OpenAI GPT-5.6 Sol, Terra & Luna: Benchmark Scores, Pricing, and Why Only 20 Organizations Can Use It

Benchmark Scores: How GPT-5.6 Performs

Terminal-Bench 2.1

This benchmark tests planning, tool use, and command-line agentic capability — a "brutal test" of a model's ability to operate autonomously in a terminal environment ^Y. Reported scores:

Sol Ultra's 91.9% represents a significant jump — roughly 4 percentage points ahead of Anthropic's best model and more than 8 points ahead of GPT-5.5 ^BI.

Other Reported Benchmarks

• ExploitBench: GPT-5.6 Sol matches Anthropic's Mythos Preview using only ~1/3 the output tokens, a major efficiency gain ^AOO.
• GeneBench v1 (long-horizon genomics analysis): GPT-5.6 Sol achieves stronger results than GPT-5.5 while using fewer tokens ^AO.
• Agent's Last Exam (code mode): GPT-5.6 Sol scored ~50.9%, the first model to clear 50% across 55 domains ^O.
• PostTrainBench Lite: GPT-5.6 Sol and Terra outperform GPT-5.5 at curating training data and executing post-training experiments ^A.
• Internal Research Debugging Eval: GPT-5.6 Sol demonstrates strong performance debugging real bugs from OpenAI's internal research experiments ^A.

Capability Risk Ratings

OpenAI rates all three GPT-5.6 models as "High" in both cybersecurity and biological risk, but none cross the "Critical" threshold ^A.

Pricing: How Much Does GPT-5.6 Cost?

Pricing is per 1 million tokens, across input and output ^HOK:

Prompt Caching

OpenAI also introduced prompt caching for GPT-5.6. Cache writes are billed at 1.25x the model's uncached input rate, while cache reads receive a 90% discount on cached-input tokens. GPT-5.6 supports explicit cache breakpoints and a 30-minute minimum cache life ^H.

Why Was the Launch Restricted to ~20 Organizations?

Primary reason: Direct U.S. government intervention. The Trump administration explicitly requested that OpenAI restrict access before release, citing national security concerns ^TAK. White House officials personally cleared the roughly 20 approved partner organizations one by one ^AI.

Broader Context

The administration considered GPT-5.6 to have "Mythos-like" capability — referring to Anthropic's Claude Mythos, whose predecessor (Claude Fable 5) was forcibly suspended worldwide by the Commerce Department on June 12, 2026, under export control rules ^A. The GPT-5.6 restriction stems from the same policy framework.

OpenAI stated publicly that this kind of government gatekeeping "keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them" ^A.

Mechanism of Restriction

• The preview is not a self-service program. Participation is limited to a small group of trusted partners with an OpenAI account representative ^H.
• Partner names were shared with the U.S. government before any access was granted ^HT.
• The models are available only through the OpenAI API and Codex, not in ChatGPT ^H.

When Will Broader Availability Come?

OpenAI says it "plans to expand availability as soon as possible" but has not announced a general-availability date ^H. Sam Altman has stated he expects broader access "within weeks" ^A. The company's system card notes: "We believe in broad access, and we plan to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks" ^D.

For now, the most powerful AI model ever released by OpenAI is only available to those Washington approves.

Key Takeaways

• GPT-5.6 Sol Ultra leads with 91.9% on Terminal-Bench 2.1, outperforming Claude Mythos 5 and GPT-5.5.
• No new mathematical discovery has been reported for GPT-5.6.
• Pricing ranges from $1/$6 (Luna) to $5/$30 (Sol) per million tokens.
• Access is restricted to ~20 U.S. government-approved organizations, with broader availability expected "in the coming weeks."
• All three models rate as "High" risk in cybersecurity and biological capabilities, but not "Critical."