What Is Gemini 3.7 Flash? Benchmarks, Price, and API Guide

TL;DR Google released Gemini 3.7 Flash on August 13, 2026—just three weeks after Gemini 3.6 Flash. Positioned as the company’s most intelligent “workhorse” model yet for coding, agentic workflows, web development, and knowledge work, it delivers substantial gains over its predecessor while offering an introductory price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026 (half the original 3.6 Flash rate). It supports a 1M-token context window, multimodal inputs (text, image, video, audio, PDF), tunable thinking levels, and strong tool use. Available via the Gemini API, Google AI Studio, Antigravity, Android Studio, Gemini Enterprise, and aggregators such as CometAPI.

Key Takeaways

• Gemini 3.7 Flash is a major iterative upgrade focused on real-world software engineering, multi-step agents, and design-adherent web development.
• Key benchmark lifts vs 3.6 Flash include FrontierCode 1.1 Main (43.6% vs 34.4%), DeepSWE v1.1 (65.3% vs ~49%), WebDev Arena Elo (1588 vs 1538), GDP.pdf (34% vs 22%), and AutomationBench (30.4% vs 17%).
• Introductory pricing of $0.75 / $3.75 per 1M input/output tokens lasts until end of 2026; standard rates of $1.50 / $7.50 apply from January 1, 2027.
• 1M context, 64K–65K max output, thinking levels (low/medium/high), function calling, code execution, computer use (preview), search grounding, and structured outputs.
• Ideal for production coding agents, long-horizon engineering tasks, UI generation from design mocks, and complex document/knowledge workflows at attractive cost.
• CometAPI provides easy OpenAI-compatible access to Gemini 3.7 Flash alongside 500+ other models with unified billing and competitive rates.

What Is Gemini 3.7 Flash?

Gemini 3.7 Flash is the newest stable model in Google's Gemini 3 Flash line. Flash models are designed to balance intelligence, speed, and cost for applications that make frequent model calls. They are especially relevant to agents because an agent may need to reason, call tools, inspect results, correct errors, and continue for several turns before completing one user request.

Google calls 3.7 Flash its most intelligent workhorse model yet for coding and agents. That positioning is important. Gemini 3.7 Flash is not presented as a replacement for every frontier reasoning model. Instead, it aims to deliver enough intelligence for demanding production workflows while retaining Flash-level latency and economics.

The model card provides the clearest technical explanation: Gemini 3.7 Flash is based on Gemini 3.6 Flash and introduces algorithmic improvements to its core reasoning foundation. This helps explain why Google shipped it only three weeks after 3.6. It is an optimization and capability release built on the same base rather than a completely new pretraining generation.

Gemini 3.7 Flash Specifications

What Is New in Gemini 3.7 Flash?

Stronger Agentic Coding

The largest improvements concern software engineering. Google says the model is better at debugging, resolving issues, generating production-ready code on the first pass, and staying effective across longer coding workflows.

This matters because coding-agent cost is not determined only by the number of tokens in one response. A model that produces a plausible but broken patch may trigger more repository scans, terminal commands, test runs, repair prompts, and human review. Higher first-pass accuracy can reduce the cost per accepted result even when two models have similar listed token prices.

Gemini 3.7 Flash also reportedly adapts better when it encounters roadblocks. Google describes more disciplined planning and tool use, better clarification of intent, and fewer retries. Those are operational improvements for IDE assistants, pull-request agents, debugging systems, and internal developer tools.

Better Web Development and UI Fidelity

Google says 3.7 Flash can create more functional layouts and more feature-complete applications in fewer prompts. It is also designed to follow visual references more closely, including screenshots, images, and complete design systems.

The Code Arena result supports this focus: 3.7 Flash reaches 1588 Elo versus 1538 for 3.6 Flash. For teams building screenshot-to-code tools, design-system assistants, landing-page generators, or full-stack prototypes, the practical question is whether the new model reduces manual CSS correction and missing-feature repair. That should be measured on your own components, breakpoints, and accessibility requirements.

More Reliable Enterprise Workflows

Gemini 3.7 Flash makes large gains on knowledge-heavy and automation benchmarks. Google highlights finance, law, biosciences, complex PDFs, and real-world business workflows.

This is a useful distinction from ordinary question answering. Enterprise agents often need to read several documents, extract evidence, compare facts, update a spreadsheet or status document, draft an email, and call internal tools. AutomationBench rises from 17.0% to 30.4%, while GDP.pdf rises from 22.0% to 34.0%. These scores are still far from perfect, but the relative gains are large enough to justify retesting workflows that previously required a more expensive model.

Improved Long-Context and Computer Use

Gemini 3.7 Flash retains the one-million-token input capacity and improves Google's GDM-MRCR v2 long-context score from 91.8% to 97.0% at 128K. It also improves OSWorld-2.0 computer-use performance from 33.8% to 47.9%.

