Every Frontier Lab Promised to Pause. Now They Only Promise to Pause If Everyone Else Does.

The Future of Life Institute published its Summer 2026 AI Safety Index on July 7. Seven outside reviewers, nine companies, 37 indicators, six domains. The coverage that followed picked the easiest sentence in the report and ran with it: nobody got an A. Anthropic finished first with a C+. OpenAI and Google DeepMind took a C. Meta got a D+. xAI, DeepSeek, and Mistral failed outright, one company each from the US, China, and Europe. Every outlet had the same headline by Tuesday afternoon.

The grades are the least interesting thing in that document. Report cards on voluntary safety practice are graded against a rubric the graders wrote, and you can argue about the rubric forever. What you cannot argue about is a company changing its own written policy. That is a fact with a diff. And the diff the panel found, sitting three bullets down the executive summary, is the actual news: four of the labs quietly rewrote what they promised to do when they get scared.

The Clause That Changed

Here is the thing that used to be true. Anthropic's responsible scaling policy said the company would pause if it crossed certain safeguard thresholds. Not pause if the industry agreed to pause. Not pause if a regulator ordered it. Pause. Unilaterally. That was the entire point of the commitment, and it was the thing safety people pointed to when they wanted to argue that voluntary governance could work at all. Google DeepMind and Meta made versions of the same pledge. OpenAI made one too.

Per the FLI panel, all four have moved. Anthropic's newest version says the company will consider pausing if competitors do the same. OpenAI attached similar conditions. Google DeepMind and Meta did not soften their pledges, they voided them. The reviewers called this "moving goalposts" and said it has "undermined safety frameworks across the board."

Read the new version carefully, because the grammar is doing a lot of work. A pause that triggers only when your competitors pause is not a red line. It is a request. It says: I will stop doing the dangerous thing at the exact moment stopping stops being expensive. Every lab holding that clause is waiting for a different lab to move first, and every lab knows every other lab is waiting. That is not a safety framework. That is nine players in a standoff, each holding a gun they have promised to lower right after somebody else lowers theirs.

The Full Scoreboard

For completeness, here is the grade matrix. Note the Existential Safety row, which is the only domain where the entire industry is underwater: no company exceeds a C-, and most score D or below. The panel acknowledged real work happening there (Anthropic's constitutional classifiers, DeepMind's monitoring commitments, Meta's loss-of-control provisions, OpenAI's push for governance institutions) and judged all of it "entirely inadequate." Their objection to the dominant paradigms of interpretability and chain-of-thought monitoring is four words long and hard to get around: detection is not prevention.

We Have Been Documenting the Mechanism All Year

Here is why I am not surprised, and why you should not be either. This site logs the pace of this industry every single day, and the pace is the cause. In one 48-hour window last week five frontier coding models shipped. The token price floor has been in freefall all year, which we tracked through the 2026 pricing war, and it is now down around a dollar per million input tokens. US startups raised $412.7 billion in the first half of 2026 and roughly 86 percent of it went to AI. Anthropic is inside an S-1 window with a $200 billion compute commitment to service.

Now put a unilateral pause clause in that environment and ask what it is actually worth. It is a promise to hand your competitors a multi-month head start, during an IPO process, in the most capital-saturated technology market in history, on the basis of an internal threshold that only you can see and only you can verify. Nobody was ever going to honor that. The labs did not defect from the pledge because they turned evil in 2026. They defected because the pledge was structurally unpayable the moment the money arrived, and 2026 is when the money arrived. FLI did not catch a conspiracy. It caught an incentive.

Max Tegmark's line in the release was that AI companies are sprinting toward a cliff, that they acknowledge the risks of artificial superintelligence and keep racing to build it anyway. He is describing a collective action problem and calling it a character flaw. The more useful framing is the boring one: a voluntary commitment that costs its maker a fortune to honor and nothing to abandon will be abandoned, on a schedule set by how much money is on the table. Roughly $412.7 billion showed up. The clause changed.

The Military Reversal

The panel flagged a second reversal on the same shape. From 2024 to 2026, Anthropic, OpenAI, Google DeepMind, and Meta all had policies broadly prohibiting military applications. All four reversed, joining xAI and Mistral in actively seeking defense partnerships. Tegmark's summary to Axios was "boy oh boy has that changed."

The panel singled out Anthropic here, which is worth sitting with given Anthropic is the company that topped the index. Reviewers criticized what they called questionable military engagements, including a reported link to the Minab school strike. I want to be careful with that one: it is a reported link, characterized by a review panel with a declared point of view, and it is not an established finding. Treat it as an allegation that has been raised, not a fact that has been settled. But the broader pattern does not depend on that single case, and the broader pattern is not in dispute, because the policy documents are public and they changed. The safest lab in the industry by this index's own measure is the same lab drawing the panel's sharpest current-harms criticism. Both things are true. That is the whole problem with a one-number ranking.

Mistral, which failed the index, pushed back on the methodology and made the one argument in this whole story I think deserves more air. Its position is that open weights move the safety decision to the deploying enterprise, and that a handful of companies deciding behind closed doors what is safe for everyone else is itself a concentration risk worth grading. That is a real argument. It is also, as we worked through in the GLM-5.2 piece, an argument that gets weaker the moment you notice most people never self-host anything and just route through the vendor's API anyway. Open weights are a real check on concentration. They are not a safety framework, and Mistral does not get to use one as a substitute for the other.

Our Take

The voluntary era is over and this report is its obituary, written by the people who most wanted it to work. That is what makes it credible. FLI is not a neutral party, its panel is stacked with researchers who have argued for aggressive action on catastrophic risk for years, and you should weight the grades accordingly. But you do not need to trust FLI's rubric to read a policy diff. Four labs changed four documents. The direction of every change was the same, and it was away from the thing that constrained them.

The panel's parting demand is that labs make the pause promise genuinely unilateral again, and push for comprehensive safety legislation as fast as possible. The first half of that is not going to happen and asking for it is a category error: you cannot ask nine competitors to solve a coordination problem by each unilaterally accepting a loss. The second half is the only lever with a mechanism behind it, which is why the interesting governance action in 2026 is not in any lab's policy PDF. It is in California's bill queue, in the EU AI Act phase-in, and in whatever comes out of the off-switch conversation in Brussels and at the G7. We keep the running list of what has actually been enacted, by jurisdiction, on the AI policy registry.

Three signposts I am watching from here. First, whether any lab restores an unconditional pause clause, which would be the single most informative thing any of them could do and which I do not expect from anyone. Second, whether the four remaining survey holdouts (Alibaba, xAI, DeepSeek, Mistral) participate in the Winter index, because five of nine responded this round and participation is the cheapest possible signal a lab can send. Third, and this is the one that actually decides it, whether a binding statutory threshold lands anywhere with real enforcement before the next index ships. If it does, the pause clause stops being a promise and starts being a law, and the coordination problem gets a coordinator. If it does not, expect the Winter 2026 grades to be worse, and expect the same nine companies to explain that safety is a shared responsibility.

The scoreboard everyone quoted says Anthropic is winning. What the report actually says is that the industry graded its own homework for three years, and just erased the hardest question.