GPT-5.6 Luna vs Terra vs Sol: How to Pick the Right Tier

GPT-5.6 went generally available on July 9, 2026 — not as one model but as three tiers: Luna, Terra, and Sol, cheapest to priciest, with the input price now spanning a 25x range after the July 30 price cut. Faced with three options, most people either grab the middle one or nervously pay for the top. Independent measurements say both habits waste money: the performance gap between tiers changes dramatically by task, and it is nowhere near proportional to the price gap.

Three names, one family

These are the prices after July 30, when OpenAI cut Luna by 80 percent and Terra by 20 percent (Sol unchanged) from launch prices of 1/6 and 2.50/15. Each tier is directly callable via its own API identifier. The naming shift matters in itself: instead of generations and variants, one generation now contains three explicit tiers, and your daily decision becomes which tier fits this job rather than whether to upgrade.

What the tiers share

Surprisingly much. All three offer a context window of about one million tokens, a 128,000-token output cap, and a knowledge cutoff of February 2026. Choosing Luna does not shrink the documents you can feed it. The only real differences are reasoning depth, quality, and price — which simplifies the decision to a single question: how hard do you need it to think? Get that one wrong, though, and you pay for it directly.

A 25x price gap, a 9-point intelligence gap

Artificial Analysis scores the tiers 52, 55, and 61 as of early August 2026. The cost-per-task column is their measured spend per index task before the price cut; after it, Luna and Terra run cheaper still. So the price varies wildly while the composite intelligence score moves 9 points. Whether 52-to-61 is worth 25x the input price depends on your task — but it is plainly not proportional.

Coding compresses the gap

On the same firm’s Coding Agent Index, measured on Codex, the nine-point spread shrinks to five. Luna delivers about 94 percent of Sol’s coding score — at roughly 20 percent of the cost in the pre-cut measurements, and a fifth of that again at current prices. Terra posts about 96 percent of Sol’s score. Why does coding flatten the curve? Because agentic coding is scaffolded: the harness finds files, applies edits, runs tests, and feeds back results. The model’s job is to turn that crank reliably, which depends less on raw reasoning depth than open-ended problem solving does. The composite index, by contrast, includes hard questions where depth pays directly.

There are domains where that 6 percent matters — architectural decisions, cross-cutting changes. But for routine refactors, test scaffolds, and code reading, the volume work of programming, lower tiers suffice. Tier selection is not about giving good work a good model; it is about routing high-volume work to cheap capacity. Running everything on the top tier is not buying performance — it is paying a fee for skipping the decision.

What OpenAI itself says

OpenAI positions Sol as its best coding model and its strongest cybersecurity model, targeting enterprise work, coding, science, and security. Terra is framed as GPT-5.5-competitive at a lower price — at launch it cost 2.50 in against the older model, and the July 30 cut widened that edge. OpenAI also claims a 54 percent token-efficiency gain for coding; note that Artificial Analysis measured only about 6 percent fewer output tokens per task on its composite index (15,000 versus 16,000 for GPT-5.5). Different measurements, different conditions — treat them as separate data points rather than a contradiction to resolve.

Access by plan

The API exposes all three tiers with no waitlist. In ChatGPT, Plus, Pro, Business, and Enterprise users get Sol at medium depth or above, with GPT-5.6 Sol Pro available on Pro and Enterprise for maximum quality. In Work mode and Codex, free and Go users are assigned Terra while paid tiers choose freely. Many users have never checked which tier they are actually on — worth a look before optimizing anything. For the previous lineup as a baseline, see our ChatGPT guide from May.

The two weeks before launch

GPT-5.6 did not appear from nowhere on July 9. It ran as a trusted-partner preview from June 26, and TechCrunch reported that the US administration had requested rollout restrictions in June over misuse concerns. We cannot say how much that moved the launch date — but the practical lesson stands regardless: availability is set by more than benchmarks, and tight coupling to one vendor imports someone else’s policy risk into your workflow, a topic we cover in our piece on dependency risk.

How to choose: a working procedure

Start at Terra. It is the middle of the range and posts about 96 percent of Sol’s coding score at a fraction of the cost; make it the default and move only with a reason. Drop to Luna for high-volume routine work — classification, summarizing, formatting, code reading — where difficulty is low and repetition is high. Step up to Sol when the cost of a wrong answer is high: architectural choices, changes spanning many files, bugs with no obvious cause.

When in doubt, compare the cost of a human redoing the work against the cost of the higher tier. Sorting a thousand reviews into positive and negative tolerates errors you can catch later — Luna. Estimating the blast radius of a config change on a live system does not — Sol. A new feature close to existing code sits in between — start with Terra and escalate if quality disappoints. In practice, write down your three most frequent tasks and assign each a tier once; a fixed table beats fresh judgment calls every time, and sharing it aligns a whole team.

Against the rest of the field

On the Artificial Analysis composite, early August 2026 reads: Claude Opus 5 at 63, Claude Fable 5 at 62, GPT-5.6 Sol at 61, Kimi K3 at 60, Claude Opus 4.8 at 57. Sol sits one point under Claude Fable 5 at roughly a third of the per-task cost — and with the top four inside three points, arguing over ranks is pointless. Change the yardstick and the order flips anyway: Epoch AI’s Epoch Capabilities Index puts GPT-5.6 above Claude Opus 5. Pick one or two metrics close to your actual work and judge there; the full month’s context is in our July timeline.

At the workbench

For makers, Luna at post-cut prices absorbs the volume work: triaging print-failure logs, reading rough dimensions off photos, extracting numbers from filament data sheets into tables. Sol or Terra takes the design core — parametric enclosure code with interference checks, tolerance suggestions with reasoning — because errors there surface only after a print, costing material and hours. And since all three tiers share the million-token context, a two-stage pattern works nicely: let Luna digest the whole project folder and summarize, then hand only the distilled decision to a higher tier. The expensive tokens go to the short judgment, not the long reading. One caveat: with a February 2026 knowledge cutoff, ask about newer printers or filaments by attaching the manufacturer’s documents yourself. The broader workflow map is in our AI 3D design roadmap.

What tiering hands you

Shared specs, three depths, a 25x price range, a 9-point composite gap that shrinks to 5 for coding: the asymmetry is the whole story. Tiering means you lose money by not choosing and save money by choosing. Default to Terra, push volume down to Luna, reserve Sol for irreversible calls — that three-way split alone improves both the bill and the results for most users. Rankings will keep reshuffling; a tier table matched to your own work will not.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.