OpenAI's GPT-5.6 Sol faces government access restrictions similar to Anthropic's Claude Mythos under new federal oversight.
OpenAI has unveiled GPT-5.6 Sol, a new generation of models built to compete with Anthropic's Claude Mythos class. The limited preview is currently available only to select partners through the API and Codex, under restrictions imposed by the US government. The same government previously pulled Anthropic's Mythos-class model Fable 5 from the market.
OpenAI has expressed clear frustration with these access controls. In an official statement, the company said: "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 is introducing a layered naming scheme similar to Claude's structure. The numerical component (x.6) denotes the generation, while Sol, Terra, and Luna represent permanent performance tiers that can evolve independently. Sol serves as the flagship tier, Terra matches GPT-5.5 performance at half the cost, and Luna functions as the budget option. Additionally, OpenAI offers a "max" mode for deeper reasoning and an "ultra" mode that delegates complex tasks to sub-agents running in parallel.
In agentic coding benchmarks, Sol demonstrates a lead over Claude Mythos 5. On Terminal-Bench 2.1, Sol scores 88.8 percent, with Sol Ultra reaching 91.9 percent, compared to Claude Mythos 5 at 88 percent and Fable 5 at 84.3 percent.
Sol also shows improvements in biology applications. On GeneBench v1, a benchmark for genomics and quantitative biology, it outperforms GPT-5.5 (30 percent versus 22 percent best case) while consuming fewer tokens.
On ExploitBench, which evaluates how well AI agents can identify and exploit real security flaws in Google's V8 JavaScript engine up to full code execution, Sol matches Mythos Preview's performance while using approximately one-third of the output tokens. On ExploitGym, a benchmark developed by UC Berkeley researchers with OpenAI and other labs, all three GPT-5.6 models show improved performance as reasoning effort increases, suggesting room for scaling with additional compute. Claude benchmarks for this test are not yet available.
OpenAI positions Sol as a cybersecurity defender rather than an attacker. The model excels at spotting and fixing vulnerabilities but does not autonomously execute full end-to-end attacks. In tests with Chromium and Firefox, Sol identified bugs and exploitation primitives but never produced an autonomous full-chain exploit. According to OpenAI's Preparedness Framework, GPT-5.6 Sol remains below the "Cyber Critical" threshold.
Pricing is set at $5 per million input tokens and $30 per million output tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. OpenAI has also redesigned its prompt caching system with explicit cache breakpoints and a guaranteed minimum lifetime of 30 minutes. Cache writes cost 1.25 times the regular input price, while cache reads retain a 90 percent discount.
Since Sol uses fewer tokens to match or exceed competitor performance across multiple benchmarks, the effective cost per task could be lower than previous generations. This addresses the trend of AI models becoming more expensive with each release—a frequent industry criticism and a competitive disadvantage against cheaper Chinese alternatives.
Sol is scheduled to go live on Cerebras in July at speeds up to 750 tokens per second.