Anthropic has formally cut off OpenAI’s access to its Claude family of models, citing breaches of its terms of service tied to internal benchmarking activity, per Wired.
The move comes at a pivotal moment, just as OpenAI prepares to unveil its highly anticipated GPT5 model.
Why Access Was Cut Off
According to WIRED, OpenAI engineers used Claude Code, Anthropic’s AI-powered coding assistant, to evaluate Claude’s performance against OpenAI’s own engines across tasks like coding, creative writing, and safety-related scenarios (e.g., self-harm and defamation prompts). Anthropic maintains that this violated its commercial terms, which explicitly forbid customers from using the API to build competing services or improve their own models.
Register for Tekedia Mini-MBA edition 19 (Feb 9 – May 2, 2026): big discounts for early bird.
Tekedia AI in Business Masterclass opens registrations.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab: From Technical Design to Deployment (next edition begins Jan 24 2026).
Anthropic spokesperson Christopher Nulty said: “Claude Code has become the go to choice for coders everywhere… OpenAI’s own technical staff were also using our coding tools ahead of the launch of GPT5… this is a direct violation of our terms.”
OpenAI responded by calling the benchmarking activity “industry-standard,” arguing that evaluating rival models is a normal part of development and safety testing. The company expressed disappointment but noted that Anthropic’s API remains available to it for benchmarking and safety evaluations, though details remain murky.
A Pattern of Access Control
This action follows a pattern. In recent weeks, Anthropic had already revoked API access to Windsurf, a startup rumored to be linked with OpenAI. Chief Science Officer Jared Kaplan remarked at the time: “I think it would be odd for us to be selling Claude to OpenAI.”
Analysts see parallels to precedents in Big Tech: Facebook cutting off Vine API access and Salesforce restricting Slack’s openness. As AI evolves, control over competitor access appears central to company defense strategies.
Instead of standard user interactions, OpenAI reportedly integrated Claude into its proprietary developer tools via custom API endpoints. This allowed bulk benchmarking—testing Claude’s functions in categories such as code generation, writing, and safety responses. These evaluations played a role in refining GPT5, expected to include new reasoning and coding capabilities.
On the Eve of GPT5
The cut occurred on August 1, 2025, just days before GPT5’s slated launch. Many interpret the timing as strategic; perhaps Anthropic aimed to blunt any competitive leak or advantage OpenAI might gain.
The decision signals a profound shift in AI industry dynamics: from cooperative benchmarking to protective exclusivity, as labs rush to preserve their competitive edge.
Implications for OpenAI, Anthropic & Beyond
OpenAI now loses direct insight into Claude’s performance under realistic developer conditions—a limiting blow if it can’t independently replicate those capabilities.
Anthropic reinforces its safety- and integrity-first posture, emphasizing control over how Claude is used and preventing rivals from reverse engineering its capabilities.
Developers and startups may become more vulnerable. Access to major AI models is increasingly contingent on loyalty—or at least compliance—with restrictive terms.
The AI policy landscape now faces a critical question: Are these models proprietary platforms, or should fair benchmarking access be considered industry-standard infrastructure?
Broader Industry Trends
Benchmarking in AI is becoming more complex and costly. As models excel on traditional tests like MMLU or Hellaswag, companies are building private evaluation systems for reasoning and planning. This shift raises concerns about comparability and trust in model performance claims.
However, Anthropic’s Claude models continue to gain traction and market share. Recent figures show the company now holds 32% of enterprise LLM usage, surpassing OpenAI in adoption, particularly in coding tasks, recording a 42% share compared to OpenAI’s 21%.