GPT-5.6 Sol Topped Mythos—but Only for 20 Partners

OpenAI launched its strongest frontier model family on June 26 — then handed the keys to roughly 20 government-vetted partners. Sol Ultra scored 91.9% on Terminal-Bench 2.1, beating Claude Mythos 5's 88% and GPT-5.5's 83.4%. The same day, Commerce let Anthropic restore Mythos for about 100 approved organizations. Fable 5 stays dark. ChatGPT subscribers get nothing.

Why June 26 matters: This is the first frontier launch where Washington picked the audience before OpenAI picked the market — and OpenAI said in its own blog post it does not want that to become permanent.

The launch:

• Sol — flagship for coding, cyber, and deep reasoning; $5/$30 per million tokens
• Terra — GPT-5.5-class performance at half the price; $2.50/$15
• Luna — fastest, cheapest tier; $1/$6
• Sol Ultra — subagent mode that hit 91.9% on Terminal-Bench 2.1
• Sol Max — deeper single-model reasoning; 88.76% on the same benchmark

OpenAI's new naming splits generation (5.6) from durable tiers (Sol, Terra, Luna) — the same product architecture Anthropic uses with Opus, Sonnet, and Haiku.

The benchmark fight:

• Terminal-Bench 2.1: Sol Ultra 91.9% vs. Mythos 5 88% vs. GPT-5.5 83.4%
• ExploitBench: Sol competitive with Mythos Preview at roughly one-third the output tokens
• Agent's Last Exam: Sol is the only model above 50% in code mode (50.9%)

OpenAI chose the scoreboard. Anthropic's cyber crown just got a challenger — on paper.

The risk ratings — all three, not just Sol:

OpenAI's system card rates every GPT-5.6 tier — Sol, Terra, and Luna — as High capability in both cybersecurity and biological/chemical risk. That is a first: smaller, cheaper models in a family have never triggered High cyber classification before.

• Sol saturated internal capture-the-flag tests at 96.7%
• Terra hit 91.84%; Luna hit 85.19%
• None crossed OpenAI's Cyber Critical threshold — Sol found bugs in Chromium and Firefox but did not autonomously build a full-chain exploit

The cheap tier is now a High-risk tier. Enterprise buyers pricing Luna for routine summarization are buying a model OpenAI classifies at the same preparedness level as its cyber flagship.

The guest list:

At Washington's request, GPT-5.6 is available only through API and Codex to a "select group of trusted partners" — roughly 20 organizations, per VentureBeat and Axios reporting. Their names were shared with the government before launch.

Sam Altman told staff the ONCD and OSTP asked OpenAI to stagger the rollout. Commerce Secretary Howard Lutnick reviewed release plans with Altman on June 25. General availability is promised in "the coming weeks" — contingent on preview clearing federal review.

The same day, Lutnick's June 26 letter let Anthropic restore Mythos 5 for Annex A organizations — about 100 approved entities. Fable 5 remains offline with no timeline.

What OpenAI said — in public:

"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them."

That line sits in OpenAI's official product announcement — not a leak, not a lawsuit. The company is shipping under a process it is actively lobbying to kill.

The safety bill:

OpenAI dedicated 700,000 A100-equivalent GPU hours to automated red-teaming for universal jailbreaks — more compute than many startups spend training a model. The safeguard stack includes real-time cyber and biology classifiers, activation-based screening on Sol and Terra, and account-level review that can pause generation mid-stream.

OpenAI reports 94.8% recall on biology misuse detection and 81.6% on cyber — meaning roughly one in five risky cyber outputs may slip through. Legitimate security research will also hit false positives. That friction is the preview's point.

The market read:

• OpenAI kept Sol at GPT-5.5 pricing ($5/$30) despite a generational capability jump
• Terra undercuts GPT-5.5 by 50% while matching its performance class
• Luna at $1/$6 makes OpenAI's cheapest tier more expensive than Chinese frontier models like GLM-5.2 ($5.80 combined) and DeepSeek V4 Flash ($0.42)
• Sol on Cerebras hardware arrives in July at up to 750 tokens per second — for select customers only

What's next:

• July 2 — deadline for agencies to stand up the voluntary model-testing framework under Trump's June 2 cyber executive order
• August — formal federal process for assessing AI cyber capabilities
• Mid-July — earliest plausible general availability if Altman's "couple of weeks" timeline holds
• IPO season — OpenAI's confidential filing means every access restriction and risk rating lands in an S-1 disclosure

Convina's view: OpenAI won the benchmark and lost the launch. Sol Ultra topping Mythos on Terminal-Bench is a real competitive blow — but a score nobody can buy is a press release, not a product. Washington now runs a two-speed frontier market: roughly 20 partners for OpenAI, roughly 100 for Anthropic's Mythos, zero for Fable, and the rest of the world waiting on a cyber executive order that does not exist yet. OpenAI's public objection is the tell. The company knows guest-list releases will not scale to a trillion-dollar IPO — and accepted one anyway because the alternative was Anthropic's global kill switch. The race is no longer who builds the smartest model. It is who gets Commerce to say yes first.