Gemini 3.7 Flash
Gemini 3.7 Flash came out on Aug 13, 2026 and became the previous generation on Sep 2, 2026, when 3.8 Flash arrived at the same price: US$ 0.75 per million input tokens and US$ 3.75 per million output until Dec 31, 2026, and twice that from 2027. It remains fully supported, with no shutdown date announced. It is worth keeping when cost per task matters more than the score: 3.8 scores higher on all 14 of Google's tests, but spends more tokens per task.
Gemini 3.7 Flash in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Google DeepMind
- API identifier
gemini-3.7-flash(stable, GA)- Release
- Aug 13, 2026
- Status
- previous generation since Sep 2, 2026; no shutdown date announced
- Recommended replacement
gemini-3.8-flash, at the same price- Input
- US$ 0.75 per million until Dec 31, 2026; US$ 1.50 from Jan 1, 2027
- Output (includes reasoning)
- US$ 3.75 per million until Dec 31, 2026; US$ 7.50 from Jan 1, 2027
- Cache read
- US$ 0.075 per million (US$ 0.15 in 2027), plus hourly storage
- Batch and Flex
- half price: US$ 0.375 and US$ 1.875 until Dec 31, 2026
- Context window
- 1,048,576 tokens
- Max output
- 65,536 tokens
- Reliable knowledge
- through March 2026; in some domains, through January 2025
- Reasoning
- low, medium (default) and high; minimal returns an error
- Input and output
- text, image, video, audio and PDF in; text only out
- Free tier
- yes, on the Gemini API and Google AI Studio
- Where it runs
- Gemini API, Google AI Studio, Google Cloud, Antigravity, Android Studio, the Gemini Enterprise app and the Spark agent in the Gemini app
Where Gemini 3.7 Flash does well, where it falls behind and who it is for
Where it does well
- Same price, fewer tokens per task: Google says 3.8 Flash spends more tokens by design and points to 3.7 when efficiency is the main constraint.
- Cheaper long video: agentic video understanding, released on Sep 1, 2026, uses up to 88% fewer tokens on long-form content, according to Google. On Google Cloud, the feature is listed as not supported on 3.8.
- Safety outside English: in the 3.8 model card, the automated multilingual safety test got 5.4 percentage points worse compared with 3.7.
- A jump over 3.6 Flash: DeepSWE v1.1 from 48.6% to 65.3%, AutomationBench from 17.0% to 30.4% and GDP.pdf from 22.0% to 34.0%, in the launch table.
- Long context: 97.0% on GDM-MRCR v2 at 128 thousand tokens, against 81.5% for Claude Sonnet 5 and 93.5% for GPT-5.6 Terra, in Google's figures.
Where it falls behind
- Below its successor on everything: in the 3.8 Flash table, 3.7 trails on all 14 tests, with the widest gaps in long-horizon coding (65.3% against 73.7% on DeepSWE v1.1) and computer use (50.6% against 59.0% on OSWorld 2.0).
- Long terminal work: 11.2% on Terminal-Bench 4.0, against 19.1% for 3.8 Flash and 51.8% for Claude Opus 5 in the same table.
- Prompt injection: in the Gray Swan chart Google published, attacks succeeded 9.2% of the time on 3.7, against 5.5% on 3.8 Flash.
- Uncertain lifespan: Google Cloud lists it among the shorter availability models, with 3.8 as the replacement. Once a date is set, the minimum notice is 45 days.
- Price doubles in 2027: US$ 1.50 and US$ 7.50 from Jan 1, 2027, as with 3.8 Flash.
Who it makes sense for
- Production apps on 3.7 that already work well and where cost per task matters: switching at the same price per token can make each task more expensive.
- Long video analysis with agentic video understanding, which Google Cloud lists for 3.7 and not for 3.8.
- Customer support in Portuguese or other non-English languages on sensitive topics, where 3.7 has the better multilingual safety score in Google's figures.
- High-volume reading of PDFs and long documents, with 34.0% on GDP.pdf and 97.0% on GDM-MRCR v2, at US$ 0.75 input until the end of the year.
When to pick another model
- New projects: the Gemini API model list already treats 3.7 as the previous generation, and 3.8 costs the same.
- Long-horizon coding agents or agents that operate screens: 3.8 is much stronger on DeepSWE, Terminal-Bench 4.0 and OSWorld 2.0.
