GPT-5.6 Is Here: Sol, Terra and Luna Explained

OpenAI shipped GPT-5.6 today, and the headline is not a single model — it is a structure. The generation arrives as three durable capability tiers: Sol, the flagship frontier-reasoning model; Terra, the balanced mid-tier; and Luna, the lightweight fast one. Each tier exposes the full reasoning-effort dial, from low up to xhigh. And as of this morning, all three are running in our arena — each at low and max effort — against the same one-shot prompts every other model in the roster received.

This post covers what the tiers are for, how the pricing ladder works, and where to watch them build the same apps live. What it deliberately does not contain is scores: our tallies are community votes and they move in real time, so the live leaderboard is the source of truth, not a paragraph frozen at publish time.

One generation, three durable tiers

The naming is the most consequential part of the launch. Instead of a flagship plus afterthought minis, OpenAI is positioning Sol, Terra and Luna as durable tiers — capability brands meant to persist across generations rather than one-off model names. The intent is that you standardize on a tier, not a version: Sol for the hardest reasoning problems, Terra as the balanced default for production work, Luna where speed and volume matter more than depth.

The pricing ladder is unusually clean. Terra costs exactly half of Sol on both input and output, Luna sits below half of Terra, and every tier keeps the same 6-to-1 output-to-input ratio, which makes cost projections straightforward. Top to bottom the spread is 5× — wide enough that picking a tier is a real engineering decision with a budget attached, narrow enough that trying the tier above is never absurd.

Effort is the second axis

Every tier supports reasoning-effort levels from low to xhigh, so “GPT-5.6” is not one price-performance point but a grid: tier × effort. Effort controls how many reasoning tokens the model spends before it answers, and as we found when we dug into thinking-effort levels across other families, the gap between the same model at low and high effort can be bigger than the gap between two adjacent tiers. Whether Terra at max effort beats Sol at low effort is exactly the kind of question a spec sheet cannot answer — and a same-prompt arena can.

Now in the arena: six configurations, one shot each

Starting today, GPT-5.6 is benchmarked in our arena at low and max effort for each tier — six new configurations in the roster. The rules are the same ones every model gets: one-shot generation from the exact same published prompt, no retries, no cherry-picking, and the output runs live in your browser across 40+ coding and design challenges — landing pages, arcade games, physics sims, emulators. You interact with the artifact the model actually produced, then vote blind before the labels reveal.

• The launch matchup: GPT-5.6 Sol vs Claude Fable 5 — the new flagship against Anthropic’s top-shelf model, on the landing-page challenge.
• The generational check: GPT-5.6 Sol vs GPT-5.5 — the cleanest way to see whether the version bump changed what actually ships.
• Or open the coding arena and build your own pairing — any GPT-5.6 tier at either effort level against anything else in the roster.

Which tier for which job

• Sol ($5/$30) — for work where correctness on a hard problem justifies flagship pricing: gnarly algorithms, multi-step refactors, long-horizon agentic runs.
• Terra ($2.50/$15) — the sensible default for production coding assistants and agents: half of Sol’s price, the same effort dial.
• Luna ($1/$6) — classification, extraction, scaffolding, high-volume pipelines; cheap enough that raising its effort is often a better trade than moving up a tier.

Those are starting points, not verdicts. For the cross-vendor version of the decision — GPT against Claude, Gemini and the open-weight cluster — our standing guide to the best AI for coding tracks how arena results shift it.

Watch the same app get built three ways

A spec sheet tells you Sol is the flagship. Watching Sol, Terra and Luna take the identical prompt and ship three different apps in front of you tells you what that actually means — where Luna cuts corners, where Terra is indistinguishable from its bigger sibling, and whether Sol’s premium shows up in the output or only on the invoice. The runs are live now. Pick a challenge, watch, vote — the tally is everyone’s votes combined, updated as they land.