GPT-5.6 in Cursor: Luna, Terra, and Sol Explained

OpenAI's GPT-5.6 generation is available in Cursor as three named sizes rather than a single model. Cursor calls them the Sol, Terra, and Luna family: they share the same tool-calling and agent support, so any of them can drive an agent loop, but they trade intelligence against price. The three cards below map that trade before the pricing detail.

All three sit in the OpenAI provider group and draw from the Other ModelsCursor's pool of supported third-party models from providers such as Anthropic, Google and OpenAI, with separate plan access and usage from Cursor Models. Press Enter for the full definition. pool. Pro includes $20 of that usage, with more on Pro Plus and Ultra. Cursor StartAn India-only individual plan centered on Cursor Models, with cloud and mobile access but without Other Models, Auto, Bugbot, Automations or the Agent SDK. Press Enter for the full definition. includes only the separate Cursor ModelsCursor's plan and usage pool for models it labels as Cursor Models, kept separate from the third-party Other Models pool. Press Enter for the full definition. pool: Grok 4.6A Cursor and SpaceXAI frontier model for complex coding and knowledge work; it is a selectable model and is separate from the Grok Bot product. Press Enter for the full definition., Grok 4.5, and Composer 2.5The current Composer release, better at long-running tasks and at judging when a job needs a light touch versus deep work. Press Enter for the full definition.. It does not include GPT-5.6 models.

Switching between the three, though, takes no agent tooling away from you. Each has access to all agent tools in Cursor and each does reasoning, tool calling and agent loops, so the only dial you move is how much thinking you are paying for. Cursor's page for Sol does name one difference that is not about price, describing its communication as clean and direct with less comment and final-message slop than many Claude models, which reads to me as a house observation rather than a measured result.

Cursor prices every GPT-5.6 model per million tokens, split across input, cache write, cache read and output. The table lists the published rates so you can see the whole spread from Luna up to Sol before deciding anything.

Read down the columns and the family still follows one price curve, but the gaps are now wide. Terra costs 10 times Luna in every column. Sol costs 25 times Luna and 2.5 times Terra. That makes Luna a genuinely different budget tier, not a small discount. If a task costs one dollar on Luna, the same token mix costs ten dollars on Terra and twenty-five on Sol.

Which multiplier you are sitting on decides how fast the included allowance drains. Cursor's own sizing guidance says limited agent users often stay inside the included $20 a month, puts daily agent users at $60 to $100 of total usage, and puts people running several agents or automations past $200. Those bands cover the whole model picker rather than this family, so they are probably better for sanity-checking a budget than for building one.

Pick by how much open-ended reasoning the task needs. Luna handles cheap, repeatable work; Terra is the everyday middle; and Sol earns its 2.5-times premium over Terra on problems that genuinely need more thinking. The table sorts common tasks by that test.

The failure people hit is quieter than the one everyone warns about. Defaulting to Sol is the loud mistake and it shows up on the invoice: 25 times Luna's rate on a batch loop, for reasoning the loop never needed. The version I would actually watch for is the opposite move. Pinning Luna to protect the budget looks disciplined, and Cursor states plainly that Luna carries smaller capacity for long, open-ended reasoning chains, so on the wrong task it returns something that looks finished and is not. Budget a review pass whenever you pin Luna to open-ended work.

My default here used to be the cheapest model that can do the job. Checking the output is part of the job, though, and that moves the default: unfamiliar work now starts on Terra for me rather than Luna. Cursor puts Terra between the other two on intelligence and price, so a first pass there shows which direction to move without much riding on the guess.

Each GPT-5.6 model has a companion Fast tier for priority processing. It is the same model on a faster queue, billed at twice the standard rates. The rows below show the naming.

Double those rates and the tiers no longer overlap the next model exactly. Luna Fast costs $0.40 input and $2.40 output, one fifth of Terra Standard. Terra Fast costs $4 input and $24 output, which is 20% below Sol Standard's $5 and $30. Sol Fast remains $10 input and $60 output. Fast now buys priority without erasing the price gap between model sizes.

That leaves one question to answer before you tick Fast on Terra: was it latency you needed, or intelligence? Sol Standard now costs 25% more than Terra Fast, and Cursor rates Sol higher on hard coding and reasoning. Fast earns its premium when waiting is itself the cost: an interactive session you are sitting in front of, or a batch that blocks something else. Left on out of habit, it still doubles the rate.

Sol is the one Cursor describes as persisting through multi-hour agent sessions, and its pricing changes once a single request carries a very large prompt. The callout has the threshold and what it does to the rate.

Put numbers on it and a Sol request over that line runs $10 per million input and $45 output, against $5 and $30 below it. Cursor's pages for Luna and Terra state no equivalent threshold, though a docs page staying silent is weaker evidence than a rule written down, so check rather than assume the whole family behaves alike.

Nor does the threshold survive a move to another provider. Claude Opus 5 documents the opposite: every prompt bills at base rates, including above 300k, with no long-context multiplier at all. So a mental model built on Sol will over-budget Opus, and one built on Opus will under-budget any Sol request that crosses 272k, by double on the input side. If a GPT-5.6 bill lands higher than the rate card implied, the usage dashboard is where I would start, since both pools show up there and in editor settings, and a Sol thread that went over the line is the first thing to rule out.

GPT-5.6 spend runs through Cursor's third-party pool. Other ModelsCursor's pool of supported third-party models from providers such as Anthropic, Google and OpenAI, with separate plan access and usage from Cursor Models. Press Enter for the full definition. access starts with Pro, and eligible individual plans come with a monthly allowance against it. The pairs below summarize how the metering works.

On an individual plan, the first four rows are the whole story.

Add a team and the last row starts to matter, plus a lever the table cannot show. Third-party requests on Teams and Enterprise carry the Cursor Token Rate of $0.25 per million on top of the model rate, and that page works through why the fee bites hardest on the cheapest model here. Cursor Router, meanwhile, exists only on those two plans.

Cursor has published a figure that puts the router directly against this family. In the Cursor Router launch measurements, pinned GPT-5.6 Sol matched Auto Intelligence on cost per commit at $6.76 while scoring lower on user satisfaction with the output, and Auto Balance came in at $4.63 with satisfaction comparable to Sol at a lower spend rate. Those are first-party measurements from enterprise accounts with thousands of users, so a small team has no basis for expecting the same numbers on its own traffic. If you are on Teams or Enterprise, run the comparison against Auto before you standardize everyone on one of these three.

Because the family shares one pool and one allowance, moving a task from Sol down to Terra or Luna stretches the monthly budget without leaving the GPT-5.6 generation. How well that works depends on how often the cheaper model has to be asked twice, and that is a count you have to keep yourself. Watch for it across a billing cycle before you make it a habit.