SITUATION REPORT: The two biggest safety-focused rivals in AI are close to grading each other’s homework.
OpenAI and Anthropic neared a deal to stress test each other’s AI models, according to The Information. The two companies compete directly for enterprise customers, developers and top researchers. Now they’re reportedly working out a formal arrangement that would let each side attack the other’s systems to find weaknesses. The Information’s headline is the only part of the report we’ve seen, so the deal’s terms, timing and scope aren’t public yet.
This matters because frontier labs almost never let a direct competitor look under the hood.
🎯 Threat Assessment: Why This Deal Matters
Stress testing, often called red teaming, means deliberately trying to make a model misbehave. Testers push it to leak dangerous information, ignore its safety rules, deceive users or take harmful actions when it runs as an agent.
Every major lab already does this internally, and most also hire outside contractors. The problem is blind spots. A team that built a model tends to test for the failures it already expects. A rival lab, with different training methods and different instincts, is more likely to find the failures nobody planned for.
Three reasons this stands out:
- Competitors make the harshest testers. OpenAI’s researchers have every incentive to find real flaws in Claude, and Anthropic’s have the same motive with GPT models. That adversarial energy is useful.
- It sets an industry precedent. If the two leading labs make cross-testing routine, Google DeepMind, xAI and Meta will face pressure to join or explain why they won’t.
- It gets ahead of regulators. Governments in the US, UK and EU keep asking for independent model evaluations. A lab-to-lab arrangement shows the industry can police itself, at least partly.
📜 Background: This Isn’t Their First Round
The two companies have tried this before. In August 2025, OpenAI and Anthropic published results from a joint exercise where each ran its internal safety and alignment evaluations on the other’s publicly available models. They looked at issues like sycophancy, misuse and how well models followed instructions under pressure. Both sides presented it as a pilot.
The relationship has also been tense. Around the same time, Anthropic cut off OpenAI’s access to Claude through its API, saying OpenAI had violated its terms of service. That’s the backdrop. These labs don’t trust each other easily, which makes a formal testing deal more notable.
What’s reportedly on the table now looks like a step past that one-off pilot and toward a standing agreement. Whether it covers unreleased models, which would be the real prize for safety, is one of the key open questions.
⚠️ Open Questions to Watch
The full reporting sits behind The Information’s paywall, and these are the details that will determine how meaningful this deal is:
- Pre-release access. Testing public models is helpful. Testing models before launch is where cross-lab review could actually stop a problem from reaching users.
- Disclosure rules. Will findings be published, shared privately or kept confidential? Transparency decides whether the public benefits or just the two companies.
- Protecting trade secrets. Each lab will want to stop the other from learning training tricks through the testing process. Expect strict limits on access.
- Who else joins. A two-party deal is a start. A broader framework involving government safety institutes would carry much more weight.
🧭 What This Means for Practitioners
If you build on GPT or Claude models, this is good news. More eyes on failure modes means fewer surprises in production, especially for agent workflows where a model takes real actions.
There’s also a practical lesson for your own AI stack. The labs are admitting that self-testing isn’t enough, so don’t rely on it either. If your team ships AI features, get someone who didn’t build them to try and break them. Swap reviewers between teams, or bring in an outside tester. The same logic that applies to OpenAI and Anthropic applies to your chatbot.
🔭 Outlook
If the deal closes, expect cross-lab testing to become a selling point in enterprise deals and a talking point in policy debates. Buyers will start asking which labs have had their models checked by a rival, and the ones that haven’t will have to answer for it.
For full details on the deal, see the original report from The Information.