Gemini 3.6 Flash: Previous Pricing and the 3.7 Upgrade

Gemini 3.6 Flash was Google's speed-tier step between Gemini 3 Flash and Gemini 3.1 ProGoogle's frontier model in Cursor for work that combines images and code, especially UI implementation and visual code analysis. Press Enter for the full definition. in Cursor. Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition. now succeeds it as the current speed-tier choice, while Cursor's pricing table keeps 3.6 listed but hidden by default. This page preserves the older model's documented behavior and rates for teams that still need a comparison or policy reference.

Cursor rated 3.6 Flash on three axes: fast speed, low cost and high intelligence. Those labels explain its original role, but they no longer make it the default Flash recommendation. Open the Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition. guide for a new evaluation.

High is not the top of Cursor's own vocabulary here. Claude Opus 5's card reads frontier on that same axis, and Cursor tiers that one medium on speed and high on cost. The safe reading of a fast, low-cost, high-intelligence card, I think, is that this model will hold a multi-step task together, and that Cursor is not pointing at it for the hardest reasoning in the picker.

Cursor publishes three strengths for this model, and the three turn out to be one argument: keeping a fast model useful once a task stops being a single edit. The list below is Cursor's, paired with what each line changes while you work.

Two of those three lines only pay off together.

A million-token ceiling is affordable to use because cached input reads at $0.15 per million, and the reasoning claim is what makes it worth putting multi-step work through the cheap tier at all. On one fresh pass through a large repository, you get the ceiling and none of the discount.

Cursor does not publish a limitations list here, the way it does for Claude Opus 5, so there is no first-party statement of where this model falls down. My first instinct was to read the tiering as a stand-in, and that is only half right: the tiering tells you where Cursor placed the model and nothing about how it behaves. Opus 5's card names a specific habit, over-elaborating in long sessions. Treat the gap as missing information and find the edges yourself, on work you can check.

Cursor's own page states the context limit two ways, and the two are not the same claim. Both figures are below exactly as published, because a large-repository workflow planned around the wrong one is expensive to discover mid-session.

The ceiling is the wrong number to design toward, whichever of the two is live on your plan. A Cursor field engineer's rule of thumb puts the first slip in quality at 50 to 60% of a filled window, which is why our CLI guide says to compact before that point rather than after it.

A big window can also hurt on its own terms. More code in view means more for the model to read past, and the file that matters gets easier to lose. The cache discount makes that trap cheaper to walk into, since keeping the whole repository loaded costs a tenth of what fresh input would. Keep the working set to what the task actually needs, and watch the context indicator the way you would on an expensive model.

Usage lands in the third-party 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, which Cursor meters apart from the 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 and which Claude Opus 5 and the rest of the third-party picker bill against too. The allowance included there runs $20 on Pro, $70 on Pro Plus and $400 on 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 Cursor Models, so it does not include Gemini 3.6 Flash. All rates below are per million tokens.

The cache-read rate is the one to plan around. At $0.15 against $1.50 for fresh input, repeated context costs a tenth of what it cost the first time, and Cursor frames the discount around exactly that case: a codebase you keep re-sending rather than a scattering of one-off prompts.

One cell in that table is a dash rather than a number. Cursor has not published a cache-write rate for this model, so do not read it as zero when you model the cost.

The allowance is what actually runs out, and a low rate only changes how fast. At $1.50 input against Opus 5's $5, the same pool lasts roughly three times as long. On an individual plan, you add on-demand usage at the same rates or move up a tier. Teams first moves a member who empties 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. to the 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; continuing with Gemini 3.6 Flash uses on-demand billing when enabled. Auto remains a Router choice with mode-specific billing, not the fallback pool.

It saves up to 90% on the input side, only on the input side, and less than that once a team plan adds the Token Rate. On an individual plan cached input reads at $0.15 per million against $1.50 fresh, so a session that keeps re-sending the same repository converges toward a tenth of what its first pass cost.

On Teams and Enterprise the arithmetic moves, and not in the direction the 90% headline suggests. Naming this model on a request adds the Cursor Token Rate of $0.25 per million, and Cursor states that fee covers cached tokens as well as input and output. Add it to both sides and the spread narrows from ten to one down to $0.40 against $1.75, a little over four to one. Budget from the second pair if your account is on either plan.

Switching to Auto clears the fee in one of its three modes only. Cursor exempts Auto Cost and the 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, while Auto Balance and Auto Intelligence carry the rate on any request the router sends to a third-party model, and this is one. A team that moves to Auto for cost reasons and leaves everybody on Balance keeps paying the $0.25 whenever a request lands outside the Cursor Models pool.

The discount is also easy to lose by accident.

Switching models mid-session breaks the cache. Which is the part that catches people out: a thread that jumps to a Pro model for one hard step and comes back pays fresh input rates on the way in, and the tenth-of-the-price arithmetic quietly stops describing your session. Unrelated one-shot prompts miss the discount for a duller reason, since there is nothing there to re-read.

If a bill lands higher than the rate card implied, start at the usage dashboard, which shows usage and token breakdowns instead of one total. On an Enterprise account there is a second line to rule out, since Cursor prices regional data residency as a 10% uplift on Model pricing for the models it lists as eligible.

Yes on all three published rates, though how much cheaper depends on which rate your work leans on. Cursor's pricing index lists Gemini 3.1 ProGoogle's frontier model in Cursor for work that combines images and code, especially UI implementation and visual code analysis. Press Enter for the full definition. at $2 input, $0.20 cache read and $12 output, against $1.50, $0.15 and $7.50 for Gemini 3.6 Flash.

Line those up and the input columns sit close together: fifty cents per million between them on fresh input, five cents on cached input. The output rates carry the real distance, $7.50 against $12. So the gap between these two is mostly a function of how much code they write, not how much they read, and a cached-heavy session narrows it to a few cents per million read. Decide on how much reasoning the work needs, then look at the output column, because that is where the difference shows up on the invoice.

Start new speed-tier work with Gemini 3.7 FlashGoogle's latest speed-tier model in Cursor for high-throughput coding, with a 1M maximum context window and a 90% cached-input discount. Press Enter for the full definition.. Keep 3.6 only when you need to reproduce an older result, validate a migration, or honor an approval that names the exact model. The older positioning below still explains what a 3.6 result was designed to do.

Model choice moves cost and reasoning depth. Every model in the picker gets the same agent tools, this one included: file and folder search, reading and editing, shell commands, the browser, fetching rules.

Team size changes who makes this call more than it changes the answer. On a small team it stays per-person, and the correction usually arrives late, once somebody's allowance is gone for the month. On Teams or Enterprise the same pick collects the Token Rate, and Cursor Router becomes available, so the tradeoff can be settled per request instead of per session. An admin chooses which routing modes members can select from Auto, so that call lands on everyone at once.

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. at $0.50 input and $2.50 output is the other end of that argument, and it bills against the 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, which Cursor describes as carrying significantly more included usage. Where the work is routine enough that reasoning depth is not the question, that is the cheaper pool to spend.