Anthropic cuts OpenAI's Claude access ahead of GPT-5

Top AI labs evaluating each other's models is standard practice, but Anthropic just decided OpenAI crossed the line.

According to a report from Wired, Anthropic has restricted OpenAI’s API access to its Claude models. The reason? Anthropic caught technical staff from the ChatGPT maker using their access to interact with Claude Code, the company's dedicated coding assistant, right ahead of the GPT-5 release.

Catching OpenAI in the act

Both companies usually maintain special developer access to each other's APIs to run industry-standard benchmarks and safety tests. That allows teams to see how their outputs stack up against the competition.

Anthropic says OpenAI went far beyond standard evals. Its commercial terms strictly forbid using Claude to build competing products, train rival models, or reverse engineer features.

Anthropic spokesperson Christopher Nulty didn't hold back, saying it was no surprise OpenAI's own engineers were relying on Claude Code ahead of GPT-5 because it has become the go-to tool for developers. Anthropic did clarify that OpenAI will keep access strictly for legitimate safety evaluations and benchmarks.

OpenAI pushed back immediately. Chief communications officer Hannah Wong defended the activity as an industry standard for safety and progress, adding that Anthropic's move was disappointing, especially since OpenAI's API remains completely open to them.

The high-stakes coding war

The timing here tells the real story. OpenAI is gearing up to launch GPT-5, with code generation expected to be a major highlight. Meanwhile, Claude has spent the past year becoming the undisputed favorite among software engineers.

Anthropic has been fiercely protective of that lead. Last month, it restricted Claude access for AI coding startup Windsurf following rumors that OpenAI was trying to acquire it. Anthropic Chief Science Officer Jared Kaplan openly questioned why they would sell Claude to OpenAI.

Anthropic CEO Dario Amodei also recently took shots at OpenAI leadership on a podcast, explaining his departure by emphasizing that working for people whose motivations aren't sincere means contributing to something bad. Throw in OpenAI's own previous accusations against DeepSeek for distillation, and the hypocrisy accusations are flying in every direction.

Why it matters

For everyday developers, this squabble won't break your workflow today, but it confirms where the AI frontier is actually being fought: automated coding.

The line between legitimate competitive benchmarking and using a rival's output to guide your next release is paper-thin. When models hit a plateau in general chat, elite coding execution becomes the primary metric that drives paid enterprise subscriptions.

We think Anthropic played this brilliantly from a PR and defensive perspective. Calling out OpenAI lets Anthropic showcase that even its fiercest rival relies on Claude Code internally, while simultaneously disrupting whatever last-minute baseline evals OpenAI was running for GPT-5.

However, the precedent is messy. If OpenAI retaliates by cutting off Anthropic, mutual benchmarking between frontier labs disappears behind closed walls. Once that trust completely evaporates, independent verification gets a lot harder for everyone else.