GPT-5.6 Sol: OpenAI’s Next-Gen Model with Enhanced Safety and Capabilities
OpenAI announced GPT-5.6, a new model family including Sol (flagship), Terra (balanced), and Luna (fast/affordable). The article frames the central tension: delivering a major leap in agentic capabilities—especially in coding, biology, and cybersecurity—while managing safety risks and navigating government oversight. OpenAI chose a limited preview with trusted partners, sharing plans with the U.S. government, to balance broad access against the need for robust safeguards before general availability. The company explicitly states this government access process should not become the default, but sees it as a short-term step toward broader release.
The technical path focuses on layered safety and new reasoning modes. Sol introduces a max reasoning effort and ultra mode using subagents for complex tasks. On benchmarks, Sol sets a state of the art on Terminal-Bench 2.1 for coding, outperforms GPT-5.5 on GeneBench v1 for biology, and matches Mythos Preview on ExploitBench with ~1/3 the output tokens. Safety measures include model-level refusal training, real-time classifiers, account-level review, and automated red-teaming using over 700,000 A100-equivalent GPU hours. OpenAI also emphasizes that Sol does not cross the Cyber Critical threshold despite improved vulnerability research capabilities.
For builders, the key takeaway is the deliberate phasing and pricing strategy. GPY-5.6 models are priced per 1M tokens: Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6 (input/output), with improved prompt caching (30-minute min cache life, 1.25x cache write cost, 90% cache read discount). Sol will also launch on Cerebras at up to 750 tokens/second in July. The article signals that safety is a moving target: safeguards may block legitimate dual-use work, and feedback during the preview will refine them. Enterprise features like privacy-preserving detection and customer-operated controls are in development.
