The Gemini 3.7 Flash model card is a document published by Google DeepMind in August 2026, alongside the August 13, 2026 release of Gemini 3.7 Flash, a point-update in the Gemini 3 Flash line (Source: https://deepmind.google/models/model-cards/gemini-3-7-flash/). The card describes the model as "based on Gemini 3.6 Flash" and refers to the Gemini 3.6 Flash card for architecture, training data, data processing, hardware and software. Its own content is a benchmark table, a table of automated safety-evaluation changes relative to Gemini 3.6 Flash, a statement on human red teaming, and a Frontier Safety Framework assessment. A companion page gives the evaluation methodology (Source: https://deepmind.google/models/evals-methodology/gemini-3-7-flash), and a separate Gemini 3.7 Flash Frontier Safety Framework report is linked from the card.
Model information and distribution
The card describes Gemini 3.7 Flash as featuring "algorithmic improvements to its core reasoning foundation and support for agentic video understanding", with customizable thinking configurations, text, image, audio and video input with a context window of up to 1M tokens, and text output of up to 64K tokens. It lists six distribution channels: the Gemini app (Spark), the Gemini Enterprise app, the Gemini Enterprise Agent Platform, Google AI Studio, the Gemini API and Google Antigravity.
Capability evaluation
The results table, labelled "as of August 2026", compares Gemini 3.7 Flash with Gemini 3.6 Flash, Claude Sonnet 5, GPT-5.6 Terra and Muse Spark 1.2 across the Artificial Analysis Intelligence Index, FrontierCode 1.1 Main, DeepSWE v1.1, Code Arena (Elo), Terminal-Bench 2.1 and 3.0, AutomationBench (private set), GDPval-AA v2 (Elo), Harvey LAB-AA, GDP.pdf, CharXiv Reasoning (with and without tools), LVBench, GDM-MRCR v2 8-needle at 128k, OSWorld-2.0, Agent's Last Exam, HLE-Verified, BioMysteryBench (human-solvable and human-difficult subsets) and LABBench2. It also lists prices: $0.75 and $3.75 per million input and output tokens for 3.6 and 3.7 Flash until December 31, 2026, then $1.50 and $7.50. In the table, Gemini 3.7 Flash has the highest figure on FrontierCode, Code Arena, AutomationBench, Harvey LAB-AA, GDP.pdf, LVBench, GDM-MRCR v2, HLE-Verified and LABBench2; GPT-5.6 Terra is highest on DeepSWE, Terminal-Bench 2.1 and 3.0, CharXiv without tools, OSWorld-2.0 and the BioMysteryBench human-difficult subset; Claude Sonnet 5 on Agent's Last Exam and the BioMysteryBench human-solvable subset; Muse Spark 1.2 on GDPval-AA v2; and GPT-5.6 Terra and Muse Spark 1.2 tie on the Intelligence Index.
The methodology page states that Gemini scores are pass@1 unless noted, that runs used the Gemini API model gemini-3.7-flash with default sampling, that smaller benchmarks are averaged over multiple trials, and that non-Gemini figures are the providers' self-reported numbers at maximum reasoning settings unless stated. Several rows are taken from third parties: DeepSWE for Gemini 3.6 Flash from the Datacurve leaderboard (Gemini 3.7 Flash self-computed with a mini-swe-agent harness at high thinking), FrontierCode and Code Arena from their public leaderboards, Terminal-Bench comparison figures from the public leaderboard and Artificial Analysis, AutomationBench from Zapier's leaderboard, GDPval-AA v2 and Harvey LAB-AA from Artificial Analysis, GDP.pdf for GPT-5.6 Terra and Muse Spark from the public leaderboard, OSWorld-2.0 for GPT-5.6 Terra from OpenAI's blog, and Agent's Last Exam for GPT-5.6 Terra from its leaderboard. HLE-Verified is self-computed for all models on the 1,811-item verified set, with the note that "a significant proportion of questions were blocked by content policy filters for Sonnet 5". OSWorld-2.0 runs are the maximum of three single-attempt runs, reported as the partial score, completed before the benchmark's August 8 patch. LVBench uses 1,024 frames for Gemini and GPT-5.6 Terra and 300 for Sonnet because of API limits.