- Agents that read email or web pages: 3.8 resists prompt injection better in the Gray Swan chart.
- Fine-tuning: Google Cloud does not offer tuning for 3.7 Flash.
What Google published at launch
3.7 Flash came out on Aug 13, 2026, three weeks after 3.6 Flash, with an introductory price at half the original 3.6 price, according to Google. The focus was coding and agents: more production-ready code, better tool use and closer instruction following.
The table compares 3.7 with 3.6 Flash, Claude Sonnet 5 and GPT-5.6 Terra. 3.7 leads seven of the eleven tests below, but loses to GPT-5.6 Terra in long-horizon coding, terminal work and computer use, and to Sonnet 5 in office work.
A word of caution: in the 3.8 Flash model card, from September, Google published different scores for the same 3.7. GDPVal-AA v2 fell from 1,525 to 1,482, and OSWorld 2.0 rose from 47.9% to 50.6%, the latter with batch tool calls.
| 3.7 Flash | 3.6 Flash | Sonnet 5 Anthropic | GPT-5.6 Terra OpenAI | |
|---|---|---|---|---|
| Input price US$ per million, no caching |
0.75 (1.50 later) | 0.75 (1.50 later) | 2.00 | 2.00 |
| Output price US$ per million |
3.75 (7.50 later) | 3.75 (7.50 later) | 10.00 | 12.00 |
| FrontierCode 1.1 Main code quality |
43.6% | 34.4% | 42.7% | 41.3% |
| DeepSWE v1.1 long-horizon coding |
65.3% | 48.6% | 53.8% | 69.6% |
| Terminal-Bench 2.1 terminal |
85.8% | 78.0% | 80.4% | 87.4% |
| AutomationBench business automation |
30.4% | 17.0% | 10.7% | 23.6% |
| GDPVal-AA v2 office work, Elo |
1,525 | 1,422 | 1,598 | 1,578 |
| GDP.pdf expert PDF |
34.0% | 22.0% | 28.0% | 24.7% |
| Harvey LAB-AA legal work |
90.7% | 85.1% | 90.1% | 85.2% |
| LVBench long video |
85.4% | 84.2% | 68.5% | 78.9% |
| GDM-MRCR v2 long context, 128 thousand tokens |
97.0% | 91.8% | 81.5% | 93.5% |
| OSWorld 2.0 computer use |
47.9% | 33.8% | not published | 50.2% |
| HLE-Verified hard knowledge |
53.6% | 51.2% | 31.0% | 51.1% |
Figures from the Gemini 3.7 Flash model card, August 2026, with the prices Google published in the same table; the 3.7 and 3.6 Flash price is introductory until Dec 31, 2026. The dot marks the best in each row.
Is 3.7 still worth using now that 3.8 costs the same?
For new projects, 3.8 Flash is the starting point: it costs the same per token and scores higher on all 14 tests Google published for both. It wins clearly in long-horizon coding, terminal work and computer use, and resists prompt injection better.
3.7 still makes sense in three cases. First, when cost per task rules: 3.8 spends more tokens by design, and Google itself suggests lowering the effort or staying on 3.7, which remains fully supported. Second, long video: agentic video understanding is listed as supported on 3.7 and not supported on 3.8 in the Google Cloud documentation. Third, support outside English, where 3.8 got 5.4 percentage points worse on the automated safety test.
The risk of staying is the timeline. The Gemini API has not announced a shutdown date, but Google Cloud lists 3.7 among the shorter availability models, with 3.8 as the replacement and at least 45 days of notice once a date is set. If you stay on 3.7, keep the migration ready.
| 3.7 Flash | 3.8 Flash | |
|---|---|---|
| Price until Dec 31, 2026 input and output, per million |
US$ 0.75 and US$ 3.75 | US$ 0.75 and US$ 3.75 |
| Price from 2027 input and output, per million |
US$ 1.50 and US$ 7.50 | US$ 1.50 and US$ 7.50 |
| DeepSWE v1.1 long-horizon coding |
65.3% | 73.7% |
| Terminal-Bench 4.0 long terminal work |
11.2% | 19.1% |
| OSWorld 2.0 computer use |
50.6% | 59.0% |
| GDPVal-AA v2 office work, Elo |
1,482 | 1,545 |
| BioMysteryBench biology, human difficult |
43.5% | 56.5% |
| HLE-Verified hard knowledge |
53.6% | 54.9% |
| Multilingual safety automated test, lower is better |
baseline | 5.4 pp worse |
| Gray Swan prompt injection, lower is better |
9.2% | 5.5% |
| Agentic video understanding on Google Cloud |
supported | not supported |
| Status | previous generation | current |
Tests from the 3.8 Flash model card, from September, where 3.7 appears with scores different from August. Multilingual safety: 3.8 model card. Gray Swan: chart in the 3.8 Flash Cyber announcement. Video: Google Cloud pages for both models, checked on Oct 1, 2026.
