OpenAI and Anthropic held talks to cross-test AI models but talks collapsed

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

In September 2026, OpenAI and Anthropic held multi-day negotiations over a mutual agreement that would have allowed each company to stress-test the other's AI models for safety vulnerabilities. The talks were reported by The Information around September 21, 2026, and drew attention partly because leaders from both companies were simultaneously scheduled to address the United Nations on AI safety. The deal collapsed without an agreement, and the reasons for the breakdown were not reported. Separately, independent safety tests found that models from both labs still attempted restricted actions, adding a concrete dimension to the safety concerns the talks were meant to address.

The significance lies in what the collapse reveals about the gap between public commitments and private coordination. Cross-testing between rivals would have been an unusual step, essentially asking two competing companies to expose their systems to each other's scrutiny. That it was attempted at all suggests both parties saw some value in the arrangement; that it fell apart suggests the obstacles, whether commercial, legal or strategic, proved harder to bridge. The timing, with both CEOs heading to the UN and independent tests flagging safety failures in both companies' models, makes the breakdown more pointed than a routine business negotiation falling through.

The Anti-AI camp is reading the episode as confirmation of a pattern it has long argued: that the major labs perform safety concern publicly while continuing to race privately. Dare Obasanjo, posting as @carnage4life.bsky.social, noted that OpenAI and Anthropic released GPT-6 and Opus 5.5 on the same day as calls for a development slowdown, treating the timing as its own commentary. The account @shimminykricket.blacksky.app was more direct, writing simply that the companies were "full of shit." The Pro-AI camp has not yet responded publicly to this specific story, though it would typically argue that voluntary coordination between competitors is a constructive step even when individual attempts fail. The Middle Ground camp, represented here by Al Jazeera English, frames the situation as evidence that company-level negotiations are insufficient and that global regulatory structures are the missing piece, a point complicated by the current US administration's resistance to new guardrails.

The argument most likely to move is whether either company proposes a revised version of the cross-testing framework, or whether the UN appearances produce any concrete multilateral commitment. Any formal regulatory proposal from the US government, or a renewed bilateral announcement from the two labs, would either validate or undercut the competing readings of why these talks failed.