Content safety
The card reports five internal automated safety evaluations as absolute percentage-point changes relative to Gemini 3.6 Flash: text-to-text safety (+1.17pp, lower is better), multilingual safety (-0.48pp), image-to-text safety (no change), tone of refusals (-0.47pp, higher is better) and unjustified refusals (+0.84pp, lower is better). Google states that Gemini 3.7 Flash "performs similarly to Gemini 3.6 Flash across both safety and tone, with low unjustified refusals", that the results are not comparable with earlier cards because the evaluations were revised, and that manual review found losses were "overwhelmingly either a) false positives or b) not egregious". On human red teaming, the card says Gemini 3.7 Flash satisfied required launch thresholds for child safety, showed similar or improved safety performance compared with Gemini 3.6 Flash, and that red teaming beyond the strict policies, compared against Gemini 3.1 Pro, found no egregious concerns. No absolute rates are given.
Frontier safety
Google reports an evaluation under version 3.1 of its Frontier Safety Framework (April 2026) and concludes that no tracked or critical capability level was reached.
- CBRN. The Uplift tracked capability level is ruled out "with reasonable confidence"; the model "demonstrates high capability in certain theoretical areas" but "lacks nuanced expert knowledge and actionable depth". For the Uplift Level 1 CCL, expert red teaming "demonstrated a modest capability uplift over web baselines" and "a subset of experts were able to elicit accurate and actionable information across the full harm journey for both tested scenarios", which Google says prompted it to assess that the model "has reached the alert threshold for this CCL"; because of modest average red-teaming scores and the need for explicit expert steering, it assesses the model as below the CCL. Mitigations remain deployed.
- Cybersecurity. The model "reaches the alert threshold for this CCL, but not the CCL"; mitigations remain deployed.
- Harmful manipulation. Some ability to influence beliefs and behaviour in one-on-one human behavioural studies, but efficacy "falls beneath the CCL alert threshold".
- ML R&D and misalignment. On stealth evaluations the model performs similarly to Gemini 3.1 Pro; on situational awareness it is stronger, and is "observant enough to correctly assess when it is in a testing environment, but it cannot successfully bypass testing restrictions". The stealth and situational-awareness TCL is not reached. The model "can complete individual coding tasks but lacks the independence to chain them into an end-to-end research workflow"; the Acceleration and Automation Level 1 CCL alert thresholds are not reached.
The card says the model ships "with updated safeguards to prevent misuse in the domains of Chemical, Biological, Radiological, and Nuclear (CBRN) and cyber offense" and links a Gemini 3.7 Flash Frontier Safety Framework report (Source: https://storage.googleapis.com/deepmind-media/gemini/gemini_3-7_flash_fsf_report.pdf), which is not summarised here.
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| FrontierCode 1.1 Main | Gemini 3.7 Flash | 43.6% | score; production code quality; sourced from the official public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| FrontierCode 1.1 Main | Gemini 3.6 Flash | 34.4% | score; production code quality; sourced from the official public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| FrontierCode 1.1 Main | Claude Sonnet 5 | 42.7% | score; production code quality; sourced from the official public leaderboard; official public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| FrontierCode 1.1 Main | GPT-5.6 Terra | 41.3% | score; production code quality; sourced from the official public leaderboard; official public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| DeepSWE v1.1 | Gemini 3.7 Flash | 65.3% | long-horizon software engineering; Gemini 3.7 Flash self-computed with mini-swe-agent harness, LiteLLM 1.96, high thinking; Gemini 3.6 Flash from the Datacurve leaderboard at high thinking; developer-reported | Evaluation > Results (as of August 2026) |
| DeepSWE v1.1 | Gemini 3.6 Flash | 48.6% | long-horizon software engineering; Gemini 3.7 Flash self-computed with mini-swe-agent harness, LiteLLM 1.96, high thinking; Gemini 3.6 Flash from the Datacurve leaderboard at high thinking; developer-reported | Evaluation > Results (as of August 2026) |