A large context window does not guarantee reliable retrieval across every token. The benchmark gain suggests better use of long inputs, but production tests should include the document order, repeated boilerplate, code structure, tables, and retrieval noise your application actually encounters.

Updated Safety Protections

Google says Gemini 3.7 Flash ships with updated safeguards for chemical, biological, radiological, and nuclear misuse and cyber offense. The model card reports broadly similar safety and tone performance to Gemini 3.6 Flash and says tracked frontier capability thresholds were not reached.

The card also lists familiar foundation-model limitations: hallucinations, occasional slowness or timeouts, incomplete knowledge, and the continuing need for jailbreak resistance improvements. Applications in legal, financial, medical, security, or other high-impact settings still need source verification and human review.

Gemini 3.7 Flash vs Gemini 3.6 Flash: Benchmark Improvements

The table below uses Google's August 2026 model card. Percentage-point changes are calculated from the published scores. Elo differences are shown as points because treating Elo as a simple percentage would be misleading.

Google and independent evaluators (including Artificial Analysis) report clear gains. Selected comparisons:

Sources: Google DeepMind model card and official blog, Artificial Analysis independent evaluations (August 2026). 3.7 Flash also competes closely with or leads mid-frontier models such as Claude Sonnet 5 and GPT-5.6 Terra on several practical coding and agentic tests while remaining significantly cheaper.

Where the Upgrade Is Largest

The largest practical gains are DeepSWE, AutomationBench, GDP.pdf, and OSWorld-2.0. Relative to the 3.6 score, those increases are approximately 34%, 79%, 55%, and 42%, respectively. Relative percentages can exaggerate changes when the baseline is low, so the absolute percentage-point improvement should remain the primary comparison.

Terminal-bench 3.0 is another striking result: 14.9% versus 5.4%. That is a 9.5-point absolute gain and roughly 176% relative improvement, but the low final score also shows that general autonomous-agent tasks remain difficult.

Where Gemini 3.7 Flash Does Not Improve

Gemini 3.7 Flash falls 0.7 percentage points behind Gemini 3.6 Flash on both CharXiv configurations. This is a useful reminder that model upgrades are not uniformly better. If your workload depends heavily on complex scientific charts, diagrams, or tool-assisted visual synthesis, keep 3.6 available during testing and compare both models on representative inputs.

Google-reported benchmarks are valuable release evidence, but they are not a substitute for an application-specific evaluation. Measure accepted outputs, latency, tool errors, retries, token use, and human-review time on at least 20 to 50 real tasks before changing production traffic.

Gemini 3.7 Flash Pricing

Google Introductory and Standard Pricing

Google lists an introductory price through December 31, 2026:

Google describes the introductory rate as half the original Gemini 3.6 Flash cost. The detailed model card currently shows the same promotional rate for both 3.6 and 3.7 Flash, followed by the same January 2027 standard rate. In other words, 3.7 provides stronger benchmark performance at the current promotional price rather than charging an immediate premium.

Thinking tokens are normally billed as output tokens. Tool use can also add model tokens or separate tool charges. Search grounding, file retrieval, URL context, computer use, and repeated agent steps can materially change the final bill.

CometAPI Pricing

At the time checked on August 14, 2026, CometAPI lists Gemini 3.7 Flash at:

For a workload using one million input tokens and 200,000 output tokens, Google's introductory token cost is $1.50: $0.75 for input plus $0.75 for output. At CometAPI's listed rates, the same token volume is $1.20. This calculation excludes tool charges, caching, retries, and any differences in request behavior.

Do not select a route on token price alone. Compare cost per accepted result, because a lower-priced route can become more expensive if it produces more failures, retries, or escalations.

How to Access Gemini 3.7 Flash

1. Google AI Studio

Google AI Studio is the quickest place to test prompts and model behavior without building a complete application. Open the Gemini 3.7 Flash AI Studio prompt, obtain an API key, and export code when the prompt is ready.

2. Gemini API

Use the stable model code gemini-3.7-flash. Google's model page recommends the Interactions API for access to current features. The model supports low, medium, and high thinking levels, but sending minimal returns an error.

Developers migrating from older Gemini generations should also review current API rules. Deprecated sampling parameters and manually prefilled model turns may not behave as they did in earlier integrations.

3. Google Antigravity and Android Studio

Google lists Gemini 3.7 Flash for agent-first development in Google Antigravity and for coding workflows in Android Studio. These surfaces are appropriate when the main task is repository navigation, application development, testing, or tool-driven code repair.

4. Gemini Enterprise and Gemini Spark

Enterprises can access the model through the Gemini Enterprise Agent Platform and Gemini Enterprise app. Google AI Pro and Ultra subscribers in more than 160 countries can use it through Gemini Spark, Google's personal agent in the Gemini app.

5. CometAPI

CometAPI exposes Gemini 3.7 Flash through both a Gemini-native request format and an OpenAI-compatible chat-completions pattern. This can be useful when a team wants one key and billing layer across Google and competing model families.

