The C+ ceiling: what the FLI Safety Index actually measures

The Future of Life Institute's Summer 2026 AI Safety Index gives its highest grade to Anthropic — a C+, 2.66 out of 4.0. That number is the story, but not for the reason most coverage has taken it. The index does not measure whether labs are safe. It measures whether their public commitments are legible enough to be graded. A C+ ceiling on that instrument tells us something about the disclosure market, not the safety one.

The ranking, briefly. Nine frontier labs, 37 indicators, six domains. Anthropic first at C+. OpenAI and Google DeepMind in the C range. Meta at D+. Z.ai and Alibaba Cloud at D-. xAI, DeepSeek, and Mistral at F. No lab scored above D on the existential-safety sub-domain — the second consecutive edition with that result.¹²³

What the index is, and what it is not. FLI grades what labs publish: policy documents, model cards, terms of service, voluntary disclosures, published research. It does not audit internal risk management. It does not test capability. The methodology is transparent about this, and prior editions have flagged the same limit.²³ The index measures the interface between the lab and the outside world, not the machinery inside.

That distinction matters, because it changes what a C+ actually means. Anthropic is, by any reasonable read, the lab that has invested most heavily in publishing its safety reasoning. Its Responsible Scaling Policy is the most detailed public commitment framework in the industry. It employs the largest visible safety research organisation. And on an instrument designed to reward exactly this, legible, structured, public commitments, it lands at 2.66.

The sharpest finding is not in the grade table. Reviewers accused Anthropic, OpenAI, Google DeepMind, and Meta of "moving the goalposts" on prior pause commitments — the language FLI uses for weakening previously stated danger thresholds.³ This is the load-bearing claim, and it deserves more weight than the letter grades.

Voluntary commitments are supposed to ratchet in one direction. As capability rises, thresholds tighten, or at minimum hold. If the four labs with the most sophisticated public safety apparatus are softening their stated thresholds as their models get more capable, the index is capturing regression, not stasis. That is a different kind of finding. It suggests the disclosure equilibrium is not merely low — it is drifting downward under commercial pressure.

The counter-case is real and worth stating. An F grade for xAI, DeepSeek, or Mistral does not mean those labs are less safe than Anthropic. It means they publish very little about safety governance. A lab with strong internal controls and minimal external documentation scores below a lab with polished public frameworks and weaker internal practice. FLI acknowledges this.²³ So the F tier is a disclosure failure, and might also be a safety failure — but the index cannot tell you which, and honest readers of it should not either.

There is also a defensible read of Anthropic's "goalpost-moving" as responsible adaptation. The RSP is explicitly a living document; updating thresholds as capabilities and evaluation methods evolve is what a serious framework is meant to do. The question is whether specific changes tightened or loosened the bar. FLI's reviewers say loosened. Anthropic would presumably say refined. Without the underlying diff, a reader cannot fully arbitrate — but the reviewers include Stuart Russell, and they were looking at the same public documents anyone else can read.³

Why the ceiling is a market-structure finding. A commercial lab writing a safety commitment is writing something a competitor, a regulator, a plaintiff's lawyer, and a customer will all read. Every enforceable sentence is a future liability. Every specific threshold is a future constraint on shipping. The equilibrium a competitive market produces is: commit to things you would do anyway; keep the hard cases in internal policy where they can be revised without a press cycle; make the public document sophisticated enough to demonstrate seriousness, vague enough to preserve optionality.

That equilibrium produces exactly a C+. Not because the labs are cynical, most of the safety researchers I read seem to mean what they write, but because the incentive gradient of writing publicly against your own future ship dates is punishing. A lab that wrote a genuine A-grade commitment framework, with hard numerical thresholds and pre-committed pauses, would be handing its competitors a shipping advantage. So no lab does. And on the instrument that measures this, no lab scores above C+.

The second consecutive zero on existential safety. Two editions now, no lab above D on that sub-domain.² Whatever one thinks of FLI's weighting of existential risk relative to near-term harms, and the fairness-and-ethics critique of that weighting is legitimate, the plateau is the finding. Commercial voluntary commitment structures have stopped improving on the dimension FLI cares about most. If you believe the risks the labs themselves describe in their own model cards and RSPs, that plateau is the argument for something other than voluntary commitment doing the work.

What to make of a C+. Not that Anthropic is unsafe. Not that xAI is unsafe. Both of those inferences over-read the instrument. What the index shows is that the disclosure market has a ceiling, that the ceiling is not rising, and that the top labs may be quietly lowering their side of it. Read as a report card on lab safety, the Summer 2026 Index is a Rorschach. Read as a report card on whether public commitments can carry the governance load the industry has asked them to carry, it is unusually clear. They cannot, and the labs writing them appear to know.

Footnotes