The best safety grade in AI is a C+

The Future of Life Institute published its Summer 2026 AI Safety Index on Tuesday. Nine frontier labs were graded across six domains. Anthropic came top with a C+. Nobody passed.¹

This is, on its face, a bad report card. It is also, if you are Anthropic in the run-up to a reported October IPO, the most useful piece of third-party marketing collateral available in the market. The interesting question is not what the grades mean in absolute terms. It is what they mean as a competitive credential when everyone else scored worse.

What was actually published. FLI graded nine labs on risk assessment, current harms, safety frameworks, existential safety, governance, and information sharing. Anthropic scored 2.66 out of 4.0 for a C+. OpenAI and Google DeepMind received C. Meta received D+. xAI, DeepSeek and Mistral received F.¹ No lab scored above D on existential safety, the second consecutive edition where that domain flatlined at the bottom.²

The panel also made a finding it labelled "moving the goalposts": that several labs, Anthropic, OpenAI, DeepMind and Meta among them, have quietly weakened earlier commitments to pause development or deployment at defined risk thresholds. The panel linked this softening to the scaling of defence contracts across the same set of firms.³

The safety-as-moat frame, and where it starts to buckle. The working thesis on Anthropic, held by a lot of the capital that has flowed into the company, is that safety posture is a durable competitive position. Responsible Scaling Policy (RSP: Anthropic's self-imposed framework tying deployment to capability-linked risk levels), ASL-3 (Anthropic Scaling Level 3: the current classification, which triggers specific safeguards under the RSP), a research programme skewed toward alignment. The bet is that enterprise buyers, sovereign customers and regulators will pay a premium, or in some jurisdictions only buy at all, from the lab that can point at the strongest safety story.³

The FLI index is the frame's first serious public test. And on the narrow question — is Anthropic's safety posture visibly better than its peers' to a third-party panel? — the answer is yes. C+ against C, D+ and three Fs is a legible spread. For a procurement officer choosing between vendors, "highest score on the only recurring public safety benchmark" is a defensible line, regardless of whether C+ is objectively any good. I would expect this grade, or a reference to it, to appear in Anthropic's road-show materials. It costs the company nothing to cite and there is no cleaner credential available.

But the "moving the goalposts" finding does something specific to the frame. If Anthropic, OpenAI, DeepMind and Meta are all softening pause commitments as defence revenue scales, the differentiation between them on the dimension that matters most to safety-conscious buyers is compressing, not widening. The floor is rising, or the ceiling is falling, depending on how you want to draw it. Either way, the gap narrows.

That is a structural claim, and Anthropic has not confirmed it. The panel is inferring from published RSP revisions and disclosures; no lab has publicly said "we have weakened our pause threshold". Anthropic maintains its RSP and ASL-3 framework as active policy.³ The reader should treat the finding as the panel's reading of the primary documents, not as an admission. But the panel is reading the same primary documents underwriters will read, and if a similar reading lands in an EU AI Act enforcement docket, it becomes a compliance problem rather than an advocacy critique.

Existential safety is the tell. No lab scored above D on the existential safety domain, in either this edition or the prior one.² This is the domain FLI, as an organisation, cares about most, and it is the one where the composite grade is doing the most concealing work. A C+ headline sits on top of a D-or-worse floor on the panel's own priority metric. If you strip out the domains where the industry has made procedural progress, published frameworks, disclosed risk assessments, and look only at the thing the panel was set up to measure, nothing has moved.

Safety-as-moat needs the differential between labs to be growing. The panel's own metric says it is not.

The regulatory-infrastructure question. The FLI index has no legal force. It is produced by an advocacy organisation with a declared mission, and the labs have not validated its rubric.⁴ What matters for market structure is whether regulators pick up the scoring framework and cite it. The six-domain structure maps closely to the categories of obligation the EU AI Act imposes on providers of general-purpose models with systemic risk. If EU enforcement bodies, or any US successor to the previous federal safety-institute arrangement, begin citing FLI scores in docket filings, the index transitions from advocacy publication to compliance benchmark. That is a different market-structure event, and it would put a real cost on an F grade.

Prior editions produced similar F grades for the same set of labs without visibly constraining their commercial traction.⁴ The regulatory-infrastructure story depends on a policy shift that has not yet occurred. It is a live possibility, not a fact.

What this means for the IPO. If Anthropic prices in October at the reported timeline, the C+ will appear in the coverage. It is the cleanest available third-party safety credential, and it is better than every peer's. The company will not need to cite it aggressively; the market will do that for it. The harder question, whether the safety differential is actually widening or compressing, will not be settled before pricing, and probably not by pricing.

The uncomfortable read, if you own the safety-as-moat thesis, is that the panel most sympathetic to that thesis is telling you the moat is narrower than the composite score suggests. C+ against three Fs looks like a spread. C+ against a D+ existential safety floor shared across the top four labs looks like something else.

Footnotes