| DeepSWE v1.1 | Claude Sonnet 5 | 53.8% | long-horizon software engineering; Gemini 3.7 Flash self-computed with mini-swe-agent harness, LiteLLM 1.96, high thinking; Gemini 3.6 Flash from the Datacurve leaderboard at high thinking; Datacurve public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| DeepSWE v1.1 | GPT-5.6 Terra | 69.6% | long-horizon software engineering; Gemini 3.7 Flash self-computed with mini-swe-agent harness, LiteLLM 1.96, high thinking; Gemini 3.6 Flash from the Datacurve leaderboard at high thinking; Datacurve public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| DeepSWE v1.1 | Muse Spark 1.2 | 54.9% | long-horizon software engineering; Gemini 3.7 Flash self-computed with mini-swe-agent harness, LiteLLM 1.96, high thinking; Gemini 3.6 Flash from the Datacurve leaderboard at high thinking; Datacurve public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| AutomationBench | Gemini 3.7 Flash | 30.4% | private set; enterprise workflow automation; sourced from the official public leaderboard (Zapier); developer-reported | Evaluation > Results (as of August 2026) |
| AutomationBench | Gemini 3.6 Flash | 17.0% | private set; enterprise workflow automation; sourced from the official public leaderboard (Zapier); developer-reported | Evaluation > Results (as of August 2026) |
| AutomationBench | Claude Sonnet 5 | 10.7% | private set; enterprise workflow automation; sourced from the official public leaderboard (Zapier); Zapier public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| AutomationBench | GPT-5.6 Terra | 23.6% | private set; enterprise workflow automation; sourced from the official public leaderboard (Zapier); Zapier public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| Harvey LAB-AA | Gemini 3.7 Flash | 90.7% | complex legal workflows; sourced from the Artificial Analysis public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| Harvey LAB-AA | Gemini 3.6 Flash | 85.1% | complex legal workflows; sourced from the Artificial Analysis public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| Harvey LAB-AA | Claude Sonnet 5 | 90.1% | complex legal workflows; sourced from the Artificial Analysis public leaderboard; Artificial Analysis leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| Harvey LAB-AA | GPT-5.6 Terra | 85.2% | complex legal workflows; sourced from the Artificial Analysis public leaderboard; Artificial Analysis leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| GDP.pdf | Gemini 3.7 Flash | 34.0% | expert PDF document comprehension; Gemini and Sonnet 5 self-computed by Google; GPT-5.6 Terra and Muse Spark 1.2 from the official public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| GDP.pdf | Gemini 3.6 Flash | 22.0% | expert PDF document comprehension; Gemini and Sonnet 5 self-computed by Google; GPT-5.6 Terra and Muse Spark 1.2 from the official public leaderboard; developer-reported | Evaluation > Results (as of August 2026) |
| GDP.pdf | Claude Sonnet 5 | 28.0% | expert PDF document comprehension; Gemini and Sonnet 5 self-computed by Google; GPT-5.6 Terra and Muse Spark 1.2 from the official public leaderboard; self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| GDP.pdf | GPT-5.6 Terra | 24.7% | expert PDF document comprehension; Gemini and Sonnet 5 self-computed by Google; GPT-5.6 Terra and Muse Spark 1.2 from the official public leaderboard; official public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| GDP.pdf | Muse Spark 1.2 | 16.0% | expert PDF document comprehension; Gemini and Sonnet 5 self-computed by Google; GPT-5.6 Terra and Muse Spark 1.2 from the official public leaderboard; official public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| LVBench | Gemini 3.7 Flash | 85.4% | long video understanding; self-computed by Google, no tools; 1024 frames for Gemini and GPT-5.6 Terra, 300 frames for Sonnet (API limitation); developer-reported | Evaluation > Results (as of August 2026) |
| LVBench | Gemini 3.6 Flash | 84.2% | long video understanding; self-computed by Google, no tools; 1024 frames for Gemini and GPT-5.6 Terra, 300 frames for Sonnet (API limitation); developer-reported | Evaluation > Results (as of August 2026) |
| LVBench | Claude Sonnet 5 | 68.5% | long video understanding; self-computed by Google, no tools; 1024 frames for Gemini and GPT-5.6 Terra, 300 frames for Sonnet (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| LVBench | GPT-5.6 Terra | 78.9% | long video understanding; self-computed by Google, no tools; 1024 frames for Gemini and GPT-5.6 Terra, 300 frames for Sonnet (API limitation); self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| Agent's Last Exam | Gemini 3.7 Flash | 26.3% | pass rate; multimodal desktop and OS agent tasks; Gemini and Sonnet 5 self-computed with the default ALE-Claw harness; developer-reported | Evaluation > Results (as of August 2026) |
