GPT-5.6: Sol, Terra and Luna — OpenAI's new model family explained

GPT-5.6: Sol, Terra and Luna — OpenAI's new model family explained

One generation, three models

On July 9, 2026, OpenAI released GPT-5.6 — but not as a single model. This time the generation ships in three flavors: Sol, Terra and Luna. It's a deliberate break from how OpenAI has named its models so far, and it says a lot about where the industry is heading.

The idea is simple but deliberate: the number (5.6) marks the generation, while the name marks a capability tier that can be updated on its own cadence. Sol can get a successor without Terra or Luna moving at all, and vice versa. That solves a problem many of us recognize — picking the "right" model has gotten harder with every release.

Sol, Terra and Luna — what sets them apart

• Sol is the flagship. Built for heavy reasoning and long-horizon agentic work — the model OpenAI calls its most capable to date.
• Terra is the balanced model. Performance on par with GPT-5.5, at half the cost — a natural default for everyday coding and agentic tasks.
• Luna is the fast, cheap variant. Built for smaller tasks at scale, where latency and price matter more than raw capability.

All three share the same technical foundation: a 1.05 million token context window, up to 128,000 tokens of output, and the same knowledge cutoff (February 16, 2026). You also pick a reasoning effort per request — from none to max — so the same model can be tuned for anything from quick answers to deep problem-solving.

Image: mikemacmarketing, Flickr (CC BY 2.0)

Pricing and performance

At launch, Sol cost $5 per million input tokens and $30 per million output tokens. Terra was priced at $2.50/$15, and Luna at $1/$6. Just three weeks later, on July 30, 2026, OpenAI cut Luna's price by 80% and Terra's by 20% — Sol stayed unchanged. A sign of how fast competition among the cheaper model tiers is pushing prices down.

Worth noting: send more than 272,000 tokens in a single request and a surcharge kicks in — double the input price and 1.5x the output price for the entire request, not just the overage.

On benchmarks like Terminal-Bench 2.1 (agentic coding), Sol edges out both Claude Mythos 5 and GPT-5.5 — narrowly, but measurably. Luna, the cheapest model in the family, still scores 62.7% on SWE-Bench Pro, which says a lot about how much capability now fits into the "fast and cheap" tier.

What the naming change actually means

This is the interesting part. By separating generation from capability class, OpenAI can update a single tier without forcing everyone to switch models at the same time. If you build a system around Terra today, the idea is that the next "Terra" should feel like a natural upgrade — not a new name you have to research from scratch.

It echoes how Anthropic has split its own models into Opus, Sonnet and Haiku — just with a generation number layered on top. If you want to see how Claude's equivalent tiers, pricing and benchmarks stack up, we've put it together in our guide to Claude.

What it means if you're building with AI

• Pick the tier for the task, not the hype. Sol for complex agentic work, Terra for everyday coding, Luna for volume.
• Budget for the long-context surcharge if you're feeding large documents or codebases through the model — it can triple your bill if you don't plan for it.
• Watch for price changes. A three-week gap between launch and discount shows the cost picture is still moving fast.
• Build with an abstraction layer. Whether you run OpenAI or Claude, a thin integration layer makes it easy to switch tiers — or providers — when the next model lands.

The model race keeps going, and the names keep multiplying. What matters is understanding what each tier is actually built for — not just what it's called.