Built for budget-conscious teams running coding agents and document-heavy workflows, Gemini 3.7 Flash pairs a 1,048,576-token context window with a 16.7-point jump in long-horizon coding accuracy over Gemini 3.6 Flash. It is the practical upgrade path for existing 3.6 Flash users, though those running terminal-heavy computer-use agents may still prefer a rival model.
Gemini 3.7 Flash is Google DeepMind's Gemini 3 model, released in August 2026 as a fast, low-cost workhorse for agentic coding, document analysis, and web-app scaffolding. It scores 65.3% on the DeepSWE long-horizon coding benchmark, ahead of comparable Flash-tier and mid-tier rivals in Google's own published evaluation suite.
Provider: Google DeepMind · Family: Gemini 3
More about Google DeepMind on HokAI
Context window: 1,048,576 tokens · Max output: 65,536
Input modalities: text, image, audio, video, tool-calls · Output: text, tool-calls
About Gemini 3.7 Flash
Google DeepMind released Gemini 3.7 Flash on August 13, 2026, three weeks after Gemini 3.6 Flash, as the mid-tier "Flash" entry in the Gemini 3 lineup, positioned below the Pro and Ultra tiers as a fast, low-cost workhorse for coding and agentic work. Google describes it as a refinement of Gemini 3.6 Flash's reasoning foundation through algorithmic improvements rather than a new pretraining run, and has not disclosed a parameter count or confirmed whether the underlying architecture uses a mixture-of-experts design. On Google's own published evals, Gemini 3.7 Flash scores 65.3% on DeepSWE v1.1 (long-horizon coding), up 16.7 points from Gemini 3.6 Flash's 49.0%, and 43.6% on FrontierCode 1.1 (production code quality), up from 34.4%. It tops Google's WebDev Arena at 1588 Elo, ahead of Claude Sonnet 5 (1541) and GPT-5.6 Terra (1523). On the third-party Artificial Analysis Intelligence Index it scores 56, trailing GPT-5.6 Terra and Muse Spark 1.2 (57 each) but ahead of Claude Sonnet 5 (55) and its own predecessor (52). On AutomationBench, Google's enterprise-workflow eval, it scores 30.4%, comfortably ahead of Claude Sonnet 5 (10.7%) and GPT-5.6 Terra (23.6%), and on the GDP.pdf document-comprehension test it reaches 34.0%, up from 22.0% for Gemini 3.6 Flash. The model accepts up to 1,048,576 input tokens and returns as many as 65,536 output tokens. On GDM-MRCR v2, Google's long-context recall benchmark, it holds 97.0% accuracy at 128K token depth, suggesting the 1M window is usable rather than nominal. Input modalities cover text, images, audio, and video; output is text only. It supports native function calling and tool use, with Google citing improved multi-step planning, better recovery when a task hits a roadblock, and higher instruction-following fidelity than Gemini 3.6 Flash. Thinking depth is tunable across low, medium, and high levels, trading latency and cost for reasoning quality, with medium as the default. Google prices the model aggressively for its introductory window, undercutting typical mid-tier competitor rates by more than half; the discount is scheduled to expire at the end of 2026, after which list price roughly doubles. Cached input carries a steep discount versus fresh input tokens. The model is available through the Gemini API and Google AI Studio, Android Studio, Google's Gemini Enterprise Agent Platform (the 2026 rebrand of Vertex AI), and the consumer Gemini Spark app for AI Pro and Ultra subscribers in more than 160 countries. There is no AWS Bedrock or Azure listing, since Google keeps Gemini distribution in-house. As a closed-weights model there are no downloadable weight files, VRAM requirements, or quantization options. Google's model card describes automated content-safety evaluation across multiple languages, manual red-teaming by teams external to the development group, and a Frontier Safety Framework assessment completed in April 2026 with deployed mitigations in the CBRN and cybersecurity domains. Named limitations include a continued risk of hallucination, potential jailbreak vulnerability despite strengthened mitigations, occasional slowness or timeout under load, and domain-specific knowledge gaps. Google does not name its red-teaming partners for this release. Gemini 3.7 Flash fits budget-conscious teams running high-volume coding agents, document automation in finance, law, and life sciences, and web-app scaffolding, where its cost and Google-reported benchmark wins over Claude Sonnet 5 and GPT-5.6 Terra are most pronounced. Teams building terminal-heavy or computer-use agent loops should weigh GPT-5.6 Terra instead, which leads on DeepSWE's hardest long-horizon tasks at 69.6%. Anyone locked into a self-hosted or air-gapped deployment mandate will need to pick a different, open-weights model, since Gemini 3.7 Flash ships as a closed-weights API product. Google states a primary training-data cutoff of March 2026, with some domains limited to January 2025. Google has pledged that paid usage through the Gemini API and its enterprise platform is excluded from model training, with only brief abuse and legal-compliance logs kept; the free AI Studio tier, by contrast, does feed submitted content back into product improvement. Eligible enterprise customers can additionally have prompts, responses, and identifying metadata stripped before logging. The platform inherits Google Cloud's SOC 2, ISO 27001, and HIPAA-eligible certifications, with HIPAA coverage requiring a signed Business Associate Agreement. Versus Gemini 3.6 Flash, the one regression Google disclosed is a 0.7-point dip on CharXiv Reasoning without tool access, to 84.5%, even as most other evals improved. Migrated prompts must drop deprecated sampling parameters (temperature, top_p, top_k) and prefilled model turns, support for which Gemini 3.6 Flash had already removed.
