Google DeepMind

Gemini 3.7 Flash

The $0.75/$3.75 launch price is real, and all six benchmark scores on this page are independent, third-party runs — it's Google's own listed rate and every score is independently verified — but that price is a four-and-a-half-month introductory offer that doubles to $1.50/$7.50 the moment 2027 starts.

Gemini 3.7 Flash’s 6 benchmark scores on this page were each verified against their sources on or after 2026-08-17.

Released
2026-08-13
License
proprietary
Context window
1M tokens
Knowledge cutoff
2026-03

The verified record

Against the 76 head-to-head comparisons Gemini 3.7 Flash shares with other tracked models: 16 real gaps, 25 inside the noise band, and 35 we will not call.

A gap counts for Gemini 3.7 Flash only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Gemini 3.7 Flash trails on 2 of them.

HLE · no toolsReasoning

±2 is noise
Ahead
Behind
Claude Fable 5 7.6 · Claude Opus 5 7.0
Tie
4 models within ±2

LiveCodeBenchContest coding

±3.1 is noise
Ahead
GLM-5.2 +19.2 · GLM-5.3 +8.2 · Claude Sonnet 5 +6.3 · GPT-5.6 Sol +6.1
Tie
9 models within ±3.1

DeepSWELong-horizon coding

±9.5 is noise
Ahead
Tie
7 models within ±9.5
Unverified
2 models — vendor-reported on one side

No verdict for Gemini 3.7 Flash anywhere on Terminal-Bench 2.1 (nothing independently confirmed on both sides); GPQA Diamond, SWE-bench Verified (saturated).

  • None of Gemini 3.7 Flash’s agentic comparisons are independently confirmed on both sides yet.

Gemini 3.7 Flash API pricing

$0.75 in / $3.75 out per 1M tokens official pricing

Gemini 3.7 Flash is one of 2 Google DeepMind models tracked on this site, at these official list prices.

Google DeepMind model pricing, official list rates
ModelIn / 1MOut / 1M
Gemini 3.7 Flash$0.75$3.75
Gemini 3.1 Pro Preview$2.00$12.00

Gemini 3.7 Flash benchmark scores

Gemini 3.7 Flash benchmark scores, provenance, and source links
BenchmarkScore
HLE(no tools)[1]
Reasoning · ±2 is noise
SWE-bench Verifiedsaturated[2]
Bug fixing — not ranked at any gap size
GPQA Diamondsaturated[3]
Expert science Q&A — not ranked at any gap size
LiveCodeBench[4]
Contest coding · ±3.1 is noise
DeepSWE[5]
Long-horizon coding · ±9.5 is noise
Terminal-Bench 2.1[6]
Terminal ops · ±10.6 is noise

Who ran these numbers: 6 of 6 independent — vals.ai (3), artificialanalysis.ai (2), deepswe.datacurve.ai (1).

  1. HLE: AA's own run at high effort, text-only subset. Beats Grok 4.6 here despite a lower overall AA index — single benchmarks and composites disagree, which is rather the point of this site.
  2. SWE-bench Verified: vals.ai run, bash-only harness (updated 2026-08-14).
  3. GPQA Diamond: vals.ai run (93.94), rank 4 (updated 2026-08-15).
  4. LiveCodeBench: vals.ai run (88.652), rank 3 — just ahead of Grok 4.6 (updated 2026-08-15).
  5. DeepSWE: 65%±2 on mini-swe-agent, high effort (board updated 2026-08-13).
  6. Terminal-Bench 2.1: AA's own run of 'Gemini 3.7 Flash (high)' (0.857677902621723 in AA's payload). Not on tbench.ai's official board, which lists only Gemini 3 Pro and 3.1 Pro.

Notes on the record

Introductory pricing through 2026-12-31; standard rate from 2027-01-01 is $1.50/$7.50, per Google's Gemini API pricing docs (https://ai.google.dev/gemini-api/docs/pricing and https://ai.google.dev/gemini-api/docs/latest-model). Output price includes thinking tokens; Google's docs describe this as three selectable reasoning tiers (low/medium/high) rather than a raw thinking-token budget.

GA per the official Gemini API changelog (2026-08-13), superseding Gemini 3.6 Flash. Knowledge cutoff March 2026 per the official DeepMind model card, which warns coverage in some domains only reaches January 2025. Google's Pro line (3.1) is still in preview — the Flash line is its only GA track. The discounted introductory rate also applies to the older Gemini 3.6 Flash, per Google's pricing page.

Per DeepMind's model card, 3.7 Flash is built on the Gemini 3.6 Flash base with algorithmic reasoning improvements, not a new pretraining run. The GA release drops support for deprecated sampling parameters (temperature, top_p, top_k, and candidate_count) that worked on earlier Gemini versions, per Google's migration/changelog docs — existing integrations calling those fields need updating.

Compare with

FAQ

Has Gemini 3.7 Flash been independently benchmarked?

Yes — every one of Gemini 3.7 Flash's six tracked scores is independently sourced. HLE, Terminal-Bench 2.1, SWE-bench Verified, GPQA Diamond, LiveCodeBench, and DeepSWE are each independent third-party runs — via Artificial Analysis, Vals.ai, and the DeepSWE leaderboard — observed 2026-08-17 and 2026-08-20, with none of the headline numbers on this page sourced from Google's own launch materials.

How much does Gemini 3.7 Flash cost, and will the price change?

$0.75 per million input tokens and $3.75 per million output tokens (thinking tokens are billed as output) — but that is introductory pricing, good only through 2026-12-31. From 2027-01-01 the rate rises to $1.50 in / $7.50 out per million tokens, per Google's Gemini API pricing page. The same discounted rate currently also applies to the older Gemini 3.6 Flash.

Does Gemini 3.7 Flash have a generally available Pro-tier sibling?

No. As of August 2026, Google's Pro line (Gemini 3.1 Pro) is still shipping only as a preview offering, not GA. Per Google's own materials, the Flash line — the track Gemini 3.7 Flash belongs to — is currently Google's only generally-available Gemini line.

What is Gemini 3.7 Flash's knowledge cutoff?

Google's official DeepMind model card lists March 2026 as the nominal cutoff, but explicitly warns that in some domains actual coverage only reaches back to January 2025 — so treat 'March 2026' as an upper bound rather than a guarantee for any specific topic.

What breaks for developers upgrading from Gemini 3.6 Flash to Gemini 3.7 Flash?

Beyond the benchmark gains, Google's own migration documentation notes that Gemini 3.7 Flash drops support for older sampling parameters — temperature, top_p, top_k, and candidate_count — that worked on earlier Gemini versions, so any integration relying on those fields needs to be updated before switching over.

Get the next verdict by email

One email per verdict — which launch-chart claims held up. No spam, unsubscribe anytime.