Gemini 3.7 Flash API Cost, Measured: Tasks Bill 2.5-8x Less Than 3.6

Gemini 3.7 Flash bills the same tasks 2.5x to 8x cheaper than Gemini 3.6 Flash, and only half of that is the advertised price cut. The intro rate of $0.75 per million input and $3.75 per million output is 50% off 3.6, and on our meter the same 6.9K-token prompt billed exactly half ($0.00529 vs $0.01057). The other half is quieter: gemini-3.7-flash burned 26% to 77% fewer thinking tokens than 3.6 on the same four measured tasks. We probed the model on day two of its release: the pricing terms, the thinking dial (one position died), the compatibility breaks Google’s docs warn about and the ones they undersell, plus cache, context, and tokenizer continuity.

TL;DR

• Intro pricing $0.75/$3.75 runs through December 31, 2026, then $1.50/$7.50; the same prompt billed exactly half of 3.6 on our meter.
• Default thinking dropped 26-77% versus 3.6 (144 vs 384 tokens on a 5-step task), so per-task bills fall 2.5-8x, not just 2x.
• The off-switch is gone: every off spelling returns a 400 (“Thinking level is unsupported”); low is now the floor.
• A two-turn tool loop halved 3.7’s deliberation on the tool-result turn (32 vs 61 tokens); refusing an unanswerable question cost both models 500.

What does Gemini 3.7 Flash actually cost?

Half of 3.6’s rate until December 31, 2026, and less than half of 3.6’s tokens on every task we measured. Google prices the launch at $0.75 per million input and $3.75 per million output as an introductory rate; from January 1, 2027 it reverts to $1.50/$7.50, exactly 3.6’s card. Cache reads are $0.075 per million, 10% of intro input. The meter agrees with the card: the same 6.9K-token prompt cost $0.0052875 on 3.7 and $0.0105675 on 3.6, a clean halving.

The larger effect is what the model no longer spends. Identical salted tasks, three runs each, median reasoning tokens and the resulting output bill:

Google’s launch line is that 3.7 “thinks more diligently”; on our tasks diligence means fewer tokens, not more. All accuracy cells stayed 3/3 on both models. Multiply the halved rate by the halved-or-better thinking and the practical upgrade math is a 61-88% smaller bill per task, before caching. The one caveat sits in the calendar: budget against the January 1 reversion, the same pattern as Sonnet 5’s intro pricing.

One task shape resists the trend, and it is the one nobody budgets for: refusing. Asked about five fabricated entities (a company, an institute, a town charter, an alloy, an award), both models declined all five, and both spent more thinking doing it than on any other task we measured, a median 511 reasoning tokens on 3.7 and 494 on 3.6. Saying “there is no record of this” cost 3.5x the thinking of solving the 5-step arithmetic chain. Retrieval-augmented pipelines that frequently hit missing data pay this tax on every miss, and it is the one place 3.7’s efficiency gain disappears. It also puts a measured floor under a launch-week worry: reviewers flagged a higher hallucination rate for 3.7, but on fabricated-entity questions it hedged 5 of 5, identical to 3.6.

Which thinking controls survive on 3.7?

Three positions and no off, exactly as documented: low, medium (the default), high. What the docs do not spell out is that 3.6’s escape hatches are dead. Every off spelling we sent, reasoning_effort: "none", "minimal", thinking_budget: 0, thinking: {"type": "disabled"}, enable_thinking: false, returned the same upstream 400: “Thinking level is unsupported: THINKING_LEVEL_MINIMAL”. A second, independent request path refuses in plainer words, “Reasoning is mandatory for this endpoint and cannot be disabled”, so this is the model, not one client’s translation layer. On 3.6-flash, measured in the same batch, none and minimal still zero the burn. The Flash line has joined the pro tier’s policy: thinking cannot be turned off, and 3.6 is now the last Flash that can.

The surviving dial, on the 5-step task (median of 3, all 3/3 correct):

Two working notes. The default-is-medium claim checks out: unconfigured runs in the dial matrix burned 154, indistinguishable from its explicit medium arm at 147. And thinking_budget is deprecated for real: every nonzero value we sent (16 through 1,024) was accepted and burned the same 135-138 tokens, a no-op, where 3.6 still maps small budgets to off and larger ones to caps. If your 3.6 integration steers cost through budgets, the 3.7 translation is: pick a level, that is the whole interface. For the cross-vendor context on which dials do anything, see our thinking-controls matrix.

What does the agent pitch cost per step?

