GPT-5.6 Launch: Sol, Terra, Luna Beat Mythos Yet Stay Behind Paywall

On Friday OpenAI released a limited preview of the GPT‑5.6 series, comprising three models positioned for different use cases: the flagship Sol, the cost‑effective Terra for high‑frequency tasks, and the lightweight Luna for speed‑critical scenarios.

Sol is equipped with two new enhanced modes: “max” allocates additional inference time for deeper reasoning on complex tasks, while “ultra” orchestrates multiple sub‑agents to surpass the limits of a single agent. Terra matches the performance of the previous GPT‑5.5 generation at roughly half the price, and Luna further reduces cost for lightweight workloads.

OpenAI reports that Sol achieves the strongest results to date on several benchmarks. On Terminal‑Bench 2.1, Sol reaches 88.8 % while the ultra variant scores 91.95 %, both surpassing Anthropic Mythos 5 and Claude Fable 5. GeneBench v1 shows higher accuracy than GPT‑5.5 with fewer tokens consumed. HealthBench Professional and HealthBench Hard see modest gains, while HealthBench and HealthBench Consensus remain roughly unchanged. In ExploitBench, Sol attains Mythos‑level performance using only about one‑third of the output tokens.

In the ExploitGym benchmark—co‑created by UC Berkeley, OpenAI and other labs—all three GPT‑5.6 models demonstrate noticeable security capability improvements as inference intensity increases.

Pricing is expressed per million tokens: Sol costs $5 for input and $30 for output, roughly half of Anthropic Claude Fable 5’s rate; Terra is $2.5/$15, and Luna is $1/$6.

The release also introduces a flexible prompt‑caching mechanism that allows custom cache breakpoints with a minimum cache lifetime of 30 minutes. Writing to the cache is billed at 1.25 × the uncached input price, while reads receive a 10 % discount.

Security is a focal point: OpenAI built a multi‑layer protection stack that trains the model to reject disallowed attack requests, deploys a real‑time classifier during generation to pause suspicious outputs for review by a larger model, and monitors account‑level risk signals across sessions. Over 700 k A100‑equivalent GPU hours were spent on automated red‑team testing targeting generalized jailbreak paths, supplemented by third‑party expert red‑team evaluations throughout the preview period.

OpenAI announced that Sol will be deployed on Cerebras hardware in July, delivering up to 750 tokens per second for early‑access customers. The preview is limited to a small set of trusted partners and institutions, a decision driven by recent U.S. government intervention in AI regulation. OpenAI cautioned that such regulatory constraints should not become a long‑term norm, as they could delay developers, enterprises, security teams, and global partners from accessing advanced tools.