QUICK COMPARISON — FLI SUMMER 2026
● FLI grade: Anthropic — C+ (2.66). OpenAI — C (2.28). Gap: 0.38 points on 4.0 scale.
● FLI rank: Anthropic #1 of 9. OpenAI #2.
● Domain lead: Anthropic leads 5/6 domains. OpenAI leads Risk Assessment.
● Safety framework: Anthropic — RSP (Responsible Scaling Policy). OpenAI — Preparedness Framework.
● Framework action taken: OpenAI paused Astra when Preparedness Framework Critical threshold was triggered (August 7).
● Pause pledge status: Both weakened prior unconditional pause pledges per FLI panel — "moving goalposts."
● Military: Both moved from banning to actively seeking defence partnerships (FLI finding).
● Existential safety: Both below C- — worst domain for both labs.
The FLI index grades institutional safety posture — policies, governance, and disclosures — not the safety of deployed products. The 0.38-point gap between Anthropic (2.66) and OpenAI (2.28) is real but narrower than the letter grade difference (C+ vs C) suggests. As Digital Applied's enterprise analysis notes, the gap between OpenAI (2.28) and Google DeepMind (2.01) is 0.27, while the gap from Google DeepMind to Meta (1.32) is 0.69 — the largest single drop on the scorecard. The three C-range labs (Anthropic, OpenAI, Google DeepMind) form a closer cluster than the letter grades imply.
Anthropic for regulated enterprise procurement: FLI C+ #1 of 9 labs, Constitutional AI, RSP, Ode sovereign JV for banks/health systems/manufacturers, Tino Cuéllar CGAO, state-by-state AI safety law push. The documented safety posture most regulated industries can point to in a vendor selection memo.
OpenAI for enterprise buyers who prioritise Risk Assessment: OpenAI leads the Risk Assessment domain specifically — broader evaluation suite and external testing engagement. The Astra pause (August 7) demonstrated the Preparedness Framework actually triggers safety action when critical thresholds are reached.
Last updated August 10, 2026. Related: FLI Safety Index full results → · OpenAI pauses Astra →