Anthropic cited OpenAI's use of Claude Code “ahead of the launch of GPT-5” but says it will provide API access to OpenAI for benchmarking and safety evaluations
Mayank Parmar / BleepingComputer :
Context & Ripple Effects
This is the operational expression of a long-running competitive tension: earlier coverage described how OpenAI had strengthened ChatGPT’s coding capabilities in response to Claude, while Anthropic differentiated itself around safety. The dispute turns that rivalry into a question of who may use a rival’s tools for model development.
Anthropic’s position creates a split regime: it objected to the reported use of Claude Code before GPT-5 while preserving narrowly framed access for benchmarking and safety work, following the reported revocation of OpenAI’s general Claude API access.
First-order effects
- OpenAI’s access to Claude becomes purpose-limited rather than a broadly available development input; benchmarking and safety evaluations remain permitted under Anthropic’s stated exception.
- Anthropic gains a clearer basis to distinguish sanctioned cross-model evaluation from competitive product-development use, as described in its account of the Claude Code use.
Second-order effects
- Other frontier-model providers are likely to formalize similar distinctions in API terms, access controls, and monitoring, because rival customers can also be direct competitors.
- Independent benchmarking becomes more dependent on explicitly authorized access, potentially separating safety comparison workflows from the tooling used to improve commercial models.
Third-order effects
- If such exceptions become standard, frontier-model APIs may evolve from general-purpose services into governed channels with different rights for developers, competitors, and evaluators.
- The episode points to a durable tension: cross-model testing can support safety assessment, while the same access can be treated as proprietary competitive intelligence; resolving it will increasingly depend on access-governance rules rather than technical capability alone.
The trend: Frontier AI labs are increasingly treating model access as a governed competitive asset, while preserving limited interoperability for safety and evaluation.