OpenAI has unveiled GPT-5.6 Sol, its newest flagship model positioned to compete directly with Anthropic’s Claude Mythos class. The launch, however, comes with a significant caveat: access is limited to select partners through API and Codex, at what the company explicitly describes as “the explicit direction of the US government.”
The restricted rollout follows a pattern established with Claude Mythos 5, which the US government forced Anthropic to disable worldwide in June. OpenAI frames the access limitations as a temporary measure, but acknowledges uncertainty about when—or if—broader release will occur.
Performance Benchmarks
GPT-5.6 Sol shows meaningful improvements over previous generations in several key areas:
Agentic Coding: On Terminal-Bench 2.1, Sol scores 88.8%, while Sol Ultra reaches 91.9%. Claude Mythos 5 scores 88.0%, and Gemini 3.1 Pro Preview trails at 70.7%.
Cybersecurity: On ExploitBench, which tests autonomous vulnerability discovery, Sol matches Mythos Preview performance while using roughly one-third of the output tokens. OpenAI characterizes the model as a “defender” rather than attacker, noting it stays below the “Cyber Critical” threshold in the company’s Preparedness Framework.
Biology Research: On GeneBench v1, Sol achieves 30% success versus GPT-5.5’s 22% best case, while consuming fewer tokens.
Pricing and Availability
The tiered pricing structure introduces three permanent performance levels:
- Sol: $5 input / $30 output per million tokens
- Terra: $2.50 / $15 (matches GPT-5.5 capability at half cost)
- Luna: $1 / $6 (budget option)
Additional reasoning modes include “max” for deeper analysis and “ultra” for parallel sub-agent orchestration. Prompt caching receives an overhaul with explicit breakpoints and guaranteed 30-minute cache lifetime, with cache writes at 1.25x input price and reads receiving a 90% discount.
Industry Implications
The Sol launch highlights an emerging tension in frontier AI development: capability advances increasingly trigger regulatory scrutiny before public release. OpenAI’s blog post states directly: “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 model’s Cerebras deployment is scheduled for July, potentially expanding access beyond current API limitations.