Previewing GPT‑5.6 Sol: a next-generation model
4 hours ago
- #Cybersecurity
- #Technology Release
- #AI Models
- Limited preview of GPT-5.6 series launched, including flagship Sol, balanced Terra (2x cheaper than GPT-5.5), and fast/affordable Luna.
- Sol features enhanced safety stack with strengthened protections for high-risk activities, cyber requests, and misuse; underwent extensive pressure-testing.
- Models to be generally available soon; initial limited preview for trusted partners shared with U.S. government, aiming for broader access while developing cyber Executive Order framework.
- Sol shows improved agentic capabilities in coding, biology, and cybersecurity; introduces new 'max' reasoning effort and 'ultra' mode with subagents for complex tasks.
- Sets new state-of-the-art on Terminal-Bench 2.1 for coding; achieves stronger results on GeneBench v1 with fewer tokens; competitive on ExploitBench and ExploitGym for cybersecurity.
- Robust safeguards include model-level refusals, real-time misuse classifiers, account-level review, and layered protections; designed to benefit defensive work while constraining offensive use.
- No single safeguard sufficient; over 700,000 GPU hours dedicated to automated red teaming for universal jailbreaks, complemented by human expert testing.
- Pricing per 1M tokens: Sol $5 input/$30 output, Terra $2.50/$15, Luna $1/$6; includes predictable prompt caching with 30-minute minimum cache life.
- Sol to launch on Cerebras in July at up to 750 tokens/second for select customers initially; models available via API and Codex to trusted partners first.