Three weeks ago, the White House asked OpenAI to slow down. Today, the US Department of Commerce gave the green light — and GPT-5.6 is now publicly available.
The release comes in three tiers. Sol is the flagship, built for agentic coding, cybersecurity, and complex multi-step workflows. On TerminalBench 2.1, Sol Ultra scored 91.91% and base Sol reached 88.76%, edging out Claude Mythos 5 at 88.0% and the previous GPT-5.5 at 83.4%. Terra sits at 84.27% — matching Claude Fable 5's coding performance — priced at $2.50/$15 per million tokens, roughly half Sol's cost ($5/$30). Luna is the budget option at $1/$6, meant for high-volume, cost-sensitive use cases.
The standout new feature is ultra mode: instead of a single model working through a task linearly, a cluster of subagents runs in parallel. It's designed for complex, open-ended work that previously required manual orchestration. Prompt caching got an update too — minimum 30-minute cache life with a 1.25x write billing multiplier.
Speed is the other story. OpenAI is deploying Sol on Cerebras infrastructure at up to 750 tokens per second, though at a steep premium: $12.50/$75 per million tokens. Access is currently limited to select customers as capacity expands.
The AI flagship market is tighter than ever. Grok 4.5 launched yesterday. Claude Mythos 5 is the benchmark to beat. GPT-5.6 Sol sits less than a percentage point above it. For builders, that's the best possible situation: the gap between the top models is shrinking, prices are falling, and picking the right tool is increasingly about fit rather than raw supremacy.
