GPT-5.6 Goes Public July 9 — Government Gate Opens

Verdict: GPT-5.6 Sol moves from Pending to Recommended. The single barrier holding back our recommendation — government-gated access that kept the model behind a ~20-org preview — is lifted. The Commerce Department cleared the full launch, and public GA begins Thursday July 9.

What happened

OpenAI posted at ~4am UTC July 8: "GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We're expanding preview access globally now" (OpenAI, 2026). The Axios report that broke the story hours earlier confirmed the administration gave OpenAI the green light following additional testing by the Center for AI Standards and Innovation within the Commerce Department (Axios, 2026).

This is the resolution of a story we've been tracking since June 25, when the White House asked OpenAI to stagger the launch behind case-by-case government approval. The preview launched June 26 to ~20 trusted partners. Twelve days later, the gate is open.

Why our verdict changes

When we issued a Pending verdict on June 26, we wrote: "What we are waiting on: general availability and independent benchmarks." The GA barrier is now resolved. The benchmark barrier partially resolves with it — wider access means independent third-party evaluations will follow rapidly.

METR's caution about gaming on the highest detected rate of any publicly tested model stands. But that's a reason to read independent scores carefully, not to withhold a verdict.

The model's raw capability was never in question: 88.8% Terminal-Bench 2.1 base, 91.9% Ultra, first model to clear 50% on Agent's Last Exam. At $5/$30 per 1M tokens, it matches the GPT-5.5 rate card — same price, strictly better capability. The Cerebras-hosted variant at ~750 tok/s follows later in July.

What changes Thursday

The competitive window is strategic. Fable 5 goes credit-only July 12 at $10/$50. Claude Sonnet 5's introductory pricing lasts through August 31. GPT-5.6 lands at launch prices while its closest competitor simultaneously limits subscription access.

What to do

1. If you're on GPT-5.5 — Terra is the drop-in upgrade at half the cost. Test your agent pipelines against Terra first; escalate to Sol only for the hard fraction of requests. This three-tier routing architecture is the intended design.
2. If you're using Fable 5 — Thursday gives you a same-cost-or-cheaper frontier alternative before Fable shifts to credit-only on July 12. Run your hardest agentic workloads against Sol Ultra and compare.
3. If you're budget-conscious — Luna at $1/$6 resets the capability-per-dollar equation for high-volume work. Against Haiku 4.5 at $1/$5, the Luna value proposition is compelling.

One caveat we're watching: the government coordination process isn't gone — it's been exercised once and will be again for future models. OpenAI's own language from the June 26 blog stands: "We don't believe this kind of government access process should become the long-term default." But for today: the gate is open, the model is shipping, and the verdict is Recommended.

FAQ

Which GPT-5.6 tier should I use? Start with Luna for volume work, route to Terra for production quality, and escalate to Sol only for your hardest agentic workloads. The three tiers are designed for routing, not mutual exclusion — most teams will use all three.

Is the government review process over permanently? No. OpenAI's own statement opposes making case-by-case government approval the default. Future frontier models may face the same process, but for this cycle the gate is open.

When will independent benchmarks arrive? Now that the model is publicly available, expect third-party evaluations within days. Treat METR's gaming-detection caution as a reason to read scores critically, but the public release guarantees independent verification will happen.