GPT-5.6: OpenAI’s most powerful AI is here, and almost no one can use it (yet)

OpenAI has unveiled its next-generation GPT-5.6 model series, consisting of three tiers named Sol, Terra, and Luna. At the U.S. government’s request, access is currently limited to roughly 20 vetted organizations. A broader rollout is expected within weeks.

The launch reflects a new dynamic in the AI industry: Washington is beginning to treat the most capable AI models as products requiring government review before wide release. The restriction follows similar U.S. measures applied to Anthropic’s Fable 5 and Mythos 5 models, as Carl Franzen reports for VentureBeat.

Three models for different needs

The three GPT-5.6 models are designed for different use cases and budgets:

• Sol is the most powerful option, built for complex reasoning, advanced coding, cybersecurity work, and long-running agentic tasks. It is priced at $5 per million input tokens and $30 per million output tokens.
• Terra balances performance and efficiency for high-volume production environments. It costs $2.50 input and $15 output per million tokens, the same price as the older GPT-5.4.
• Luna is optimized for speed and everyday tasks where cost and responsiveness matter most. It is the most affordable option at $1 input and $6 output per million tokens.

OpenAI is introducing two new capabilities with this release. A “max” reasoning setting gives Sol more time to work through difficult problems. An “ultra” mode assigns parts of a complex task to multiple subagents working in parallel, rather than relying on a single agent to handle everything sequentially.

In benchmark testing, Sol set a new record on TerminalBench 2.1, which measures performance on command-line tasks. It also outperformed its predecessor GPT-5.5 on biology and genomics tasks while using fewer tokens. On ExploitBench, a cybersecurity evaluation, Sol matched the performance of Anthropic’s Mythos Preview model while generating roughly one-third as many output tokens.

Safety layers and government oversight

OpenAI says GPT-5.6 Sol does not cross what it calls the “Cyber Critical” threshold in its internal preparedness framework. In testing against the Chromium and Firefox codebases, the model identified vulnerabilities and building blocks for exploits but could not autonomously produce a working full-chain attack. All three models did, however, cross OpenAI’s “High” cyber risk threshold in internal testing.

To address these risks, OpenAI describes a multi-layered safeguard system. It includes trained refusals inside the model itself, real-time classifiers that monitor outputs as they are generated, and the ability to pause generation when elevated risk is detected. A larger reasoning model then reviews the flagged content before it reaches the user. Repeated flags can trigger a review of an account’s broader activity.

OpenAI also dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming, focused on finding systemic attack patterns rather than single-prompt workarounds. Human expert testing is continuing during the preview period.

OpenAI is openly critical of the current situation, stating in its announcement: “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 frames the limited release as a short-term measure while it works with the Trump administration to establish a repeatable framework for future model releases under a planned cyber Executive Order.

Sources

• Previewing GPT-5.6 Sol: a next-generation model – OpenAI
• OpenAI unveils GPT-5.6 Sol, Terra and Luna models — but only accessible to limited preview partners for now, per US Gov – VentureBeat
• OpenAI releases powerful new GPT-5.6 model under restrictions – Axios