What happened
OpenAI, Anthropic, Google and Meta submitted classified frontier-model benchmark packages to the US AI Safety Institute on July 21, according to people familiar with the process. The August 1 deadline was set by the June executive order for covered models including GPT-5.6 Sol and Mythos.
Submissions include cybersecurity red-team results, biorisk evals and agentic misuse scenarios reviewed in secure facilities. Outcomes could influence whether GPT-5.6 and Mythos graduate from preview to wider release. None of the labs commented publicly on submission details.
Why it matters
The classified benchmark process is the enforcement backstop behind the voluntary release standards announced July 14. Labs can meet the deadline on paper while outcomes still delay general availability if reviewers flag issues.
Developers waiting on gated models should treat August not as a guaranteed unlock date but as the start of a review period with uncertain timing.
MintedBrain take
Plan Q3 roadmaps without assuming Sol or Mythos go wide on August 2. Keep mid-tier and open alternatives production-ready so gating politics do not block shipping.
Discussion
Sign in to comment. Your account must be at least 1 day old.