GPT-5.6 Sol, Terra, and Luna: How to Choose

OpenAI’s GPT-5.6 family has three tiers: Sol for the hardest work, Terra for balanced everyday tasks, and Luna for cost-efficient volume. The short answer is to start most workloads on Terra, push proven high-volume routines down to Luna, and reserve Sol for tasks where its extra capability measurably cuts failures or review time. The right route depends on your prompts, tools, output length, and acceptable error rate, so treat OpenAI’s numbers as a map and your own test set as the territory.

What changed

OpenAI launched GPT-5.6 for general availability after a limited preview, across ChatGPT, Codex, and the API. The company describes Sol as its new flagship, Terra as “a lower-cost model with performance competitive with GPT-5.5,” and Luna as its fastest and most affordable model.

One naming detail matters for anyone building routing logic. OpenAI says the number identifies the generation, while Sol, Terra, and Luna are durable capability tiers that can advance on their own cadence. In practice, that means a tier name is not a fixed model. Pin specific model identifiers in production and record which version produced which result.

Pricing also moved quickly. At launch, OpenAI listed per-1M-token prices of $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. On July 30, OpenAI cut Luna’s price by 80% and Terra’s by 20%. On August 21, it dropped Sol’s API and credit pricing by over 20% for the next three months. Any cost model built in early July is already out of date, which is the strongest argument for keeping prices in configuration rather than in code.

Launch figures at a glance

The table below uses OpenAI’s published launch prices and a few figures from its own benchmark tables. These are vendor-reported results on vendor-selected evaluations, not our tests.

Measure (OpenAI-reported)	Sol	Terra	Luna
Launch input price per 1M tokens	$5	$2.50	$1
Launch output price per 1M tokens	$30	$15	$6
Artificial Analysis Intelligence Index v4.1	58.9	55	51.2
Artificial Analysis Coding Agent Index v1.1	80	77.4	74.6
Terminal-Bench 2.1	88.8%	87.4%	84.7%
OSWorld 2.0 (computer use)	62.6%	50.2%	45.6%

On coding and terminal tasks, the gap between tiers is a few points. On computer use, the gap is much wider: Sol leads Terra by more than 12 points on OSWorld 2.0. If your workload involves operating software interfaces, the cheaper tiers may cost more in failed runs than they save in tokens.

New settings and API features

GPT-5.6 also added ways to spend more compute on demand. OpenAI says max gives the model more time than xhigh to reason, check, and revise. A setting called ultra coordinates four agents in parallel by default, trading higher token use for stronger results and faster time to result. In the API, developers can build similar experiences with the multi-agent beta in the Responses API.

Programmatic Tool Calling in the Responses API lets the model write and run programs that coordinate tools and filter intermediate data, and OpenAI says it is Zero Data Retention compatible. Prompt caching changed too: explicit cache breakpoints, a 30-minute minimum cache life, cache writes billed at 1.25x the uncached input rate, and cache reads at a 90% discount.

Why it matters

Tier choice is now a bigger lever on cost than it used to be. With launch output prices five times apart between Sol and Luna, sending every request to the flagship is an expensive default. But a cheaper model can still be the more expensive choice if it creates more retries, longer outputs, or more human correction.

OpenAI’s customer quotes point at the same trade-off. Notion’s co-founder said many agents running GPT-5.5 perform just as well on Terra “for half the cost and 16% fewer tokens.” That is one customer’s claim about its own agents, but it shows the kind of saving routing can unlock when a smaller tier is good enough.

Caching rules change the math as well. A cache write now costs more than a normal input token, so caching only pays off when a prompt prefix is reused enough times within its life. Teams with long, stable system prompts should model this explicitly instead of assuming caching is always a discount.

Total cost should include input and output tokens, cache writes and reads, tool calls, latency, failed runs, and reviewer time. A model that is cheaper per token but needs a second pass is not cheaper per finished task.

What to test

Build a task set with three bands. The first holds deterministic work such as field extraction and format conversion. The second holds analysis with known source material. The third holds ambiguous, long-running work with several constraints, including any computer-use or browsing tasks. Our guide to running a repeatable model comparison covers how to structure and date that suite.

Then work through this checklist:

• Run Sol, Terra, and Luna on every task, and set the acceptance rule before you look at outputs.
• Record cost per accepted result, not cost per call. Include retries and reviewer minutes.
• Check current prices on OpenAI’s pricing page before each cost estimate, since the tiers have already been repriced.
• Test effort settings separately. A lower tier at higher effort may beat a higher tier at default effort for your tasks, or it may not.
• If you rely on caching, measure your real hit rate against the 30-minute minimum cache life and the 1.25x write cost.
• Confirm which tiers and settings your ChatGPT or Codex plan actually exposes. OpenAI says Free and Go users get Terra in ChatGPT Work and Codex, while paid plans can choose among all three.
• Store model identifiers and evaluation dates with every result, and rerun the suite after a model update or major prompt change.

Route each task to the least expensive tier that passes your rule consistently. Escalate only after a failure or when the risk of an error justifies the higher price.

What to watch next

Watch whether Sol’s temporary price cut becomes permanent after its three-month window, and whether the tier names get new underlying models before the next generation number. Both would change routing decisions without any change to your code, which is exactly why the evaluation suite matters more than any single launch table. For a parallel example of price cuts reshaping tier choice at another vendor, see our look at Claude Haiku 5.5 pricing and model selection.

The conclusion

Use Luna for proven, high-volume routines, Terra as the general starting point, and Sol where its added capability, especially on computer use and long, multi-step work, removes enough failure or review to justify the price. OpenAI’s benchmarks suggest the tiers are closer on coding than on interface work, but those are vendor figures. Your acceptance rate on your own tasks is the number that should decide the route.