Gemini 3.7 Flash: What's New, Pricing, Coding and AI Agent Capabilities

Google launched Gemini 3.7 Flash on August 13, 2026, positioning it as the most intelligent workhorse model yet for coding and agents. The release comes just three weeks after Gemini 3.6 Flash, and is a direct result of developer feedback and algorithmic innovations.

Google describes the model as targeting developers building production-grade applications, agentic pipelines, and complex multi-step coding workflows. On launch day, Google introduced introductory pricing at $0.75 per million input tokens, roughly half the rate that will apply after the introductory period ends. 3.7 Flash delivers substantial improvements across software engineering, knowledge work, and web development workflows.

This article covers everything verified from official sources: what changed from 3.6 Flash, the full current pricing structure, technical specifications, how thinking levels work, coding benchmark evidence, agentic capabilities, and who the model is likely a good fit for.

What Is Gemini 3.7 Flash?

Gemini 3.7 Flash (gemini-3.7-flash) is generally available and ready for production use. It is Google's most capable Flash model, built for complex coding, agentic workflows, and reliable multi-step execution.

Gemini 3.7 Flash is the next iteration in the Gemini 3 model family, featuring algorithmic improvements to its core reasoning foundation. It supports customizable thinking configurations to control the mix of quality, cost, and latency. The model is based on Gemini 3.6 Flash.

The model is natively multimodal, accepting text, image, video, audio, and PDF as inputs, and producing text as output. Its primary intended workloads include code generation and debugging, multi-step agentic workflows, web development, and knowledge-intensive document tasks.

In Google's Gemini lineup, Flash models occupy the performance-and-efficiency tier, designed for production workloads where both intelligence and scalability matter. Gemini 3.7 Flash sits at the top of that tier as the most capable current Flash model.

What's New in Gemini 3.7 Flash?

Google reports substantial improvements across coding, web development, knowledge work, and agentic task execution compared to Gemini 3.6 Flash.

Better Coding and Software Engineering

3.7 Flash shows strong gains over 3.6 Flash in coding tasks like debugging and issue resolution. It also achieves higher first-pass code accuracy and has improved performance in generating production-ready code as seen in FrontierCode 1.1 Main (43.6% vs 34.4%) and DeepSWE v1.1 (65.3% vs 49.0%).

The Google DeepMind model card, which contains the more detailed technical evaluation, records the Gemini 3.6 Flash score on DeepSWE v1.1 as 48.6%. The blog post rounds this slightly to 49.0%. Both are Google sources; the model card is the more detailed technical document.

A more disciplined execution means less manual oversight and fewer retries across engineering workflows.

Improved Web Development

In web development, 3.7 Flash generates more functional layouts and feature-complete apps in fewer prompts. For UI generation, the model shows high design adherence and parity based on a reference input, whether it's a screenshot, an image, or a full design system. It outperforms 3.6 Flash on Arena.ai's WebDev Arena with an Elo score of 1588 vs 1538.

Better Knowledge Work

For knowledge-dense fields like finance, law, and biosciences, 3.7 Flash delivers improved reasoning and accuracy. It significantly outperforms 3.6 Flash on the GDP.pdf benchmark (34.0% vs 22.0%), an eval for testing a model's ability to process complex documents. According to Google's model card evaluation, the model also scores 90.7% on Harvey LAB-AA, a benchmark for complex legal workflows, compared to 85.1% for 3.6 Flash.

It is important to note that benchmark performance in a domain does not guarantee expert-level accuracy on every real-world professional task. Outputs in high-stakes fields such as law and finance should always be reviewed by qualified professionals.

Improved Agentic Workflows

3.7 Flash also surpasses 3.6 Flash in AutomationBench, demonstrating it can more effectively complete real-world business workflows (30.4% vs 17.0%). AutomationBench uses a private evaluation set; the percentage represents task completion on that benchmark, not a general claim about what fraction of enterprise workflows the model can automate.

Gemini 3.7 Flash delivers a noticeably improved developer experience over 3.6 Flash. It better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. It thinks more diligently, putting in more effort into multi-step planning and tool calls.

Gemini 3.7 Flash Pricing

The figures below are drawn from the official Google Gemini API pricing page as of August 15, 2026. All prices are per million tokens unless otherwise noted.

Standard (Pay-as-you-go)

Batch API (50% off Standard)

Priority Inference

On the Free tier, input and output tokens are free of charge and context caching is also available at no cost. Content on the Free tier is used to improve Google's products.

A few points developers should keep in mind. First, output price includes thinking tokens there is no separate line item for thinking. Second, introductory pricing expires on December 31, 2026; after that, $1.50/1M input tokens and $7.50/1M output tokens will apply. Third, Google also reduced Gemini 3.6 Flash to the same introductory rate through December 31, 2026, meaning both models currently cost the same on the Standard tier.

