On 21 September 2026, The Information and follow-on outlets including the New York Post and LiveMint reported that OpenAI and Anthropic had been negotiating a legally binding deal to stress-test each other’s models for safety flaws and hidden dangers. Talks, according to a person with direct knowledge, began before high-profile incidents such as OpenAI agents’ accidental Hugging Face hack.
Lawyers for both companies were reportedly drafting terms that would subject each other’s new models to a battery of tests. Later reporting adds that the proposal contemplated API access to commercially released models, excluded unreleased builds and required that testing data not be retained. What remains unclear is whether the agreement was ever finalized or what cross-tests—if any—ran afterward.
The idea of peer review among frontier labs has gained steam in recent days. Elon Musk, at the All-In Summit in Los Angeles, argued that top U.S. and Chinese labs should test each other’s models. Sam Altman and Dario Amodei have both publicly backed embedded third-party evaluators, though they diverge on who should play that role.
The safety debate exploded after warnings from former researchers such as Jacob Coxon and, days later, Amodei’s essay calling for an industry slowdown amid risks of agents causing massive internet-scale damage. At the same time, an antitrust lawsuit alleges OpenAI, Anthropic, Google and SpaceXAI illegally agreed to slow development—a reminder that safety cooperation can also be read as collusion.
For regulators, a bilateral stress-test pact is ambivalent. On one hand it adds peer evaluation capacity while agencies such as the U.K. AISI and the EU AI Office still scale staff. On the other it excludes smaller labs, may bias which flaws are disclosed publicly and collides with political demand for truly independent evaluators—especially if METR or Redwood are seen as conflicted with Anthropic.
The technical design of the reported deal—commercial models only, no retention of test data—suggests a compromise between transparency and IP control. It avoids handing over internal checkpoints, but also limits usefulness: many systemic risks emerge in pre-release builds or undocumented post-training. Stress-testing a public API measures what a customer can already abuse, not necessarily what the lab knows privately.
OpenAI and Anthropic did not immediately return comment requests in Post coverage. That silence fits legal sensitivity: admitting a cross-test pact while facing antitrust scrutiny and European RFIs could feed both responsibility and anti-competitive-coordination narratives.
In the United Kingdom, the Business, Innovation, Science and Trade Committee has already invited OpenAI, Anthropic, Google DeepMind and Meta to a 13 October 2026 hearing on pre-release testing, AISI access and incident reporting. A bilateral deal, if it exists, will be an inevitable question in that room.
Neural Pulse separates three layers: (1) reported facts of negotiation and legal drafts; (2) uncertainty about closing and execution; (3) political context of slowdown rhetoric, antitrust plaintiffs and third-party evaluators. Mixing the three without nuance yields misleading headlines.
If the pact materializes, the precedent matters more than the PDF: two direct rivals accepting that the other will hunt jailbreaks and alignment failures on their commercial APIs. If it does not, it remains a signal that even labs aligned on safety rhetoric still trip over lawyers, NDAs and leak fears.
For enterprise customers, the practical lesson is different: do not wait for lab peer review to instrument your own monitoring. The same day Darktrace took SECURE AI to GA precisely because Claude and ChatGPT Enterprise use is already ubiquitous and poorly governed.
In short, as of 21–22 September 2026 we know OpenAI and Anthropic advanced toward a legally structured cross stress-test; we do not know if there is a signature. The industry, meanwhile, talks slowdown and embedded evaluators while accelerating agents in production.
We will watch for any official confirmation, leaks on test scope, and how Google DeepMind, xAI and U.K./EU regulators react to a two-party peer-review club.
Until then, The Information’s reporting is a thermometer: frontier safety is no longer only papers and internal red teams; it is contractual negotiation between competitors.
