GPT-5.6 Sol, Terra and Luna: Which OpenAI Tier Should You Actually Pay For?

Most model releases give you one thing to evaluate. GPT-5.6 gave you three.

OpenAI announced a limited preview on 26 June 2026 and made the family generally available on 9 July: Sol, Terra and Luna — one generation, three capability tiers, spanning a price range that is now roughly 25x from top to bottom.

Then, three weeks later, OpenAI cut two of the three prices and left the third alone. That single decision changed which tier is the right default for most work, and a lot of the coverage written in July has not caught up.

This guide has the current numbers, what each tier is actually for, and the one calculation that should decide your choice — which is not the one most people run.

The naming convention, briefly

This is the release where OpenAI stopped shipping model names that told you nothing.

Under the new scheme, the number is the generation and the name is the tier. GPT-5.6 is the generation. Sol, Terra and Luna are durable capability tiers that can each advance on their own cadence — so a future Terra can improve without waiting for a whole-family version bump.

It is a better system than "4o versus o1 versus 4.1", and it survives contact with the next release: when GPT-6 Astra arrived in September 2026, it slotted in above Sol rather than replacing the naming logic.

The three tiers

OpenAI's own positioning for Terra at launch was that it matched GPT-5.5 — the previous generation's flagship — at roughly half the cost. That framing is the useful one: each tier is roughly a generation's worth of capability cheaper than the one above it.

Shared specifications

The tiers differ in capability and price, not in shape:

This uniformity is the underrated feature. Because context, output ceiling and cutoff are identical across tiers, you can route a request to a cheaper tier without touching your prompts or your retrieval design. Tier selection becomes a runtime decision rather than an architectural one — which is exactly what you want if you are trying to control costs on a live product.

Pricing, after the cut

Per million tokens, standard service tier. Verified 9 September 2026.

What changed on 30 July

OpenAI attributed the cuts to efficiency gains made during GPT-5.6's own development — including the model rewriting and optimising production code and improving token generation.

The consequence matters more than the cause. At launch the spread between Luna and Sol was about 5x. After the cut it is about 25x. Tier selection went from a modest optimisation to the single largest cost lever you have, and any cost analysis written before 30 July is using numbers that no longer exist.

Service tier multipliers

These stack on the list rates above:

Batch at half price is the easiest saving in the table and the one most teams forget. If your work is not latency-sensitive — overnight generation, bulk classification, evaluation runs — you are paying double for nothing.

The 272K cliff

Every tier carries a 1.05M context window, but standard pricing only holds up to 272,000 input tokens. Above that line the entire request re-bills at 2x input and 1.5x output.

In practice:

A 280,000-token request costs about double a 270,000-token one. If you are feeding large document sets, chunk deliberately around that boundary rather than discovering it on an invoice.

The calculation that should decide your tier

Here is where most tier comparisons go wrong. They compare price per token, when the number that matters is cost per completed task.

An independent long-horizon coding benchmark ran both Sol and Terra across 100+ tasks:

Terra used 2.7x more output tokens to reach a lower pass rate. Run the arithmetic at current post-cut prices:

• Sol: 20,968 output tokens × $30/M = $0.63 per attempt; at a 63.7% pass rate that is ~$0.99 per resolved task
• Terra: 55,594 output tokens × $12/M = $0.67 per attempt; at a 40.7% pass rate that is ~$1.64 per resolved task

Sol is roughly 40% cheaper per finished job, despite costing 2.5x more per output token. The July price cut narrowed that gap — at Terra's old $15 output rate the figure was about $2.05 per resolved task — but it did not close it.

The lesson generalises beyond this one benchmark: a weaker model on a task it cannot reliably finish will talk more, retry more, and cost more. Cheap per token is not cheap per outcome. The place Luna and Terra genuinely win is work where the pass rate is near-identical across tiers — classification, extraction, routing, short summaries — and there the 25x price gap is close to free money.

Getting it in ChatGPT

Worth noting for anyone comparing subscriptions: the free tier of ChatGPT now runs Luna, a model that costs OpenAI $0.20 per million input tokens to serve. That is the economic reason free access got more generous — not generosity.

Where GPT-5.6 sits now

GPT-6 Astra replaced Sol as OpenAI's flagship on 4 September 2026 at $10 input and $50 output. That reshuffles the whole ladder rather than retiring anything:

Sol is now the sensible middle of OpenAI's range rather than its ceiling — which is usually when a model becomes interesting to buy. If Astra's premium is not measurably returning value on your workload, Sol is the step down to try before you drop to Terra.

For a cross-vendor view, Claude Sonnet 5 at $2/$10 sits almost exactly on Terra's price point, and Claude Fable 5.1 at $10/$50 matches Astra's. The tiers line up closely enough across vendors that price is rarely the deciding factor — your own evaluation set is.

Practical recommendations

Default to Luna for anything high-volume where you can verify the output cheaply — classification, tagging, routing, extraction, first-pass drafting. At $0.20 per million input tokens the cost is close to a rounding error.

Default to Terra for customer-facing text, support replies, document analysis and content at scale. It is the tier most business workloads should sit on.

Reach for Sol when the task is long, multi-step, or expensive to get wrong — and measure cost per completed task, not per request, before you decide it is too expensive.

Always check Batch. Half price for the same model, for any work that can wait.

Watch the 272K line. It is the most common source of surprise invoices in this family.

Related reading

• GPT-6 Astra Explained — the model that replaced Sol at the top of the range
• Claude Fable 5.1 Explained — Anthropic's competing frontier model
• How to Use OpenAI Codex — OpenAI's agentic coding system
• ChatGPT Review — our full review of the consumer product and its plans
• AI Tool Pricing Index — our verified, regularly re-checked pricing dataset

Pricing, specifications and benchmark figures were verified on 9 September 2026 against OpenAI's published materials and independent trackers. Model pricing in this family has already changed once since launch — check OpenAI's own pricing page before budgeting against these numbers.