OpenAI is working with rivals Anthropic and Alphabet’s Google DeepMind on AI safety, Bloomberg reported Tuesday, citing the ChatGPT maker’s global policy lead. The collaboration covers shared safety standards and independent evaluation of frontier models, and it lands days after the CEOs of OpenAI and Anthropic both publicly called for ways to slow the pace of frontier development.
The three labs have spent two years criticizing each other’s safety practices. A joint effort on standards does not erase that history, but it signals that the industry’s leading companies now see a common problem worth solving together: nobody can slow down alone.
The prisoner’s dilemma in plain terms
The logic is uncomfortable but simple. If OpenAI delays a model release for months of safety testing, Google or Anthropic can ship a better model first and take the customers, developers, investment and talent. Chinese firms might keep developing regardless. The result is a race where every player would prefer slower conditions but no player can afford unilateral restraint.
Anthropic CEO Dario Amodei laid out a three-step deceleration plan on September 12: independent evaluators for AI companies, joint safety standards among major firms, and international expansion of those standards. OpenAI CEO Sam Altman has expressed willingness to participate in independent evaluation. Bloomberg reported that Altman told staff any slowdown would be coordinated with other companies, and acknowledged some rivals might not follow.
The acceleration is measurable. According to Artificial Analysis, the median release interval for frontier models across OpenAI, Google, Anthropic, Meta and xAI fell from 37.5 days in 2023 to 11 days this year. OpenAI’s own median dropped from 170.5 days to 49. Anthropic’s fell from 126 days to 71.5. AI is speeding up AI: labs delegate coding, debugging and experiment analysis to AI agents, so the same staff produces more research, which improves the next model, which automates more research. Each lab now ships models at a pace no internal review process was designed for.
What the collaboration actually covers
Details remain thin. OpenAI declined to describe scope beyond safety cooperation, and neither Anthropic nor Google DeepMind has issued a statement of its own. Based on the CEOs’ public proposals, likely areas include third-party evaluation of frontier capabilities before release, shared incident reporting, and common benchmarks for dangerous capabilities in cyber, bio and autonomy.
Precedent exists. The labs already participate in the UK AI Safety Institute’s pre-deployment testing and the US AI Safety Institute’s voluntary agreements. What is new is direct lab-to-lab coordination on standards rather than working through government intermediaries. That could move faster, and it avoids the problem of standards frozen into regulation before anyone knows what works.
Skeptics inside the field point out the obvious conflict of interest: companies setting their own safety rules, even jointly, are marking their own homework. Amodei’s plan addresses that with independent evaluators, but funding, access and methodology all remain undefined. There is also an antitrust shadow. Three companies holding most of the frontier model market agreeing on release practices is exactly the kind of coordination competition regulators examine. Whether that pushes the work into government-affiliated channels, as the UK and US institutes allow, is an open question.
The public clash that frames it
The cooperation news broke the same week the industry’s sharpest safety disagreement went public. At Salesforce’s Dreamforce conference on Monday, Nvidia CEO Jensen Huang argued the industry needs no new AI laws and that engineering discipline at the company level is enough. Hours earlier, Amodei had urged peers to slow frontier development and pause releases of systems that fail safety thresholds.
The disagreement is not cosmetic. Huang’s position reflects the computing industry’s standard posture: regulation follows maturity, and premature rules entrench incumbents. Amodei’s position is that the release interval math makes that posture obsolete when capabilities arrive faster than institutions can evaluate them. Anthropic researcher Jacob Coxon quit this week saying developers were “playing with our lives.” OpenAI chief scientist Jakub Pachocki wrote in a recent internal message that AI poses serious dangers.
EU regulators are watching. ChatGPT, Reddit and Roblox were added Monday to the EU’s list of digital services facing heightened scrutiny under the Digital Services Act, and the bloc is preparing rules requiring mandatory human oversight of AI hiring systems from 2026. If US labs set joint standards, they will still have to reconcile them with Brussels, and with whatever the UK’s institute concludes from its own testing rounds.
OpenAI has other distractions. Bloomberg and other outlets report the company is preparing a confidential IPO filing in the coming days or weeks, working with Goldman Sachs and Morgan Stanley, with a public debut potentially targeted for this month. Public-market investors will ask hard questions about safety spending, and a documented cross-lab framework is easier to defend than ad hoc practices.
Why it matters for markets
The AI capex story runs through everything this year: Nvidia’s earnings, data center leases like Anthropic’s $19 billion TeraWulf deal, the bond market’s repricing. A credible safety slowdown, coordinated across the three biggest labs, would change release cadences and potentially procurement plans. That is why the story moved past the tech press on Tuesday.
For now, treat the announcement as a signal rather than a regime. No standard has been published, no evaluator named, no release delayed. What changed is that the three labs building the frontier have stopped pretending the race is purely competitive. The next useful milestone would be a named independent evaluation body with access commitments. Until then, the prisoner’s dilemma is still running, the players have just started talking to each other.