Google has launched Gemini 3.7 Flash, a new artificial intelligence model designed specifically for coding, software engineering and AI-agent workflows. The company describes it as its most intelligent Flash model yet for coding and agents, with substantial improvements over Gemini 3.6 Flash across software development, enterprise automation, document processing and complex multi-step tasks.
The model is attracting attention because of its combination of performance and cost. Google’s published benchmark results show Gemini 3.7 Flash leading several tests, including FrontierCode 1.1 Main, Code Arena and Google’s private AutomationBench. However, it does not lead every agent benchmark: GPT-5.6 Terra remains ahead on DeepSWE v1.1, Terminal-Bench 2.1 and OSWorld-2.0.
Gemini 3.7 Flash Targets AI Agents
Gemini 3.7 Flash was introduced on August 13, only three weeks after Gemini 3.6 Flash. Google said the rapid release was driven by developer feedback and algorithmic improvements rather than simply a larger model.
The focus is increasingly on AI agents that can perform multi-step tasks instead of simply responding to individual prompts.
These systems can plan actions, use tools, write and test code, navigate software environments and complete business workflows with less human intervention.
Gemini 3.7 Flash At A Glance
| Metric | Gemini 3.7 Flash |
|---|---|
| Launch date | August 13, 2026 |
| Primary focus | Coding and AI agents |
| Context window | 1,048,576 tokens |
| Input price | $0.75 per 1M tokens* |
| Output price | $3.75 per 1M tokens* |
| FrontierCode 1.1 | 43.6% |
| DeepSWE v1.1 | 65.3% |
| Code Arena | 1,588 Elo |
| Terminal-Bench 2.1 | 85.8% |
| AutomationBench | 30.4% |
| OSWorld-2.0 | 47.9% |
*Introductory pricing through December 31, 2026. Google says prices will rise to $1.50 per 1M input tokens and $7.50 per 1M output tokens from January 1, 2027.
Gemini 3.7 Flash Beats Its Predecessor Across Key Benchmarks
The biggest improvement is compared with Gemini 3.6 Flash.
On FrontierCode 1.1 Main, which measures production-code quality, Gemini 3.7 Flash scored 43.6%, compared with 34.4% for the previous model.
On DeepSWE v1.1, which evaluates long-horizon software engineering, the score increased from 48.6% to 65.3%.
Gemini 3.7 Flash Vs Gemini 3.6 Flash
| Benchmark | Gemini 3.7 Flash | Gemini 3.6 Flash | Improvement |
|---|---|---|---|
| FrontierCode 1.1 | 43.6% | 34.4% | +9.2 pts |
| DeepSWE v1.1 | 65.3% | 48.6% | +16.7 pts |
| Code Arena | 1,588 | 1,538 | +50 Elo |
| Terminal-Bench 2.1 | 85.8% | 78.0% | +7.8 pts |
| AutomationBench | 30.4% | 17.0% | +13.4 pts |
| OSWorld-2.0 | 47.9% | 33.8% | +14.1 pts |
| GDP.pdf | 34.0% | 22.0% | +12 pts |
The improvements are particularly notable in tasks requiring multiple steps, coding and interaction with tools.
Gemini 3.7 Flash Leads Several AI Agent Tests
Google’s benchmark table shows Gemini 3.7 Flash performing strongly against competing models.
It achieved a 1,588 Elo rating on Code Arena’s web-development evaluation, ahead of GPT-5.6 Terra at 1,523 and Claude Sonnet 5 at 1,541.
Gemini 3.7 Flash also scored 30.4% on AutomationBench, compared with 23.6% for GPT-5.6 Terra and 10.7% for Claude Sonnet 5.
Selected Agent And Coding Benchmarks
| Benchmark | Gemini 3.7 Flash | Claude Sonnet 5 | GPT-5.6 Terra | Leader |
|---|---|---|---|---|
| FrontierCode 1.1 | 43.6% | 42.7% | 41.3% | Gemini |
| DeepSWE v1.1 | 65.3% | 53.8% | 69.6% | GPT |
| Code Arena | 1,588 | 1,541 | 1,523 | Gemini |
| Terminal-Bench 2.1 | 85.8% | 80.4% | 87.4% | GPT |
| AutomationBench | 30.4% | 10.7% | 23.6% | Gemini |
| OSWorld-2.0 | 47.9% | — | 50.2% | GPT |
The results show that Gemini 3.7 Flash is highly competitive, but the claim that it simply “tops” all AI-agent benchmarks would be too broad. It leads some evaluations while trailing other models on several important tests.
Terminal-Bench Shows The Model Still Has Competition
Terminal-Bench 2.1 evaluates an AI model’s ability to perform coding tasks in a terminal environment.
Gemini 3.7 Flash scored 85.8%, a substantial improvement over Gemini 3.6 Flash’s 78%.
However, GPT-5.6 Terra scored 87.4%, meaning Google’s model did not take the top position on this particular benchmark.
