OpenAI Launches GPT-5.6 Sol Under Strict Government Restrictions
OpenAI has released its GPT-5.6 Sol model to compete with Anthropic’s Claude Mythos, but strict US government access controls limit its availability.
- OpenAI’s new GPT-5.6 lineup features three tiers: the flagship Sol, the mid-tier Terra, and the budget-friendly Luna.
- Initial benchmarks show Sol outperforming Anthropic’s Claude Mythos 5 in agentic coding, scoring 88.8% on Terminal-Bench 2.1 compared to Mythos’s 88.0%.
- Access to the model is currently restricted to select partners via API and Codex due to US government mandates, a policy OpenAI publicly criticizes as unsustainable.
OpenAI has launched its new flagship model, GPT-5.6 Sol, designed to directly challenge Anthropic’s Claude Mythos class. However, due to strict US government security mandates, the model is currently locked behind a highly restricted preview available only to select partners through the API and Codex. OpenAI has publicly criticized this government-enforced gatekeeping, calling the restricted access pipeline unsustainable for developers, enterprises, and global security defenders.
Why it matters
The GPT-5.6 generation introduces a layered naming structure similar to Anthropic’s setup, according to reports from The Decoder. The “5.6” designation represents the core model generation, while the performance tiers are divided into Sol (the high-end flagship), Terra (matching GPT-5.5 performance at half the operating cost), and Luna (the lightweight, budget-friendly option). Additionally, OpenAI has introduced a “max” mode for deeper reasoning and an “ultra” mode designed to delegate complex workflows to parallel sub-agents.
In performance benchmarks shared by OpenAI, the flagship Sol model edges out Anthropic’s Claude Mythos 5 in agentic coding tasks. On the Terminal-Bench 2.1 benchmark, the standard Sol model scored 88.8 percent, while the Sol Ultra variant reached 91.9 percent, compared to Claude Mythos 5’s 88.0 percent and Google’s Gemini 3.1 Pro Preview at 70.7 percent. This intense rivalry highlights the ongoing battle documented in our ChatGPT vs Claude comparison, where both labs consistently leapfrog each other on technical capabilities.
What it means for you
Beyond coding, GPT-5.6 Sol shows notable efficiency gains in cybersecurity and biology. On ExploitBench—which evaluates an agent’s ability to identify and exploit vulnerabilities in Google’s V8 JavaScript engine—Sol matched the performance of Anthropic’s Mythos Preview while utilizing roughly a third of the output tokens. In genomics and quantitative biology via GeneBench v1, Sol achieved a 30 percent success rate, up from GPT-5.5’s 22 percent. Despite these capabilities, OpenAI emphasizes that Sol behaves primarily as a defensive tool, failing to generate autonomous full-chain exploits in tests with Chromium and Firefox, keeping it below the “Cyber Critical” threshold in its internal Preparedness Framework.
Frequently asked questions
Why is access to GPT-5.6 Sol restricted?
Access is limited to select partners via the API and Codex due to direct mandates from the US government, which previously pulled Anthropic’s Fable 5 model from the market.
How does GPT-5.6 Sol compare to Claude Mythos in coding?
According to OpenAI’s benchmarks, GPT-5.6 Sol scored 88.8% on Terminal-Bench 2.1 (and 91.9% in Ultra mode), outperforming Anthropic’s Claude Mythos 5, which scored 88.0%.
What are the different tiers in the GPT-5.6 model lineup?
The lineup includes three permanent performance tiers: Sol (the flagship), Terra (matching GPT-5.5 capabilities at half the cost), and Luna (the budget-friendly option).
To see how these two AI giants stack up across other models, read our full ChatGPT vs Claude head-to-head comparison.
ChatGPT vs Claude (2026): Which Is Better? (Tested) →Source: The Decoder. Published June 26, 2026.
Ali has hands-on tested 50+ AI tools and tracks model releases daily. Every verdict here comes from real, paid usage — never vendor demos or sponsored placements.
AI Tools Worth is independent and unsponsored. Some linked guides contain affiliate links — they never change our verdicts.