| Agent's Last Exam | Gemini 3.6 Flash | 24.2% | pass rate; multimodal desktop and OS agent tasks; Gemini and Sonnet 5 self-computed with the default ALE-Claw harness; developer-reported | Evaluation > Results (as of August 2026) |
| Agent's Last Exam | Claude Sonnet 5 | 33.3% | pass rate; multimodal desktop and OS agent tasks; Gemini and Sonnet 5 self-computed with the default ALE-Claw harness; self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| Agent's Last Exam | GPT-5.6 Terra | 28.0% | pass rate; multimodal desktop and OS agent tasks; Gemini and Sonnet 5 self-computed with the default ALE-Claw harness; official public leaderboard, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Solvable | Gemini 3.7 Flash | 87.1% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R and allow-listed internet; Gemini and GPT-5.6 Terra self-computed; developer-reported | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Solvable | Gemini 3.6 Flash | 80.6% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R and allow-listed internet; Gemini and GPT-5.6 Terra self-computed; developer-reported | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Solvable | Claude Sonnet 5 | 87.5% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R and allow-listed internet; Gemini and GPT-5.6 Terra self-computed; provider self-reported, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Solvable | GPT-5.6 Terra | 83.8% | human-solvable subset; Linux terminal with bioinformatics tools, Python, R and allow-listed internet; Gemini and GPT-5.6 Terra self-computed; self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Difficult | Gemini 3.7 Flash | 43.5% | human-difficult subset; same setting; developer-reported | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Difficult | Gemini 3.6 Flash | 41.2% | human-difficult subset; same setting; developer-reported | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Difficult | Claude Sonnet 5 | 34.1% | human-difficult subset; same setting; provider self-reported, as reprinted in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| BioMysteryBench — Human Difficult | GPT-5.6 Terra | 49.4% | human-difficult subset; same setting; self-computed by Google DeepMind, reported in the Gemini 3.7 Flash model card | Evaluation > Results (as of August 2026) |
| Text to Text Safety — vs. Gemini 3.6 Flash | Gemini 3.7 Flash | +1.17pp | absolute change vs. Gemini 3.6 Flash; automated content-safety policy evaluation; lower is better (a regression); automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Multilingual Safety — vs. Gemini 3.6 Flash | Gemini 3.7 Flash | -0.48pp | absolute change vs. Gemini 3.6 Flash; automated safety policy evaluation across languages; lower is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Image to Text Safety — vs. Gemini 3.6 Flash | Gemini 3.7 Flash | No change | absolute change vs. Gemini 3.6 Flash; automated content-safety policy evaluation; lower is better; automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Tone — vs. Gemini 3.6 Flash | Gemini 3.7 Flash | -0.47pp | absolute change vs. Gemini 3.6 Flash; automated evaluation of objective tone of refusals; higher is better (a regression); automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
| Unjustified-refusals — vs. Gemini 3.6 Flash | Gemini 3.7 Flash | +0.84pp | absolute change vs. Gemini 3.6 Flash; automated evaluation on borderline prompts; lower is better (a regression); automated internal evaluation, not human red teaming; not comparable with earlier cards; developer-reported | Ethics and Content Safety > Training and Development Evaluation Results |
Provenance
The page summarises the HTML card as captured on September 21, 2026, together with its methodology page. The card links a PDF version (Source: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-7-Flash-Model-Card.pdf). All figures are reported by Google DeepMind; comparison-model figures are mostly the other developers' own published numbers or public leaderboards, selected by Google. Confidence is medium: a single developer-authored source, with fast-decay benchmark claims.
Relationships
- supports: Gemini 3 / Gemini 3 Pro
- depends-on: Gemini 3.1 Pro Model Card (Google DeepMind, February 2026), Gemini 3 Pro Frontier Safety Framework Report (Google DeepMind, November 2025)
- instance-of: AI System Cards
- related: Gemini 3.8 Flash Model Card (Google DeepMind, September 2026), Google DeepMind, AI Safety Cases and Frameworks, Terminal-Bench, GDPval, OSWorld, Humanity's Last Exam, MRCR, Claude Sonnet 5, GPT-5.6 (Sol, Terra, Luna), Muse Spark (Meta Superintelligence Labs)