This is API pricing. It is entirely separate from consumer-facing subscriptions such as Google AI Pro or Ultra, which govern end-user access to the Gemini app and Gemini Spark.

Gemini 3.7 Flash Technical Specifications

The following specifications are drawn from the Gemini API model documentation, the latest model guide, and the Google DeepMind model card.

Gemini 3.7 Flash Thinking Levels

Gemini 3.7 Flash gives developers flexible control over latency and intelligence by adjusting the model's thinking level. Three levels are available: Low, Medium, and High. The default is Medium.

Low thinking effort reduces time-to-answer for latency-critical tasks like incident response pipelines, real-time chat, writing drafts, and fast data analysis.

Medium (default) offers the best quality for most tasks. It is recommended for complex code and agentic use cases, with higher first-pass accuracy.

High thinking effort maximizes the model's ability to think and use tools. It is best for complex reasoning, hard math, and the most difficult coding and agent tasks. It allows extended thoughts and function calls, with higher token consumption and cost.

The practical implication: because thinking tokens are billed at the standard output token rate, High is meaningfully more expensive and slower than Low or Medium. For latency-sensitive production pipelines, Low may be the right choice even if the task seems complex, because a faster partial answer is sometimes more useful than a slower comprehensive one. Developers should test all three levels against their specific workloads rather than assuming High is always best.

Is Gemini 3.7 Flash Good for Coding?

Based on Google's published evaluations, the answer is yes with important caveats about what benchmarks can and cannot tell you. Google's evaluation results reflect controlled benchmark conditions; actual performance on a specific codebase will depend on its complexity, language, test coverage, the tools provided, the prompting approach, and the thinking level configured.

Code Generation

3.7 Flash achieves higher first-pass code accuracy and has improved performance in generating production-ready code as seen in FrontierCode 1.1 Main (43.6% vs 34.4%). FrontierCode 1.1 Main measures production code quality.

Debugging

Google's announcement describes improved debugging and issue resolution, but provides this at a qualitative level. Developers should expect complex debugging tasks on large unfamiliar codebases to still require iteration.

Software Engineering Tasks

According to Google's published model card evaluation, 3.7 Flash scores 65.3% on DeepSWE v1.1 (long-horizon software engineering) versus 48.6% for 3.6 Flash. On Terminal-bench 2.1, measuring agentic terminal coding, Google's evaluation shows 85.8% versus 78.0%.

Web Development

The WebDev Arena result (Elo 1588 vs 1538) reflects improved ability to generate functional web layouts and applications from design references. Google specifically notes stronger design adherence when working from screenshots, images, or design systems.

Long, Multi-Step Coding Tasks

Terminal-bench 3.0, which measures general agent capabilities in coding contexts, shows 14.9% for 3.7 Flash versus 5.4% for 3.6 Flash. This is a substantial relative improvement, though the absolute scores indicate these are still genuinely hard tasks.

To summarize: Google's benchmarks show meaningful improvements across coding tasks. Benchmark scores should be read as evidence of directional improvement, not as guarantees of production performance.

Gemini 3.7 Flash for AI Agents

A critical distinction applies here. Gemini 3.7 Flash is a model. An AI agent is a full system: a model combined with tools, instructions, permissions, orchestration logic, and application code. The model contributes to agent performance but does not determine it alone.

Multi-Step Planning

3.7 Flash thinks more diligently, putting in more effort into multi-step planning and tool calls. This is reflected in the Terminal-bench 3.0 result (14.9% vs 5.4%), which tests general agent capabilities across extended task sequences.

Tool Calling

Function calling and tool use are fully supported capabilities. Due to its improved performance and reasoning, Gemini 3.7 Flash is now the new default model powering the Antigravity agent in Gemini Managed Agents and Google Antigravity SDK. This is one of the strongest signals of Google's confidence in its agentic performance.

Handling Roadblocks

3.7 Flash better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. In practical terms, a model that handles unexpected situations mid-workflow reduces the need for developer-built error-recovery logic and produces more reliable agent behavior across runs.

Instruction Following

Improved instruction fidelity means the model is more likely to do what the developer intended on the first attempt. This reduces the need for elaborate prompt engineering workarounds and produces more consistent results across agent loop iterations.

Business Workflow Automation

3.7 Flash surpasses 3.6 Flash in AutomationBench, demonstrating it can more effectively complete real-world business workflows (30.4% vs 17.0%). AutomationBench uses a private evaluation set from Zapier. The percentage represents task completion on that specific benchmark, not a general claim about all enterprise tasks.

Gemini 3.7 Flash and Gemini Spark

Two things that are easy to conflate deserve a clear distinction.

Gemini 3.7 Flash is the underlying AI model, available through the Gemini API, Google AI Studio, and other developer platforms.

Gemini Spark is a consumer product, an agent experience built on top of the model.