Gemini-Native CometAPI Request

model="gemini-3.7-flash"

curl "https://api.cometapi.com/v1beta/models/${model}:generateContent" \
  -H "x-goog-api-key: $COMETAPI_KEY" \
  -H "Content-Type: application/json" \
  -X POST \
  -d '{
    "contents": [{
      "parts": [{
        "text": "Review this implementation plan and identify the three highest-risk assumptions."
      }]
    }],
    "generationConfig": {
      "maxOutputTokens": 4096
    }
  }'

OpenAI-Compatible Python Request Through CometAPI

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["COMETAPI_KEY"],
    base_url="https://api.cometapi.com/v1",
)

response = client.chat.completions.create(
    model="gemini-3.7-flash",
    messages=[
        {
            "role": "user",
            "content": "Design a migration test plan from Gemini 3.6 Flash to 3.7 Flash.",
        }
    ],
)

print(response.choices[0].message.content)

Use the Gemini-native route when you need provider-specific controls that are not represented by the compatibility layer. Use the OpenAI-compatible route when portability and rapid model comparison are the priority. Confirm supported parameters in the live documentation before production use.

Best Use Cases for Gemini 3.7 Flash

Coding and Software Engineering Agents

Gemini 3.7 Flash is a strong candidate for debugging, issue resolution, repository analysis, test generation, refactoring, and code review. Its DeepSWE, FrontierCode, and terminal benchmark improvements make coding the clearest upgrade case.

Web and UI Generation

Use it for screenshot-to-code workflows, design-system implementation, interactive prototypes, and full-stack web application generation. Evaluate visual fidelity, responsive behavior, accessibility, and feature completeness rather than judging only the first screenshot.

Enterprise Document Agents

The one-million-token context window, PDF input, GDP.pdf improvement, function calling, and file search support make the model relevant to financial reports, legal documents, scientific literature, policy libraries, and internal knowledge systems.

Multimodal and Computer-Use Automation

Gemini 3.7 Flash can reason over text, images, video, audio, and PDFs, while preview computer use enables UI-driven workflows. Keep human approval around irreversible actions, permissions, transactions, and high-impact decisions.

High-Volume Multi-Step Workflows

The model is designed for agentic execution at Flash economics. It may reduce the need to send every complex task to a more expensive model. A practical routing strategy is to start with 3.7 Flash, validate the result, and escalate only ambiguous or high-risk cases.

FAQs

What is Gemini 3.7 Flash?

Gemini 3.7 Flash is Google's August 2026 stable Flash model for coding, AI agents, web development, multimodal understanding, and enterprise knowledge work. It is based on Gemini 3.6 Flash and adds algorithmic improvements to core reasoning.

What is the Gemini 3.7 Flash context window?

Gemini 3.7 Flash supports up to 1,048,576 input tokens and up to 65,536 output tokens. It accepts text, image, video, audio, and PDF input and returns text.

How much does Gemini 3.7 Flash cost?

Google's introductory price through December 31, 2026 is $0.75 per million input tokens and $3.75 per million output tokens. Google lists $1.50/M input and $7.50/M output beginning January 1, 2027. CometAPI lists $0.60/M input and $3.00/M output at the time of publication.

Is Gemini 3.7 Flash better than Gemini 3.6 Flash?

It is better on most Google-published benchmarks, especially production coding, long-horizon software engineering, workflow automation, PDF comprehension, long-context retrieval, and computer use. It is slightly worse on CharXiv chart reasoning, so the answer depends on the workload.

Is Gemini 3.7 Flash good for coding agents?

Yes. Coding and agents are its main target use cases. Google reports 65.3% versus 48.6% for 3.6 Flash on DeepSWE v1.1 and 43.6% versus 34.4% on FrontierCode 1.1 Main.

Does Gemini 3.7 Flash support image generation?

No. It can understand image inputs but produces text output. Image generation and audio generation are not supported by this model.

Which thinking levels does Gemini 3.7 Flash support?

It supports low, medium, and high thinking. Google's developer page says minimal thinking is not supported and returns an error.

How do I use Gemini 3.7 Flash through CometAPI?

Create a CometAPI key, choose gemini-3.7-flash, and call either the Gemini-compatible /v1beta/models/{model}:generateContent endpoint or the OpenAI-compatible /v1/chat/completions endpoint. Check the live model page for current pricing and parameter support.

Conclusion

Gemini 3.7 Flash represents a rapid, focused improvement in Google’s workhorse line: stronger coding and agentic capabilities, better design adherence for web and UI work, improved complex-document and knowledge-work performance, and attractive introductory pricing that makes high-volume production agents more economical. With a 1M-token context, robust tool suite, and tunable reasoning, it is well-suited for real-world software engineering and multi-step automation.

Whether you access it directly through Google’s developer tools or via a unified platform such as CometAPI, Gemini 3.7 Flash is ready for production evaluation today. Explore the live model catalog and competitive rates at cometapi.com, claim free trial credits, and start building more reliable, cost-effective agents with the latest Flash model from Google.