What changes when migrating to 3.8 Flash
Price, the 1,048,576 token window, the 65,536 token output, reasoning levels and input types are the same. Google's migration guide for 3.8 asks for these code changes:
- Change the identifier to
gemini-3.8-flash. - Remove
temperature,top_pandtop_kfrom the config; Google Cloud still lists default values for them, so check your platform. - Use
thinking_levelinstead ofthinking_budgetand removecandidate_count. - Standardize multi-turn conversations on
previous_interaction_idand remove prefilled model turns. - Review function calling: in generateContent, every function response needs
call_idandname. - Before switching, run ten real tasks on both models at medium effort and compare accuracy and cost per task, already at the 2027 price, including long video and non-English support if your app has them.
Examples of using Gemini 3.7 Flash
A recording of about 45 minutes with audio, the limit documented on Google Cloud, becomes minutes with decisions and owners. With agentic video understanding, the model fetches only the segments, frames or audio it needs, and Google cites up to 88% fewer tokens on long-form content.
An assistant answers customers in Portuguese, with your own moderation on top. On the automated multilingual safety test, 3.7 scores better than 3.8, according to the successor's model card, and the price per token is the same.
Annual reports and contracts in PDF go into the same request, up to 1,048,576 tokens; on Google Cloud, files of up to 50 MB and 3,000 pages. 3.7 scored 34.0% on GDP.pdf, against 22.0% for 3.6 Flash, and 97.0% on the GDM-MRCR v2 long context test.
Where Gemini 3.7 Flash stands out and where it falls behind, by the numbers
Gemini 3.7 Flash has published results on 23 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Nov 16, 2025 and Nov 11, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 23 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 94.8% | #6 / 79 | #6 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 97.2% | #31 / 81 (tied with 1) | #31 / 168 | 100.0% GPT-5.5 |
| Dtbench | 96.8% | #9 / 68 | #9 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 50.38 | #20 / 68 | #20 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 14.3% | #39 / 78 (tied with 1) | #39 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 59.8% | #8 / 75 | #8 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 71.6% | #25 / 68 | #25 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 84.6% | #11 / 49 | #11 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 904 | #45 / 62 | #48 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,587 | #22 / 73 | #22 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 95.5% | #12 / 49 (tied with 1) | #12 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 69.2% | #10 / 63 | #10 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 66.0% | #5 / 63 | #5 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 37.0% | #17 / 61 (tied with 1) | #17 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 58.0% | #21 / 70 (tied with 2) | #21 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 36.6% | #25 / 58 | #25 / 67 | 100.0% GPT-6.1 Sol |
| Apex Agents | 67.8% | #5 / 49 | #5 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 43.6% | #18 / 43 | #18 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 23.8% | #16 / 37 | #16 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 26.7% | #27 / 34 (tied with 1) | #27 / 34 | 83.3% Claude Opus 5.5 |
| DeepSWE | 65.5% | #12 / 26 | #12 / 26 | 74.1% GPT-6 Astra |
| FrontierSWE | 20.3% | #15 / 20 | #15 / 20 | 65.5% GPT-6 Astra |
| Rli | 5.0% | #8 / 12 | #8 / 15 | 21.3% Claude Opus 5.5 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Gemini 3.7 Flash with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 5, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
Gemini 3.7 Flash or Kimi K3?
Kimi K3 is the model from another lab closest to Gemini 3.7 Flash in capability. Across the 22 tests both took, this is the score:
| Test | Gemini 3.7 Flash | Kimi K3 |
|---|---|---|
| Chess puzzles | 66.0% | 36.0% |
| ALE-Bench | 904 | 1,524 |
| CritPt | 14.3% | 23.4% |
| ProofBench | 58.0% | 87.0% |
| Game puzzles | 37.0% | 26.0% |
| ARC-AGI 2 | 84.6% | 60.4% |
| SimpleQA Verified | 69.2% | 50.6% |
| Apex Agents | 67.8% | 50.6% |
| Furniture Assembly | 26.7% | 34.2% |
| FrontierSWE | 20.3% | 25.9% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Gemini 3.7 Flash costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
DeepSWE
The score scale does not start at zero.
