Analysis
OpenAI is negotiating a legally binding agreement with Anthropic under which each company would run safety stress tests on the other's AI models, according to The Information. As of September 21, the two labs were described as having "neared" a deal, though it remained unclear whether terms were fully finalized.
A Formal Sequel To An Informal Test
This isn't the first time OpenAI and Anthropic have opened their models to each other. In August 2025, the two labs ran an informal cross-evaluation, handing each other access to their best models to run safety evaluations on a rival's technology. That earlier round produced a genuinely useful data point: Anthropic's Claude models refused roughly 70% of uncertain or potentially harmful queries, reflecting a more conservative design philosophy, while OpenAI's o3 model matched or outperformed Claude Opus 4 on core alignment benchmarks despite refusing less often. Neither company came out of that round with a clean "safer" label -- the two labs simply optimize differently, and the exercise made that difference visible rather than resolving it.
“In August 2025, the two labs ran an informal cross-evaluation, handing each other access to their best models to run safety evaluations on a rival's technology.”
Why Make It Binding Now
Moving from an informal, one-time swap to a reportedly legally binding arrangement is a meaningfully higher level of commitment between two companies that compete directly for the same enterprise customers, the same developer mindshare and, increasingly, the same pool of late-stage capital. A binding agreement implies recurring testing cycles and some mechanism for acting on what each side finds in the other's model, rather than a single published comparison that both companies can cite selectively in their own marketing.
Part Of A Wider Industry Push
The talks aren't happening in isolation. Pulse previously covered weeks of parallel safety coordination among OpenAI, Anthropic and Google DeepMind, including a proposal -- reportedly first floated by DeepMind's Demis Hassabis, developed further by OpenAI's Chris Lehane, and endorsed by Sam Altman -- for an industry-funded, FINRA-style self-regulatory body to test powerful AI systems before release. A bilateral stress-testing pact between the two most valuable AI labs reads as a smaller, faster-moving piece of that same broader effort: frontier labs increasingly prefer negotiated, industry-run safety mechanisms they help design, over waiting for binding federal rules that don't yet exist.
What The Cooperation Doesn't Resolve
What this kind of pact doesn't address is who arbitrates a disagreement if OpenAI's stress test flags a serious issue in an Anthropic model, or vice versa -- there is no disclosed enforcement mechanism, no independent referee, and no public reporting requirement comparable to what California's SB 53 already imposes on both companies. A voluntary bilateral agreement between two competitors, however well-intentioned, is still self-policing by the companies with the most to lose from an unfavorable public result being disclosed at all.
What To Watch
The gap between "neared a deal" and an actually signed, binding agreement is exactly where past industry safety commitments -- including earlier voluntary AI pledges from 2023 -- have stalled or been quietly narrowed before signing. Whether this pact produces a published joint report, similar to the August 2025 cross-evaluation, or stays a private arrangement between the two companies, will determine whether it functions as real accountability or as a talking point for both labs' next round of safety messaging.