Lead Analysis
Google, OpenAI and Anthropic form SAFA, an AI standards authority
An IAEA-style, industry-run body to set model-safety standards and run frontier benchmarks — launch targeted for late 2026 or early 2027
Saturday, September 26, 2026: Google, OpenAI and Anthropic on September 25 revealed plans for the Standards Authority for Frontier AI (SAFA) — an independent, industry-run body that would set model-safety standards and run frontier benchmarks, filling gaps the labs say governments have left open.
The plans, reported by The Information on September 25, mark the first formalised cross-lab standards body of the frontier era. SAFA would operate independently of government oversight, modelled on the FINRA-style self-regulatory organisation DeepMind chief Demis Hassabis floated in a July essay: common standards to assess frontier AI models, shared benchmarks, and a verification mechanism. Launch is targeted for late 2026 or early 2027, and representatives of the three companies have been meeting for months.
The timing is the story. The announcement lands days after Anthropic CEO Dario Amodei and, this week, OpenAI CEO Sam Altman publicly called for a slower frontier after models at Anthropic, Google, OpenAI and Meta reportedly launched cyber attacks; four days after a proposed class action alleged the labs illegally coordinated to slow AI; and one day after the White House’s Office of the National Cyber Director asked OpenAI and Anthropic to withhold new models from the UK’s AI Security Institute pending US review. Rival labs are self-coordinating precisely as governments fragment over who tests first.
For Indian enterprises the read is threefold. First, a private standards body will define what “safe frontier model” means before most regulators do — MeitY, the RBI and the IndiaAI Mission can adopt or contest SAFA’s norms, but the body will set the baseline Indian procurement teams evaluate against. Second, US-first release sequencing — the UK, once a privileged early tester, is now deprioritised — signals India’s test access will be secondary too, strengthening the case for domestic evaluation capability. Third, SAFA’s benchmarks would give Indian enterprises a vendor-neutral yardstick at a moment when a four-lane Q4 model pipeline is about to land.
Markets were quiet against the governance news. Friday’s session saw a partial rebound from Thursday’s crude-driven rout: Sensex gained 315.20 points to 73,895.74, Nifty 77.40 to 23,140.50, the rupee firmed 19 paise to 95.80, and Brent hovered near $105.5. IT stayed weak, with Infosys slipping below ₹1,000 to ₹998 and Wipro hitting a fresh 52-week low — the AI-slowdown narrative and two-decade-high US bond yields are compressing tech multiples even as AI capital expenditure accelerates. OpenAI’s DevDay on September 29 is the next catalyst.
