By Google, OpenAI, and Anthropic are building a self-regulatory body for frontier AI. They’re calling it the Frontier AI Standards Agency — FASA — and they want it running by end of 2026 or early 2027. No government oversight. Entirely voluntary. And to lead it, they’ve approached the man who left the White House three months ago declaring “there will not be an FDA for AI.” If that sounds like a contradiction, that’s because it is.
What FASA Actually Does
The proposal is modeled on FINRA, the financial industry’s self-regulatory organization. Pillars include shared pre-release model evaluations, standardized incident reporting, auditor qualification standards, and formal definitions of what “voluntary safety commitments” actually mean. All three labs published their own internal safety frameworks in June 2026 — the coordinated timing suggests FASA was in planning well before the announcement.
The catch is the part borrowed least from FINRA. FINRA operates under SEC supervision and can suspend firms, impose fines, and exclude members from the market. FASA has none of those powers. It is a voluntary body whose members are the same companies that will be evaluated. This is the Frontier Model Forum problem, version two.
The Frontier Model Forum Problem
The Frontier Model Forum already exists. Google, Microsoft, OpenAI, and Anthropic formed it in 2023 with a $10M safety fund and a mandate to advance responsible AI development. It has published best practices. It has never stopped a single release, required a single evaluation, or enforced a single standard. FASA is being positioned as structurally different — more formal, with auditor certifications and pre-release requirements. But the same structural conflict applies: the companies being evaluated fund and staff the body doing the evaluating.
Sriram Krishnan, the man they want to lead FASA, spent 18 months as the Trump administration’s top AI policy adviser. His explicit position: no licensing, no central regulator, and direct opposition to any government body that would “put sand in the gears” of AI progress. He left in June 2026. The three biggest AI labs now want him to run their private version of exactly what he argued against. His answer to the offer will tell you everything about whether this body has any real authority.
The Impact on Developer Roadmaps
Here is where developers need to pay attention. When pre-release safety evaluations become standard — even voluntary ones — frontier model releases slow down. The evidence is already accumulating. At the G20 Innovation Ministerial, Sam Altman said next-generation models will be “sobering for everybody,” pairing it with an acknowledgment that OpenAI is deliberately throttling capability in what it ships. Enterprise CIOs who built 2027 technology roadmaps around a steady cadence of frontier releases are now dealing with uncertainty they did not budget for.
The evaluation bottleneck is already real. METR and Redwood Research spent six days on-site evaluating an OpenAI agent deployment — the full dataset arrived in their final two days. Apollo Research had three days with GPT-6 Astra, only two with chain-of-thought access. If FASA standardizes pre-release evaluations, access windows will be longer, which means release windows will be longer. Model capability at launch may also be more conservative than what the underlying model supports, constrained to whatever passed the safety threshold at evaluation time.
What to Do Right Now
FASA has not launched. Krishnan has not accepted the role. But the behavior change — slower releases, tighter pre-release access, more constrained capabilities at launch — is already happening regardless of whether FASA formalizes it. Adjust accordingly.
- Build provider abstraction now. Use LiteLLM, PortKey, or a custom router. Any application wired directly to a single model provider is exposed to capability gaps, safety-hold delays, and pricing changes with no fallback. Design for substitution from day one.
- Plan for capability freezes. If your 2027 product roadmap depends on a specific model capability — extended context, real-time multimodal inference, or sub-100ms latency — build fallback logic. The capability may ship later than planned or with more constraints than expected.
- Check your SLAs. Your AI vendor agreements were almost certainly written before “3-month safety evaluation delay” was a realistic event. If you have uptime or capability guarantees, verify whether they account for regulatory or evaluation holds.
- Watch the auditor certifications. METR and Apollo Research are the current de facto evaluators. Whoever FASA certifies will start appearing on model cards. An “evaluated by FASA-certified auditor” badge will become a procurement signal — both for enterprises buying AI services and for developers choosing which models to build on.
The broader pattern is worth naming directly: frontier labs are coordinating to define the standards they will be held to, staffed by people they choose, on timelines they control. That is the definition of self-regulation, and the history of self-regulation across industries is not encouraging. Watch whether FASA gets enforcement power. Until it does, treat it as a signal about slower release cadence — not a guarantee of safer models.