Pricing
Google prices Gemini 3.7 Flash at $0.75 per 1M input tokens and $3.75 per 1M output tokens through the end of 2026, an introductory rate that is half of the predecessor's launch price. Cached input tokens get a 90% discount, roughly $0.075 per 1M. Standard pricing of $1.50 input and $7.50 output per 1M tokens takes effect on January 1, 2027.
Key Features
- Tunable Thinking Levels: Low, medium, and high thinking configurations trade reasoning depth against latency and cost per request, with medium as the default.
- 1M-Token Context Window: Accepts up to 1M tokens of input and returns as many as 65,536 tokens of output, holding 97.0% recall accuracy at 128K depth on Google's GDM-MRCR v2 test.
- Multimodal Input: Reads text, images, audio, and video in a single request, positioned by Google as its workhorse model for document-heavy business workflows.
- Native Tool Use and Function Calling: Plans and calls tools across multi-step agent loops, with Google citing better recovery from roadblocks and higher instruction fidelity than the prior release.
- Discounted Cached Input: System prompts and long documents that get reused across calls are billed at a fraction of the standard input rate once cached, per Google's pricing page.
Pros
- Beats Claude Sonnet 5 (1541 Elo) and a leading GPT rival (1523 Elo) on Google's WebDev Arena leaderboard for web-app generation.
- Scores 43.6% on FrontierCode for production code quality, a meaningful jump from its predecessor's 34.4%.
- Nearly triples enterprise workflow automation accuracy versus its predecessor, per Google's AutomationBench results.
- 1M-token context window holds accuracy deep into long documents, per Google's own long-context evaluation.
Cons
- Trails GPT-5.6 Terra on long-horizon agent and terminal tasks (69.6% on DeepSWE's hardest benchmark).
- Closed weights: available only through Google's API and enterprise platforms, with no self-hosting or fine-tuning of the base model.
- The launch discount is temporary: list pricing roughly doubles once the introductory window closes in early 2027.
- Slight regression on tool-free visual reasoning, dipping 0.7 points to 84.5% on CharXiv Reasoning.
Benchmarks
- gdp pdf: 34
- deepswe v1 1: 65.3
- gdpval aa elo: 1525
- automation bench: 30.4
- gdm mrcr v2 128k: 97
- webdev arena elo: 1588
- frontier code 1 1: 43.6
- charxiv reasoning no tools: 84.5
- artificial analysis intelligence index: 56
- artificial analysis price blended per m: 1.35
- artificial analysis speed tokens per sec: 340.1
Frequently Asked Questions
How much does Gemini 3.7 Flash cost per 1M tokens?
Through the end of 2026, Google charges $0.75 for every 1M input tokens and $3.75 for every 1M output tokens processed by Gemini 3.7 Flash, an introductory rate exactly half of what the predecessor launched at. A 90% discount applies to cached input, bringing it to about $0.075 per 1M, and the rate doubles on both ends of the meter starting January 2027.
How does Gemini 3.7 Flash compare on benchmarks vs Claude Sonnet 5?
On Google's AutomationBench for enterprise workflows, Gemini 3.7 Flash scores 30.4% versus Claude Sonnet 5's 10.7% and a leading GPT rival's 23.6%, a lead Google attributes to better multi-step planning and tool use. Claude Sonnet 5 still wins on Agent's Last Exam multimodal desktop tasks, posting 33.3% against Gemini 3.7 Flash's 26.3% pass rate, so the better choice depends on whether the workload is document-heavy or computer-use-heavy.
Is Gemini 3.7 Flash open source or proprietary?
Gemini 3.7 Flash is fully proprietary: Google has not released model weights, so access is limited to the Gemini API, Google AI Studio, the Gemini Enterprise Agent Platform, and the Gemini Spark consumer app. There is no self-hosting, no fine-tuning of the base weights, and no air-gapped deployment option.
Does Gemini 3.7 Flash train on user data?
On the paid Gemini API and the Gemini Enterprise Agent Platform, Google commits not to train on customer prompts and responses, retaining them only briefly for abuse detection and legal compliance. The free Google AI Studio tier does use submitted content to improve Google products, and enterprise customers can additionally request zero data retention on the Enterprise Agent Platform.
Who is Gemini 3.7 Flash best for, and who should avoid it?
Gemini 3.7 Flash fits budget-conscious teams running coding agents, enterprise document automation, and web-app scaffolding, where its low cost and Google-reported benchmark wins matter most. Teams building terminal-heavy or computer-use-heavy agent loops should look at a stronger long-horizon agent model instead, and teams that require self-hosted or air-gapped infrastructure will need an open-weights model.