GPT-5.6 Released — OpenAI's Sol, Terra, Luna Variants Explained

GPT-5.6 is the biggest AI news of the month. OpenAI shipped it in limited preview with three named variants and a release process unlike anything we've seen before: a US government review gate determines who can use it. This is genuinely new territory.

The three variants — what each one is for

• Sol — flagship model. Maximum capability across every benchmark. Slower and most expensive. Use it for hard problems where quality is everything: research, complex agentic flows, multi-hour reasoning.
• Terra — balanced. Roughly Sol's capability on most tasks, at a fraction of the cost and latency. The default for most workloads. Most people will use Terra 90% of the time.
• Luna — fast and cheap. Smaller, distilled from Sol. Designed for high-volume work where speed matters: chat, autocomplete, classification, simple RAG.

What's actually improved

• Coding — Sol scores around 92% on SWE-Bench Verified (up from ~79% on GPT-5), making it competitive with Claude Opus 4.8 for the first time.
• Long reasoning — Sol's extended thinking mode now runs to roughly 90 minutes (up from ~10 min on GPT-5), closing the gap with Anthropic.
• Cybersecurity — meaningful jumps on red-team benchmarks; OpenAI explicitly highlights this as a focus area.
• Scientific tasks — physics, chemistry, molecular biology benchmarks up significantly. Sol is now the default for serious scientific research workflows.
• Agent workflows — tool-use reliability jumped to ~95%, multi-step task completion is now in usable territory.

The US government review process

This is the part of the release that nobody saw coming. GPT-5.6 access is gated by a US government review process — OpenAI confirmed the gate exists but won't disclose specifics on what triggers approval. Reading between the lines: capability-based dual-use review, similar to what NIST and the AI Safety Institute have been signalling for 18 months. Approved organisations include major US government contractors, big banks, and a handful of named research labs.

What it means for everyone else

• You probably can't use GPT-5.6 today. ChatGPT Pro users get a "preview access" toggle for Terra; Sol and Luna are not yet broadly available.
• API access is restricted to approved organisations only — no waitlist, no self-serve. If you're not pre-approved, the API returns "model not available".
• Expect broader rollout in tiers: first to US-domiciled paying organisations, then EU/UK with separate regulatory review, then general availability.
• In the meantime: GPT-5 remains fully available. Claude Opus 4.8 and Claude Mythos 5 (for approved orgs) are the alternatives if you need frontier capability today.

How to think about Sol vs Terra vs Luna

If you do get access, the three-variant strategy mirrors what Anthropic, Google, and Meta all converged on: a tiered family. Sol for the hardest 5% of problems. Terra for everything else. Luna for high-volume cheap work where latency matters. Most teams will route 90% of traffic to Terra and reserve Sol for specific high-value flows.

Pricing (estimated — official numbers pending)

• Sol — ~$15/M input, ~$90/M output (parity with Opus 4.8 territory)
• Terra — ~$5/M input, ~$25/M output (mid-tier, cheaper than current GPT-5)
• Luna — ~$0.25/M input, ~$1.25/M output (Haiku / Gemini Flash territory)
• Prompt caching: 50% discount on cache reads (down from 75% rumoured)

Bottom line

GPT-5.6 is a real jump in capability — coding, reasoning, agents, science all measurably better. But the restricted rollout makes it less of a "what should I switch to" announcement and more of a "what's coming when it opens up." For now: GPT-5 and Claude Opus 4.8 remain the daily drivers. Watch the rollout closely.