Frontier models now launch under government review

Last week the US government became an active gatekeeper for frontier AI models. Not as a hypothetical or a policy paper. As a working process that both OpenAI and Anthropic had to navigate in real time.

I started a new job the same week, so my attention was split. But this is the story that matters from June 22 to 28. It is the week the US AI review regime went live.

OpenAI releases GPT-5.6 under a government-gated preview

OpenAI launched GPT-5.6 on June 26 as a three-model family: Sol (flagship), Terra (balanced), and Luna (fast and cheap). Sol scored 96.7% on OpenAI’s internal cyberattack benchmark, crossing the “high” risk threshold in the company’s own preparedness framework. The White House had asked OpenAI to stagger the release over cybersecurity concerns. The result was a limited preview to roughly 20 pre-approved partner organizations, with the government approving each one on a case-by-case basis.

OpenAI made its position clear in the launch post. “We don’t believe this kind of government access process should become the long-term default,” the company wrote. “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” The company framed the arrangement as a short-term step toward broader availability in the coming weeks.

The pricing was aggressive. Sol costs $5 per million input tokens and $30 per million output tokens. Terra is half that. Luna is $1 and $6 respectively. At those prices, OpenAI is competing directly on cost while the government handles the access list.

Anthropic gets a narrow clearance for Mythos 5

On the same day, Commerce Secretary Howard Lutnick authorized Anthropic to restore Claude Mythos 5 for roughly 100 US organizations operating and defending critical infrastructure. Mythos 5 had been taken offline two weeks earlier under export controls that the administration invoked after the model’s initial release.

The authorization letter said Anthropic had worked with the government to address risks, and that “appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model.” Fable 5, the subscription-facing model that millions of developers have been locked out of since June 12, was not included. Reports indicated Pentagon and NSA sign-off was the remaining step, with access expected to follow within days.

The result was a tiered access structure that had never existed before: one model for critical infrastructure, a different model for paying subscribers, with the government deciding who gets what.

Google Gemini 2.5 Pro with Deep Think lands in the gap

Google DeepMind launched Gemini 2.5 Pro with Deep Think reasoning mode on June 22, and the timing was strategic. With Fable 5 suspended and GPT-5.6 locked behind a government preview, Google’s model was the most capable AI system freely accessible to the general public.

Deep Think uses extended parallel inference, exploring multiple reasoning paths before committing to a response. The model scored 82.4% on GPQA Diamond, surpassing Fable 5’s 79.1% and GPT-5.5’s 76.3%. Google had the frontier to itself, even if only temporarily. The model is available through Google AI Ultra subscriptions and the Gemini API for developers.

OpenAI and Broadcom unveil Jalapeño, a custom inference chip

OpenAI and Broadcom introduced Jalapeño on June 24, a custom LLM inference processor designed in about nine months with AI assistance. OpenAI said early lab testing running GPT-5.3-Codex-Spark showed substantially better performance per watt than current alternatives, with full benchmarks still pending. The target is gigawatt-scale deployment by late 2026.

This is OpenAI’s clearest move toward owning its compute stack end to end. Custom silicon means OpenAI can differentiate on cost and latency in ways that API competitors cannot replicate. For anyone building on OpenAI, the direction matters: if inference costs keep dropping on custom hardware, the gap between OpenAI and everyone else running on Nvidia GPUs could widen.

DeepMind loses four elite researchers in one week

Google DeepMind lost four of its most prominent researchers in a single week. Noam Shazeer, co-author of the 2017 “Attention Is All You Need” paper and co-lead of the Gemini project, joined OpenAI as Lead for Architecture Research. Nobel laureate John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, joined Anthropic. Jonas Adler and Alexander Pritzel, both AlphaFold contributors, also went to Anthropic. A fifth departure, senior safety researcher Arthur Conmy, followed the next day.

The aggregate market-cap loss tied to the departures was roughly $270 billion from Alphabet’s valuation. CEO Demis Hassabis publicly noted that talent movement between leading labs is expected, and that Google has “by far the biggest and broadest research bench” in AI. But losing the co-author of the Transformer architecture in the same week as a Nobel laureate is not ordinary turnover.

What this week meant

The clearest pattern across all of these stories is that AI model releases are no longer just product decisions. They are regulatory events. OpenAI’s staged rollout, Anthropic’s tiered access, Google’s opportunistic timing, the chip investment, the talent war: every signal points toward a market that increasingly resembles national security infrastructure more than ordinary SaaS.

For developers, the immediate impact is uncertainty. Which models will be available next month? Will government review slow down the release cadence of the best models? Will non-US alternatives like GLM-5.2 fill the gap while Western labs wait for clearance? These are not theoretical questions anymore. They are the operating environment.

FAQ

What is the US government’s new AI review process?

The administration’s June 2 executive order requires frontier AI companies to submit advanced models for government review up to 30 days before release. The GPT-5.6 preview and Mythos 5 restoration were the first operational tests of that framework, creating a de facto licensing regime for the most capable models.

Which GPT-5.6 models were released and what do they cost?

OpenAI released three models: Sol ($5/$30 per million tokens), Terra ($2.50/$15), and Luna ($1/$6). Sol is the flagship for complex reasoning and agentic tasks. All three were initially limited to about 20 government-approved partner organizations.

Why was Anthropic’s Mythos 5 taken offline and then partially restored?

Mythos 5 was taken offline on June 12 under Commerce Department export controls citing national security concerns around its cybersecurity capabilities. It was partially restored on June 26 for approximately 100 US organizations operating critical infrastructure, after Anthropic worked with the government to implement safeguards.

How does Google Gemini 2.5 Pro with Deep Think compare?

Gemini 2.5 Pro with Deep Think scored 82.4% on GPQA Diamond, surpassing both Fable 5 (79.1%) and GPT-5.5 (76.3%). Its Deep Think mode uses parallel reasoning paths before committing to an answer. The model was freely accessible while its main competitors were under government access restrictions.