OpenAI confirms GPT-5.6 Sol, Terra, and Luna drop this Thursday

While everyone has been waiting for AI to get marginally smarter, OpenAI just changed the game on cost and speed.

OpenAI announced the rollout of three new model variants in the GPT-5.6 family: flagship Sol, mid-tier Terra, and cost-optimized Luna. The lineup is launching publicly across ChatGPT, Codex, and the API, following an expanded global preview.

Halving the bill while beating benchmarks

OpenAI clearly built this lineup around efficiency and real-world agent work. Flagship Sol scored 53.6 on the Agents' Last Exam (beating Claude Fable 5 by 13.1 points) and hit 80 on the Artificial Analysis Coding Agent Index, edging past Fable 5 by 2.8 points.

More importantly, OpenAI claims Sol achieves roughly half the cost, half the output tokens, and half the latency. Testers putting 25 billion tokens through the model reported seeing it build a Minecraft-style game autonomously over days, spend six days refining an Excel clone via Computer Use, and manage live Supabase instances in the browser without explicit prompts.

The two-week regulatory hold

Behind the scenes, the release reportedly arrived after a two-week government delay. How frontier models like Sol actually get cleared for public release remains completely opaque.

Researchers told TechCrunch they have virtually no visibility into the evaluation criteria. While OpenAI says Sol underwent safety testing with third parties including the UK AI Security Institute (AISI), SecureBio, and Irregular, and Sam Altman confirmed discussions with senior U.S. officials, neither side has published a concrete framework.

Why it matters

We think OpenAI is making the right bet at the right time. For months, enterprises and developers have hit a wall where complex multi-step agents simply burn too many tokens and cost too much money to justify in production. By slashing latency and output token counts in half while keeping top-tier reasoning, Sol makes running autonomous background agents economically viable.

If you build software with AI agents, testing Luna for routine work and Sol for complex execution is an immediate no-brainer. For Anthropic and Google, the pressure is no longer just matching benchmark intelligence; it is about matching cost-performance on a per-task basis.

The real friction ahead is trust and governance. Pushing powerful agentic capabilities into production while clearing regulatory gates through private meetings rather than open standards creates an uncomfortable precedent for accountability across the entire industry.