The Frontier Lab Non-Compete

Anthropic has revoked OpenAI's access to the Claude API. The reported trigger: OpenAI's technical staff were using Claude Code, Anthropic's coding agent product, internally, in the run-up to GPT-5. Anthropic says this violates its commercial terms, which prohibit using its services to build a competing product or to train competing models. OpenAI says the decision is disappointing but that it respects Anthropic's right to make it. Anthropic adds that it will preserve OpenAI's access for safety benchmarking and evaluations, which is the polite version of we're cutting you off from the product but not from the bit that protects us reputationally.

This is a small story that is also a structurally interesting one, and I want to walk through why.

What the terms actually say. Anthropic's commercial terms of service include a use restriction that bars customers from using the service to "build a competing product or service, including to train competing AI models" or to "reverse engineer or duplicate" the services.¹ This is standard frontier-lab boilerplate, OpenAI's usage policies contain a near-identical clause, and so do Google's. Every major model provider has written into their TOS that you cannot use their model to build a model that competes with theirs. The clause is universally present and, until now, has been used mainly against small players doing obvious distillation.

What's novel here is the enforcement vector: one frontier lab invoking the clause against another frontier lab, for what appears to be product-development use rather than weight-extraction. OpenAI engineers using Claude Code to write code, in the ordinary way that engineers everywhere are using Claude Code to write code, is, under a strict reading of the clause, a competing-product use, because OpenAI's product competes with Anthropic's. Under a loose reading it's just two companies using each other's developer tools, which they have been doing quietly for years.

What this is a case of. The frame is model weight lineage, the proposition that the defensible IP at frontier labs sits in the weights and in the data and methods that produced them, and that the legal architecture around weights is being constructed in real time, mostly through contracts rather than statutes. There is no developed body of law on what it means to "train a competing model" using another model's outputs. There is no clean doctrine on whether using a coding agent to write the training infrastructure for your own model counts. The TOS clause is doing the work that, in older industries, would be done by patents, trade secrets, or non-competes.

The defensible perimeter at the frontier is being drawn by terms-of-service clauses, enforced selectively, between four or five companies that all use each other's products.

That perimeter is porous and discretionary. Anthropic chose to enforce now, against OpenAI, in a particular week. It did not enforce against OpenAI six months ago, when presumably the same engineers were using the same tools. The decision is a signal as much as a sanction.

The signal. Two readings, not mutually exclusive.

The first is ai safety as market position. Anthropic's preserved carve-out for "safety benchmarking and evaluations" is not incidental. It lets Anthropic say, accurately, that it has not cut off the safety-relevant access; it has cut off the commercial access. This maps onto Anthropic's broader posture, safety as the thing that is non-negotiable, commerce as the thing that is. The shape of the announcement is itself a market move: we are the lab that distinguishes between these categories, and we will enforce the distinction even at cost to a customer relationship. It plays well to the enterprise and government buyers Anthropic is courting, who would like to believe their model vendor has a spine on principle.

The second reading is more prosaic: GPT-5 is imminent, Anthropic's coding product is one of its strongest revenue lines, and letting OpenAI's launch team use Claude Code to ship a competitor is a bad commercial trade regardless of TOS interpretation. This is the inference economics reading. Claude Code is, by available indicators, a meaningful contributor to Anthropic's run-rate, the company has disclosed Claude Code revenue running at roughly $400m annualised as of mid-year, growing fast.² Code generation is the highest-margin, highest-retention use case in the current market. You do not let your largest competitor use your highest-margin product to ship against you, if you can find a clean reason not to.

Both readings are probably true. Anthropic gets to enforce a principle and protect a margin in the same gesture, which is the kind of alignment of incentives that makes a decision easy.

What OpenAI's response tells you. OpenAI's statement, disappointing, but we respect their right, is the only response available. Escalating would invite scrutiny of whether OpenAI's own TOS is being applied symmetrically (it is not; Anthropic engineers, presumably, can still use ChatGPT). Litigating would surface the question of what "competing model" means under the clause, which is a question no frontier lab wants a court to answer, because the answer might be expansive enough to capture half the industry's workflows.

So OpenAI eats it, switches its engineers to internal tooling or to one of the open-weight coding models, and notes the precedent.

What to watch. Three things.

First, whether other frontier labs follow. If Google or xAI invokes similar clauses against OpenAI or Anthropic in the next quarter, the enforcement pattern hardens into a norm and the frontier becomes a more legibly closed system, five labs that do not use each other's products, with the smaller ecosystem using all of them. If no one follows, this reads as an Anthropic-specific move tied to Claude Code's commercial importance.

Second, the carve-out language. Anthropic preserved access for "benchmarking and safety evaluations." That phrase will get tested. Does it cover red-teaming? Capability evaluations that double as competitive intelligence? The carve-out is the seam where the next dispute happens.

Third, whether this accelerates the open-weights substitution at the frontier labs themselves. If you are an OpenAI engineer who can no longer use Claude Code, your fallback is your own internal tooling or an open-weight coding model, Qwen, DeepSeek, one of the Llama derivatives. The frontier labs cutting each other off is, marginally, a tailwind for the open-weight coding stack, because it removes the most capable proprietary substitute from the menu for exactly the customers most able to evaluate alternatives.

A small story, then. But the shape of it, discretionary enforcement of a universal clause, between two companies that need each other to maintain the legitimacy of the clause for everyone else, is the shape of how the frontier is going to govern itself for a while.

Footnotes

¶