AI agents · OpenClaw · self-hosting · automation

Quick Answer

FLI Summer 2026 AI Safety Index: Anthropic C+ vs OpenAI C

Published:

FLI Summer 2026 AI Safety Index: Anthropic C+ vs OpenAI C vs Google C

On July 7, 2026, the Future of Life Institute released its Summer 2026 AI Safety Index, and every frontier lab failed. Anthropic topped the ranking at C+. OpenAI and Google DeepMind each scored C. Meta got C-. xAI and Chinese labs (Zhipu, DeepSeek) sat at D+ and below. No lab scored above C- in existential safety. The report is the most comprehensive third-party evaluation of frontier AI safety in July 2026 and the most damning.

Here is what the grades mean, who leads where, and what it says about the industry.

Last verified: July 16, 2026

The Grades

LabOverallRisk AssessmentCurrent HarmsSafety FrameworkExistential SafetyGovernanceInformation Sharing
AnthropicC+B-C+B-C-B-C+
OpenAICB-CCD+CC
Google DeepMindCC+CC+D+C+C
MetaC-CC-D+DD+D+
xAID+D+DDFDD-
Zhipu / DeepSeekDDDFFFF

(Exact sub-scores per FLI Summer 2026 report; letter grades reproduced with best-effort accuracy from published summaries.)

Six Evaluation Domains

FLI evaluated across:

  1. Risk Assessment — how comprehensively the lab evaluates its models pre-deployment
  2. Current Harms — how the lab addresses in-production harms (bias, misuse, privacy, misinformation)
  3. Safety Framework — whether a formal safety framework exists and is used as a deployment gate
  4. Existential Safety — how the lab reasons about worst-case, catastrophic, and civilizational risks
  5. Governance and Accountability — organizational structure, oversight, and transparency of decisions
  6. Information Sharing — publication of research, model cards, evaluations, and incident reporting

Anthropic led or shared the lead in 5 of 6 (risk assessment shared with OpenAI). Every lab scored below C- on existential safety.

What Anthropic Got Right

  • Published detailed model cards for Claude Sonnet 5 and Opus 4.8 — including capability evals, safety evals, and failure modes
  • Responsible Scaling Policy is a real deployment gate — Anthropic’s ASL levels actually delay releases (Sonnet 5 shipped June 30, 2026 after ASL-4 evaluation)
  • Petri open-source red-team framework — externally reproducible safety evals
  • Long-term benefit trust with independent trustees — Ben Bernanke joined in June 2026; the trust has real governance authority
  • Interpretability research — significant investment in mechanistic interpretability publications

Even so, Anthropic scored C+. The FLI called that a passing-but-not-satisfactory grade. In absolute terms, no lab satisfies the FLI evaluators.

What OpenAI Got Wrong (and Right)

Right:

  • Broadest external evaluation partnerships — Apollo, METR, and multiple academic red-teams
  • GPT-Red announcement (July 15) — self-play RL red-teaming is a real safety investment
  • Preparedness Framework — documented capability evaluations
  • Broad risk assessment suite

Wrong:

  • Governance changes moving away from safety commitments — capability team splits, restructuring, and departures of key safety researchers
  • Weaker transparency — some evaluations not published; model cards less detailed than Anthropic’s
  • Weakened prior public commitments on halting development at capability thresholds

The FLI explicitly noted OpenAI is moving in the wrong direction on governance while shipping faster.

What Google DeepMind Got Wrong (and Right)

Right:

  • High-risk bio and cyber threat testing under the Frontier Safety Framework
  • DeepMind’s research culture — long tradition of publishing safety-relevant work
  • Frontier Safety Framework v3.0 (June 2026) — structured evaluation protocol

Wrong:

  • Weaker on existential safety than Anthropic — DeepMind’s public existential-safety research is real but less integrated with deployment decisions
  • Limited transparency on Gemini 3.5 Pro / Flash evaluations compared to Anthropic’s Claude model cards
  • Governance opacity — as part of Alphabet, decision-making is less transparent than a pure lab structure

What the Report Says About Existential Safety

This is the FLI’s headline concern. Quote-adjacent from published summaries:

No company scored above C- in existential safety. Anthropic, OpenAI, Google DeepMind, and Meta have weakened or abandoned previous pledges to halt development if certain risk thresholds were reached and have softened their stance on military AI applications between 2024 and 2026.

Concrete examples the FLI cites:

  • OpenAI’s evolution on military and defence work
  • Google DeepMind’s Project Nimbus and military-adjacent deployments
  • Meta’s Muse Spark deployment cadence
  • Anthropic’s Claude for Government beta (July 15, 2026)
  • All labs softening prior “pause if capability threshold X” pledges

The FLI position: frontier labs are shipping faster and taking on more high-stakes deployments (defence, government, critical infrastructure) while their safety practices have not scaled proportionally.

What This Means for Buyers and Builders

For enterprises evaluating AI vendors:

  • Anthropic remains the strongest safety story. For workloads where safety posture matters to your board, insurers, or regulators, Anthropic has the best third-party evidence.
  • OpenAI’s GPT-Red is a positive signal — the July 15 announcement, one week after the FLI report, is the most concrete OpenAI response to safety criticism in 2026.
  • Google DeepMind is competitive on safety but less compelling to boards that read model cards
  • Meta and xAI should be scrutinized carefully for safety-sensitive workloads
  • Chinese labs should be evaluated per your regulatory environment — DeepSeek and Zhipu offer excellent performance-per-dollar but the FLI safety evaluation is unforgiving

For AI startups building on these APIs:

  • Cite the FLI report in your own safety and compliance documentation
  • Do not assume the model provider handles safety for you — deploy your own eval suite (Petri, PyRIT)
  • Watch how each lab responds over the next 6-12 months — the response separates who is serious from who is doing PR

How the FLI Evaluates (Methodology)

FLI’s evaluation involved:

  • Public documents review — model cards, safety frameworks, published research
  • Structured questionnaires — sent to each lab
  • Independent expert panel scoring — academic and industry safety experts scoring each domain
  • Cross-verification against public incidents — reported harms, jailbreaks, alignment failures
  • Comparison against 2024 baseline — 2026 report explicitly tracks direction of change

The method is public. The panel is disclosed. This is not a black-box ranking.

What Comes Next

  • September 2026: FLI mid-cycle update expected
  • December 2026: Winter 2026 AI Safety Index — full re-evaluation
  • EU AI Act GPAI Article 55 evaluations — will use similar structure, feed into regulatory action
  • US AISI and UK AISI evaluations — parallel government evaluations continue

Expect the industry to respond to poor grades not by fundamentally changing safety posture, but by:

  1. Publishing more model cards and safety artifacts (visible responses)
  2. Announcing new safety tools (GPT-Red is the template)
  3. Formalizing governance structures (independent boards, ethics reviews)
  4. Continuing to ship on the same 2-3 month frontier release cadence

Whether that adds up to real safety improvement or safety theater will be the story of H2 2026.

The Frame

The FLI Summer 2026 AI Safety Index is not a report card that will lead any lab to slow down. It is a diagnostic that says: the pace of frontier deployment has outrun the pace of safety scaling. Anthropic scores highest not because it is safe in absolute terms — it is not — but because it is the least worst.

If you build on frontier AI in July 2026, deploy defensively. Assume model providers are shipping faster than they can safely evaluate. Run your own red-team evaluations. Do not rely on the vendor’s word.

Sources