OpenAI GPT-5.6 Sol launch: benchmarks, pricing, and limited access

OpenAI has unveiled GPT-5.6, a new generation headlined by the flagship model Sol and two cheaper tiers, Terra and Luna, and opened a limited preview for select partners at the direction of the US government. The company says Sol matches or beats Anthropic's Claude Mythos 5 in several benchmarks while using fewer tokens on security tests.

What exactly did OpenAI release?

GPT-5.6 is a generation with three permanent performance tiers: Sol (flagship), Terra (roughly equivalent to GPT-5.5 at half the cost), and Luna (budget). The release also adds a "max" mode for deeper reasoning and an "ultra" mode that runs complex tasks across parallel sub-agents. OpenAI charges per million tokens: Sol is $5 input and $30 output, Terra is $2.50 input and $15 output, and Luna is $1 input and $6 output. The company also changed prompt caching: cache writes cost 1.25x the regular input price, cache reads receive a 90 percent discount, and cached entries have a guaranteed minimum lifetime of 30 minutes.

How does Sol perform against Claude Mythos on benchmarks?

OpenAI's benchmark numbers place Sol at or above Mythos on key coding and biology tests. On Terminal-Bench 2.1, Sol scores 88.8 percent while Sol Ultra scores 91.9 percent, Claude Mythos 5 scores 88.0 percent, and Anthropic's Fable 5 scores 84.3 percent. On GeneBench v1, a genomics and quantitative biology test, Sol beats GPT-5.5 with a best-case 30 percent versus 22 percent. On ExploitBench, which measures finding and exploiting real security flaws, OpenAI says Sol matches a Mythos Preview's performance while using roughly a third of the output tokens; Sol does so at around 150,000 output tokens. The company also reports that Mythos 5 still leads at around 80 percent on one ExploitBench metric, though comparable efficiency data for Mythos 5 is not provided.

What about security behavior and preparedness?

OpenAI frames Sol as its most capable cybersecurity model yet but emphasizes defensive use. In tests with Chromium and Firefox, Sol found bugs and exploitation primitives but did not produce an autonomous full-chain exploit, and OpenAI says GPT-5.6 Sol remains below the "Cyber Critical" threshold in its Preparedness Framework. The company positions the model as better at spotting and fixing flaws than at running end-to-end attacks, while Anthropic's Mythos achieved autonomous full-chain exploits in a different benchmark.

Why did OpenAI limit access and what did it say?

The preview is restricted to select partners at the explicit direction of the US government, the same authority that previously removed Anthropic's Fable 5 from the market. OpenAI criticized the policy in a public statement: "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." The company argues that token efficiency gains in Sol could lower effective cost per task compared with prior generations, countering trends of rising model costs.

Why it matters

GPT-5.6 bundles raw performance wins, a new per-tier pricing structure, and token-cost improvements that could lower the real price of many tasks. The restricted access approach creates a split between a small set of early partners and the wider developer and enterprise communities, which may slow adoption and third-party security research even as capability diffuses across benchmarks.

What to watch

OpenAI plans to make Sol available on Cerebras in July at up to 750 tokens per second. Track whether the US government expands or relaxes partner-only access and whether published comparison data for Mythos 5 includes the same efficiency metrics OpenAI cites for Sol.