Terminal-Bench 2.1
GPT-5.6 Terra 87.4%
████████████████████████████████████████
Gemini 3.7 Flash 85.8%
███████████████████████████████████████
Claude Sonnet 5 80.4%
████████████████████████████████████This highlights the importance of looking at multiple benchmarks rather than relying on a single score.
AI Agents Are Becoming The Main Battleground
The launch reflects a broader shift in the AI industry.
Earlier generations of AI models were primarily judged by their ability to answer questions, summarize information or generate text.
The competition is now increasingly focused on whether models can actually perform work.
An AI agent can potentially take a goal such as “build this application,” break it into smaller tasks, write code, test the result, identify errors and make corrections.
From Chatbots To AI Agents
Traditional AI
↓
User Prompt
↓
AI Response
↓
User Takes Action
AI Agent
↓
User Goal
↓
Planning
↓
Tool Use
↓
Execution
↓
Testing
↓
Correction
↓
Completed TaskGemini 3.7 Flash has been designed around this second model of interaction.
Coding Is A Major Strength
Software engineering is one of the areas where Google reports some of the largest gains.
FrontierCode 1.1 Main measures production-code quality, while DeepSWE focuses on longer software-engineering tasks.
Gemini 3.7 Flash’s 65.3% score on DeepSWE represents a significant improvement over its predecessor’s 48.6%. However, GPT-5.6 Terra remains ahead at 69.6%.
The model’s improved ability to handle roadblocks, clarify intent and follow instructions is also intended to reduce the amount of manual intervention required from developers.
Web Development Performance Also Improves
Gemini 3.7 Flash recorded a 1,588 Elo score on Code Arena’s web-development evaluation.
That compares with 1,538 for Gemini 3.6 Flash.
The improvement indicates that the model is becoming better at generating functional web interfaces and handling more complex development instructions.
Google said the model can also orchestrate sub-agents for more complicated web-development workflows.
Web Development Improvement
| Model | Code Arena Elo |
|---|---|
| Gemini 3.7 Flash | 1,588 |
| Claude Sonnet 5 | 1,541 |
| Gemini 3.6 Flash | 1,538 |
| Muse Spark 1.2 | 1,535 |
| GPT-5.6 Terra | 1,523 |
Gemini’s 50-point improvement over its predecessor is one of the more visible gains in Google’s benchmark results.
Enterprise Automation Is Another Focus
Google is also positioning Gemini 3.7 Flash as a model for business automation.
Its private AutomationBench score rose to 30.4%, compared with 17% for Gemini 3.6 Flash.
The test is designed to measure the ability to complete real-world business workflows. Google says the improvement reflects the model’s stronger planning and tool-use capabilities.
AutomationBench Results
Gemini 3.7 Flash 30.4%
██████████████████████████████
GPT-5.6 Terra 23.6%
███████████████████████
Gemini 3.6 Flash 17.0%
█████████████████
Claude Sonnet 5 10.7%
███████████Because AutomationBench is a private evaluation, the results should be viewed as Google’s reported benchmark rather than an independently reproducible public leaderboard.
Long-Context Performance Is Strong
Gemini 3.7 Flash also has a context window of approximately 1 million tokens.
This allows the model to process very large amounts of information in a single context, potentially making it useful for lengthy codebases, large documents and complex research workflows.
On the GDM-MRCR v2 eight-needle evaluation with a 128K context, Gemini 3.7 Flash achieved a 97% average, compared with 91.8% for Gemini 3.6 Flash and 93.5% for GPT-5.6 Terra.
Long-Context Results
| Model | GDM-MRCR v2 |
|---|---|
| Gemini 3.7 Flash | 97.0% |
| GPT-5.6 Terra | 93.5% |
| Gemini 3.6 Flash | 91.8% |
| Claude Sonnet 5 | 81.5% |
Long context can be particularly useful for agentic coding because an agent may need to keep track of large codebases, documentation and previous actions.
The Model Is Also Designed For Knowledge Work
Google has expanded Gemini 3.7 Flash beyond programming.
The model scored 1,525 Elo on GDPVal-AA v2, which evaluates knowledge work.
It also recorded a 90.7% score on Harvey LAB-AA, a benchmark for complex legal workflows, slightly above GPT-5.6 Terra’s 85.2% and close to Claude Sonnet 5’s 90.1%.
This suggests Google is positioning Flash as a general-purpose work model rather than a specialized coding system.
Document Understanding Gets A Major Upgrade
Gemini 3.7 Flash also improved substantially on Google’s GDP.pdf benchmark.
The model scored 34%, compared with 22% for Gemini 3.6 Flash.
The test measures the ability to understand complex documents and extract useful information from them.
This capability could be useful for AI agents working with financial reports, legal documents, research papers and enterprise records.
Document And Research Performance
| Benchmark | Gemini 3.7 Flash |
|---|---|
| GDP.pdf | 34.0% |
| Harvey LAB-AA | 90.7% |
| HLE-Verified | 53.6% |
| BioMysteryBench | 87.1% |
| LABBench2 | 82.1% |
| LVBench | 85.4% |
The range of evaluations demonstrates Google’s effort to make the model capable of handling different types of professional workflows.