Tool selection is unchanged from 3.6; what got cheaper is the deliberation between steps. We ran a two-turn loop against two functions (look up an incident, then restart the service it names), three runs per model:

Both models chose get_incident first and restart_service after, 3/3, with identical prompt tokens. The gap is the second turn, where 3.7 thinks about half as much before committing to the next call. That is the step an agent repeats: a 20-step loop at these medians bills about $0.0043 of output on 3.7 against $0.0131 on 3.6, and the ratio holds after the January price reversion because it comes from tokens, not rates. This is the mechanism behind Google’s agentic benchmark jump being pitched as a cost story: fewer deliberation tokens per hop, multiplied by many hops.

Where the model lands against rivals is benchmark-dependent and worth reading before switching a coding agent over: independent roundups put 3.7 Flash ahead on FrontierCode (43.6% vs Sonnet 5’s 42.7%) and near 3x Sonnet 5 on AutomationBench, while GPT-5.6 Terra leads Terminal-bench and Sonnet 5 leads the desktop-task exam, per the same roundup and launch-week coverage. The cost floor is where 3.7 is unambiguous.

What breaks when you upgrade from 3.6?

Two hard 400s and several things that fail quieter than the docs suggest. Google’s migration note says to remove temperature, top_p, top_k, candidate_count, and prefilled model turns. Measured:

The asymmetry is worth noticing: the thinking dial validates loudly (an unsupported level is a clean 400 naming the value), while sampling params and output caps swallow whatever you send. If your client library sets temperature by default, nothing breaks today; if it prefills assistant turns for guided output, it broke before you upgraded.

Do cache, context, and tokenizer carry over?

Yes, with one slowdown. The implicit cache hits at the same shape as 3.6, 4,076 of 6,905 tokens cached on the second call of an identical prompt, cutting the call’s cost 52%, and reads price at 10% of intro input. But the build lag grew: 3.6 served a hit 4 seconds after the priming call, 3.7 returned nothing at 4 seconds and hit at 30. Rapid-fire duplicate traffic lands before the cache is warm; the write-side economics are otherwise unchanged.

Context accepted 708,912 input tokens in one call and answered a needle question correctly, consistent with the 1M claim and with no long-context price tier on the rate card. The text tokenizer is byte-identical across four generations, gemini-3.1-pro-preview, 3.5-flash, 3.6, and 3.7 all counted the same mixed English-Chinese-code corpus at 50 tokens, so token budgets port unchanged. Image inputs still bill the flat 1,089 tokens at every size we measured across the Gemini line. Neither generation returns its reasoning text: reasoning_content came back empty on every call, so the thinking you pay for stays invisible on both, a continuing complaint in launch-week threads. Structured output holds: strict json_schema returned valid, correct JSON in 3/3 runs, and reasoning_effort: "low" drove the extraction’s thinking to zero, the same single-step safe zone our dial matrix found everywhere.

FAQ

How much does the Gemini 3.7 Flash API cost?

$0.75 per million input tokens and $3.75 per million output as an introductory rate through December 31, 2026; from January 1, 2027 it reverts to $1.50/$7.50, the same card as Gemini 3.6 Flash. Cache reads are $0.075 per million. On identical prompts our meter billed 3.7 at exactly half of 3.6.

Can you turn off thinking on Gemini 3.7 Flash?

No. Every off spelling returns a 400 (“Thinking level is unsupported”); the dial is low, medium (default), high, and low still burned 133 reasoning tokens on our 5-step task. Gemini 3.6 Flash remains the newest Flash where reasoning_effort: "none" works, at 3.7’s post-intro price.

Is Gemini 3.7 Flash actually cheaper than 3.6 per request?

Cheaper than the 50% headline suggests: it also thinks 26-77% less on the same tasks, so our measured per-task bills dropped 61-88% (a 2-hop word problem fell from $0.0020 to $0.0002). After December 31 the rate halves back to parity and only the thinking savings remain.

Will code written for Gemini 3.6 Flash run on 3.7?

Mostly. Sampling params (temperature, top_p, top_k) are still accepted despite the docs’ removal list, and oversized max_tokens is silently clamped. The two hard breaks: prefilled assistant turns return a 400 (as they already did on 3.6), and every thinking off-switch now 400s, so cost steering must move from thinking_budget to the three levels.

Measured 2026-08-15 through the Synthorai gateway, two days after release: dial and off-switch matrix (7 effort values, 5 budgets, 2 off params, n=3, gemini-3.6-flash rerun in the same batch for every comparison), four-task reasoning-tax sweep, strict-JSON structured output, deprecated-parameter probes, output-cap and 708K-context acceptance, implicit-cache pairs at two wait times, a fixed-corpus tokenizer comparison across four Gemini generations, a two-turn function-calling loop (n=3 per model), and a five-question fabricated-entity refusal probe. Task dollar figures price measured completion tokens at each model’s current output rate. Two claims were cross-checked on a second, independent request path: the oversized max_tokens acceptance and the absence of an off position. Intro pricing and dates are Google’s published terms; behavior may change as the rollout matures.

← Back to blog