Gemini Spark, available to Google AI Pro and Ultra subscribers in over 160 countries, will be using Gemini 3.7 Flash starting today. Spark is Google's personal AI agent that runs 24/7, taking action on your behalf while under your direction. This model update makes Spark more efficient for knowledge work with improved tool use for Google Workspace apps, delivering improved accuracy and output quality for complex, multi-skill workflows.

Spark is a consumer product experience. Its availability, feature set, and pricing are governed by Google AI subscription terms and are separate from the developer API.

Where Can You Use Gemini 3.7 Flash?

The following are officially confirmed access points, drawn from Google's announcement and the Google DeepMind model card.

Gemini 3.7 Flash is distributed through the Gemini App via Spark, the Gemini Enterprise App, Gemini Enterprise Agent Platform, Google AI Studio, the Gemini API, and Google Antigravity. Google's announcement additionally confirms availability through Android Studio for Android developers.

The model is accessible to developers immediately via the API using the model ID gemini-3.7-flash, and can be explored in Google AI Studio without requiring a paid API key on the free tier.

Gemini 3.7 Flash vs Gemini 3.6 Flash

All benchmark figures below are from Google DeepMind's published evaluation (August 2026). Google's methodology link is deepmind.com/models/evals-methodology/gemini-3-7-flash.

These results are from Google's published evaluation. They are not independently verified by this publication.

Who Should Use Gemini 3.7 Flash?

Based on Google's documentation and benchmark evidence, the following groups may find Gemini 3.7 Flash a good fit.

Developers building production coding tools: The FrontierCode and DeepSWE gains indicate stronger code quality for software engineering tasks. The model is worth evaluating for any workflow currently using 3.6 Flash for code generation or debugging, given that it currently costs the same.

AI agent and automation builders: The AutomationBench and Terminal-bench improvements, combined with documented gains in multi-step planning and roadblock handling, make it a strong candidate for agentic pipelines. Google has also made it the default model in the Antigravity agent framework.

Web developers: The WebDev Arena improvement and Google's specific documentation of stronger design adherence suggest this model may be useful for generating web UIs from design references, screenshots, or design systems.

Enterprises handling knowledge-intensive document workflows: The GDP.pdf and Harvey LAB-AA results support knowledge work applications in legal, financial, and life sciences contexts, though professional review remains essential for high-stakes outputs.

Teams moving from 3.6 Flash: Through December 31, 2026, 3.7 Flash delivers measurably better benchmark performance at the same price as the discounted 3.6 Flash, making it the practical default for most existing Flash workloads.

Gemini 3.7 Flash: Key Strengths and Considerations

Strengths

Coding and agentic benchmarks show consistent, meaningful gains over 3.6 Flash across production code quality, long-horizon software engineering, web development, and enterprise workflow automation. Introductory pricing through December 31, 2026 is the same as the current discounted 3.6 Flash rate, so switching has no current cost penalty. The model supports a wide range of developer tools: context caching, function calling, structured outputs, search grounding, Google Maps grounding, code execution, file search, and computer use (the latter currently in Preview).

Considerations

Introductory pricing ends December 31, 2026. Standard rates of $1.50/$7.50 per million input/output tokens take effect on January 1, 2027. Developers building cost models for workloads that extend into 2027 should use the post-introductory rates for planning purposes.

Thinking level affects both cost and latency. Thinking tokens are billed at the standard output rate, so High thinking can meaningfully increase per-request costs. High is not automatically best for every task.

Benchmarks have real limits. Google's evaluation results reflect specific controlled benchmarks. Real-world performance on specific codebases, document types, and agent workflows will vary.

The knowledge cutoff date for Gemini 3.7 Flash is March 2026. Users can expect updated information for some domains while in others they may experience the model's knowledge limited to January 2025, in line with the Gemini 3 Model Family. For time-sensitive applications, search grounding or retrieval-augmented approaches may be necessary.

Gemini 3.7 Flash may exhibit some of the general limitations of foundation models, such as hallucinations. Outputs in critical applications should be verified.

Effective agent performance depends on the full system design, not the model alone. Tool definitions, orchestration logic, environment configuration, and prompting all affect agent outcomes.

Conclusion

Gemini 3.7 Flash is Google's newest Flash-tier model, launched August 13, 2026, with a clear focus on coding and AI agents. Google's published benchmarks show consistent gains over Gemini 3.6 Flash across production code quality, long-horizon software engineering, web development, enterprise workflow automation, and complex document comprehension.

For developers, the introductory pricing at $0.75/$3.75 per million input/output tokens through December 31, 2026, makes 3.7 Flash cost-equivalent to the current discounted 3.6 Flash, effectively making it the sensible default for existing Flash workloads with no upfront cost change. Standard pricing of $1.50/$7.50 applies from January 1, 2027, which developers should factor into production cost planning.

The model may be a good fit for teams building coding tools, agentic pipelines, UI generation systems, and knowledge-intensive document workflows. As with any model, benchmark improvements reflect controlled evaluation conditions and should not be assumed to transfer one-for-one to every real-world application.