OpenAI has launched GPT-5.6 Sol, a new flagship model positioned to challenge Anthropic‘s Claude Mythos, though access remains restricted to select partners under US government rules the company describes as unsustainable.
The release introduces a three-tier hierarchy: Sol, Terra, and Luna. Sol acts as the premium tier, while Terra matches the performance of the previous GPT-5.5 generation at half the cost. Luna serves as the budget option. This structure mirrors the naming conventions used by Anthropic.
OpenAI states it has built Sol specifically to compete with the Mythos class. The limited preview is available via the API and Codex platforms only because the US government directed this restriction. Anthropic previously had its Mythos-class model, Fable 5, removed from the market by the same authorities.
The company’s response to the access limitations is direct. “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.”
Performance benchmarks suggest Sol matches or exceeds Claude Mythos 5 in specific areas. In agentic coding, Sol leads the comparison. In cybersecurity, the two models perform at similar levels.
Performance against Claude Mythos
On Terminal-Bench 2.1, which measures coding capability, Sol scores 88.8 percent. The Sol Ultra variant reaches 91.9 percent. Claude Mythos 5 scores 88 percent, while Fable 5 trails at 84.3 percent.
The model also shows improvements in biology. On GeneBench v1, a test for genomics and quantitative biology, Sol outperformed GPT-5.5 with a 30 percent score against a best-case 22 percent. It achieved this while consuming fewer tokens.
Security testing on ExploitBench reveals Sol matches the performance of the Mythos Preview model. However, Sol uses roughly a third of the output tokens required to reach the same results. ExploitBench tests how well AI agents identify and exploit real security flaws in Google’s V8 JavaScript engine.
On ExploitGym, a benchmark developed by UC Berkeley researchers alongside OpenAI and other labs, all three GPT-5.6 models show improvement as reasoning effort increases. This suggests the models can scale with more compute. Anthropic has not yet released numbers for this specific benchmark.
OpenAI describes Sol as its most capable cybersecurity model to date, emphasising its role as a defender rather than an attacker. The system excels at spotting and fixing flaws but does not run full end-to-end attacks autonomously. Anthropic’s Mythos model achieved that capability in a different test.
During tests involving Chromium and Firefox, Sol identified bugs and exploitation primitives. It did not produce an autonomous full-chain exploit. OpenAI notes that GPT-5.6 Sol remains below the “Cyber Critical” threshold in its Preparedness Framework.
Pricing and technical updates
Costs are calculated per million tokens. Sol charges $5 for input and $30 for output. Terra charges $2.50 for input and $15 for output. Luna charges $1 for input and $6 for output.
OpenAI has updated 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. Cache reads continue to receive a 90 percent discount.
Because Sol uses fewer tokens to match or beat competitors across several benchmarks, the effective cost per task may be lower than previous generations. This approach counters the trend of AI models becoming more expensive with each release and addresses a competitive weakness against cheaper Chinese alternatives.
Sol is scheduled to launch on Cerebras hardware in July, capable of processing up to 750 tokens per second.
What it means
Developers and security teams can now access a model that claims to handle complex coding and security tasks more efficiently than current leaders. However, the ability to use these capabilities is currently controlled by government policy. OpenAI argues this restriction harms the ecosystem by limiting access for those who require the tools for defence and development work.
