SpaceX officially closes its Cursor acquisition. 15 Aug 2026. https://techcrunch.com/2026/08/15/spacex-officially-closes-its-cursor-acquisition/
  33. TechCrunch. Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market. 8 Sep 2026. https://techcrunch.com/2026/09/08/cognition-hits-48b-valuation-signaling-investors-believe-ai-coding-is-far-from-a-winner-take-all-market/
  34. TechCrunch. Cursor admits its new coding model was built on top of Moonshot AI's Kimi. 22 Mar 2026. https://techcrunch.com/2026/03/22/cursor-admits-its-new-coding-model-was-built-on-top-of-moonshot-ais-kimi/ (headline read, body not opened).
  35. Cursor. Cursor 2.0 and Composer. 29 Oct 2025. https://cursor.com/blog/2-0 ; Online RL for Cursor Tab. 12 Sep 2025. https://cursor.com/blog/tab-rl
  36. TechCrunch. Anthropic bundles Claude Code into enterprise plans. 20 Aug 2025. https://techcrunch.com/2025/08/20/anthropic-bundles-claude-code-into-enterprise-plans/
  37. TechCrunch. Anthropic says Claude Code subscribers will need to pay extra for OpenClaw usage. 4 Apr 2026. https://techcrunch.com/2026/04/04/anthropic-says-claude-code-subscribers-will-need-to-pay-extra-for-openclaw-support/
  38. TechCrunch. Anthropic has acquired the dev tools startup used by OpenAI, Google and Cloudflare. 18 May 2026. https://techcrunch.com/2026/05/18/anthropic-has-acquired-the-dev-tools-startup-used-by-openai-google-and-cloudflare/
  39. TechCrunch. Anthropic launches code review tool to check flood of AI-generated code. 9 Mar 2026. https://techcrunch.com/2026/03/09/anthropic-launches-code-review-tool-to-check-flood-of-ai-generated-code/
  40. TechCrunch. Anthropic launches Claude Design. 17 Apr 2026. https://techcrunch.com/2026/04/17/anthropic-launches-claude-design-a-new-product-for-creating-quick-visuals/
  41. TechCrunch. Anthropic is giving startups a free year of Claude Team and $1,000 in credits. 6 Oct 2026. https://techcrunch.com/2026/10/06/anthropic-gives-startups-a-free-year-of-enterprise-service-and-1000-in-token-credits/
  42. TechCrunch. OpenAI launches apps inside of ChatGPT. 6 Oct 2025. https://techcrunch.com/2025/10/06/openai-launches-apps-inside-of-chatgpt/
  43. TechCrunch. OpenAI launches GPT-6.1 Sol (29 Sep 2026) and Anthropic releases Opus 5.5 with lower prices and Fable-level performance (22 Sep 2026). https://techcrunch.com/tag/openai/ and https://techcrunch.com/tag/anthropic/ (headlines only).
  44. Federal Trade Commission. FTC staff report on AI partnerships and investments. 17 Jan 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study
  45. Wikipedia. ChatGPT, OpenAI and Anthropic articles, fetched 11 Oct 2026. https://en.wikipedia.org/wiki/ChatGPT (pointers to dates, not proof; items resting only on these are marked reported).
  46. Endor Labs. GPT-5.5 sets a new code security record, with Cursor, not Codex, in Agent Security League. 27 Apr 2026. https://www.endorlabs.com/learn/gpt-5-5-sets-a-new-code-security-record-with-cursor-not-codex-in-agent-security-league
  47. OpenAI. API pricing. Fetched 10 Oct 2026. https://developers.openai.com/api/docs/pricing

All writing