Other models from Google DeepMind
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Gemini 4 Argon | Sep 30, 2026 | 164.56 | 1,608 | n/a | Limited access |
| Gemini 3.8 Flash | Sep 2, 2026 | 157.20 | 1,551 | 75.8 | API access |
| Gemini 3.7 Flash | Aug 13, 2026 | 157.22 | 1,551 | 78.8 | API access |
| Gemini 3.6 Flash | Jul 21, 2026 | 154.24 | 1,530 | 73.6 | API access |
| Gemini 3.5 Flash-Lite | Jul 21, 2026 | 145.15 | 1,467 | 63.9 | API access |
| Gemini 3.5 Flash | May 19, 2026 | 154.38 | 1,526 | 74.6 | API access |
| Gemma 4 31B IT | Apr 2, 2026 | 142.86 | n/a | n/a | Open weights (restricted use) |
| Gemma 4 26B A4B | Apr 2, 2026 | 141.81 | 1,495 | n/a | Open weights (unrestricted) |
| Gemini 3.1 Pro | Feb 19, 2026 | 154.67 | 1,528 | 77.0 | API access |
Have you used Gemini 3.7 Flash? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Gemini 3.7 Flash
How much does Gemini 3.7 Flash cost?
Until Dec 31, 2026, US$ 0.75 per million input tokens and US$ 3.75 per million output, with reasoning billed as output. From Jan 1, 2027, US$ 1.50 and US$ 7.50. Batch and Flex cost half, and cache reads cost US$ 0.075 per million this year. It is the same price as 3.8 Flash in every mode.
Will Gemini 3.7 Flash be shut down?
No date has been announced. On Oct 1, 2026, the Gemini API deprecations page listed no shutdown date for gemini-3.7-flash. On Google Cloud it is among the shorter availability models, with 3.8 Flash as the replacement, and Google promises at least 45 days of notice when it sets a date.
Gemini 3.7 Flash or 3.8 Flash?
For new projects, 3.8 Flash: same price and a higher score on all 14 tests Google published for both. 3.7 is worth it when cost per task matters more, because 3.8 spends more tokens, when the workflow depends on agentic video understanding, and for support outside English, where 3.8 got 5.4 percentage points worse on the automated safety test.
What changes in the code when migrating to 3.8?
Change the identifier to gemini-3.8-flash, remove temperature, top_p and top_k, use thinking_level instead of thinking_budget, remove candidate_count and standardize conversations on previous_interaction_id. Price, context window and max output do not change.
Does Gemini 3.7 Flash have a free tier?
Yes, on the Gemini API and Google AI Studio. On the free tier, Google may use the data to improve its products; on the paid tier, it does not. For customer data, use the paid tier.
Where does Gemini 3.7 Flash rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Gemini 3.7 Flash ranks #18 out of 267 models, with an ECI of 157.22. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Gemini 3.7 Flash better than Kimi K3?
Across the 22 tests both models took and Epoch AI compiled, Gemini 3.7 Flash comes out ahead on 10, behind on 11 and tied on 1. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Gemini 3.7 Flash best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Gemini 3.7 Flash are: GPQA Diamond (#6 of 79); Chess puzzles (#5 of 63); Apex Agents (#5 of 49). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Google, Gemini 3.7 Flash announcement
- Google DeepMind, Gemini 3.7 Flash model card
- Google DeepMind, Gemini 3.7 Flash evaluation methodology (PDF)
- Gemini API, Gemini 3.7 Flash page
- Gemini API, pricing
- Gemini API, model deprecations
- Gemini API, release notes (agentic video understanding)
- Gemini API, latest model guide (migration to 3.8)
- Google Cloud, Gemini 3.7 Flash page
- Google Cloud, Gemini 3.8 Flash page
- Google Cloud, model versions and lifecycle
- Google DeepMind, Gemini 3.8 Flash model card
- Google, Gemini 3.8 Flash and 3.8 Flash Cyber announcement (Gray Swan chart)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
← See the full ranking