Gemini 3.6 Flash in Cursor: Speed-Tier Pricing and Cached Context

Gemini 3.6 Flash is Google's newest speed-tier model in Cursor's picker. Cursor's positioning is precise about where it sits: between Gemini 3 Flash and Gemini 3.1 Pro on price, with reasoning improvements over earlier Flash models. The recommended use is high-throughput coding tasks that need stronger reasoning than Gemini 3 Flash, but at a lower cost than the Pro models.

It is one of the few models in the picker that Cursor tiers as fast and low-cost while still rating its intelligence high, so the model card is worth reading before the pricing.

Cursor publishes three strengths for this model, and all three point at the same job: keeping a fast model useful once the task stops being a single edit. The list is Cursor's, with what each one changes in practice.

Cursor does not publish a limitations list for this model the way it does for Claude Opus 5, so there is no first-party statement of where it falls down. Treat the absence as missing information rather than as an endorsement, and read the tiering instead: a model Cursor rates fast and low-cost is not the one it points at for the hardest reasoning in the picker.

This is worth pinning down because Cursor's own page states it two ways, and the two are not the same claim. Both figures are reproduced below exactly as published, because planning a large-repository workflow around the wrong one is an expensive mistake to discover mid-session.

Gemini 3.6 Flash draws from the third-party Other Models pool, the same pool as Claude Opus 5 and the other third-party models, and separate from Cursor's first-party pool. Individual plans include at least $20 of Other Models usage each month, more on higher tiers. All rates are per million tokens.

The cache-read rate is the number to plan around. At $0.15 per million against $1.50 standard input, cached context costs a tenth of fresh context — which is why Cursor frames the discount as the reason to use this model on repeated large-codebase context rather than on one-off prompts. A session that re-sends the same repository context many times converges toward the cached rate; a session made of unrelated one-shot prompts never gets the benefit.

The clearest way to place this model is against its two neighbours, since Cursor defines it by exactly that gap. Each row below follows from Cursor's stated positioning rather than from a general preference.

Like every model in Cursor's picker, it has access to the full agent tool set — searching files, reading and editing them, running shell commands, browsing, fetching rules. The model choice changes cost and reasoning depth, not what the agent is allowed to do.