Gemini 3.7 Flash Is Cheaper During Its Introductory Period
Price is another major part of Google’s strategy.
Through the end of 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens.
Google says those prices will double from January 1, 2027.
Gemini 3.7 Flash Pricing
| Period | Input / 1M Tokens | Output / 1M Tokens |
|---|---|---|
| Through Dec. 31, 2026 | $0.75 | $3.75 |
| From Jan. 1, 2027 | $1.50 | $7.50 |
The temporary discount could make the model particularly attractive to developers building high-volume AI agents.
For agentic applications, where one user request can generate many model calls, inference costs can become a major part of the economics.
Cost Matters More For AI Agents
Traditional chatbot applications may require only one or two model responses for each user interaction.
Agents can require many more calls because they may plan, call tools, inspect results, make corrections and continue working.
That means even a modest difference in token prices can become significant at scale.
User Goal
↓
Agent Planning
↓
Model Call
↓
Tool Call
↓
Model Evaluation
↓
Another Model Call
↓
Code Execution
↓
Correction
↓
Final ResponseGoogle’s lower introductory pricing is therefore designed to make it economically viable for developers to run more complex agent workflows.
Google Is Building An Agent Ecosystem Around Gemini
Gemini 3.7 Flash is available through Google Antigravity, the Gemini API via Google AI Studio and Android Studio.
Enterprise users can access the model through Gemini Enterprise Agent Platform and the Gemini Enterprise app. Individuals can also access it through Spark in the Gemini app for Google AI Pro and Ultra subscribers in supported countries.
This gives Google multiple channels through which developers and businesses can adopt the model.
Gemini 3.7 Flash Availability
| Platform | Availability |
|---|---|
| Google Antigravity | Yes |
| Gemini API | Yes |
| Google AI Studio | Yes |
| Android Studio | Yes |
| Gemini Enterprise Agent Platform | Yes |
| Gemini Enterprise app | Yes |
| Gemini Spark | Available to eligible subscribers |
The strategy allows Google to target both individual developers and large enterprises.
Google Is Not Leading Every Benchmark
Despite the strong results, Gemini 3.7 Flash does not dominate every evaluation.
On DeepSWE v1.1, GPT-5.6 Terra scored 69.6%, ahead of Gemini’s 65.3%.
GPT-5.6 Terra also led Terminal-Bench 2.1 with 87.4%, compared with Gemini’s 85.8%.
On OSWorld-2.0, which evaluates agentic computer use, GPT-5.6 Terra scored 50.2%, compared with Gemini’s 47.9%.
Where Competitors Still Lead
| Benchmark | Leader | Score |
|---|---|---|
| DeepSWE v1.1 | GPT-5.6 Terra | 69.6% |
| Terminal-Bench 2.1 | GPT-5.6 Terra | 87.4% |
| OSWorld-2.0 | GPT-5.6 Terra | 50.2% |
| GDPVal-AA v2 | Muse Spark 1.2 | 1,628 |
| Agent’s Last Exam | Claude Sonnet 5 | 33.3% |
The results show that AI-agent performance remains highly dependent on the type of task being tested.
Safety Remains Part Of The Release
Google said Gemini 3.7 Flash ships with updated safeguards covering misuse related to chemical, biological, radiological and nuclear threats, as well as cyber offense.
The company said these safeguards are intended to improve protection against harmful applications while supporting legitimate uses.
This is increasingly important as models become capable of taking actions rather than simply generating text.
A more capable agent can potentially automate useful tasks, but the same ability can increase the consequences of poorly controlled tool access.
What Gemini 3.7 Flash Means For Developers
For developers, the model’s most important advantage may be the combination of coding capability, tool use, long context and relatively low inference costs.
A developer building an AI coding agent could potentially use the model for code generation, debugging, repository analysis and automated workflows.
The one-million-token context window can also reduce the need to repeatedly summarize large projects before asking the model to act.
Potential Developer Applications
| Use Case | Gemini 3.7 Flash Advantage |
|---|---|
| Coding agents | Strong software-engineering scores |
| Web development | High Code Arena score |
| Enterprise automation | 30.4% AutomationBench |
| Document analysis | Improved GDP.pdf performance |
| Legal workflows | Strong Harvey LAB-AA result |
| Long codebases | 1M-token context |
| Multimodal agents | Supports text, image, audio and video |
| AI applications | Tool calling and structured outputs |
The practical value will ultimately depend on real-world reliability rather than benchmark scores alone.
The Bigger Picture
Gemini 3.7 Flash shows how the AI race is increasingly moving from conversational intelligence toward systems capable of completing multi-step work. Google’s latest model delivers major improvements over Gemini 3.6 Flash in coding, web development, enterprise automation, document understanding and long-context tasks. Its 1,588 Code Arena score and 30.4% AutomationBench result lead the comparison table published by Google, while its 65.3% DeepSWE and 85.8% Terminal-Bench scores place it close to the strongest competing models.
However, Gemini 3.7 Flash is not the undisputed leader across every AI-agent benchmark. GPT-5.6 Terra remains ahead on DeepSWE v1.1, Terminal-Bench 2.1 and OSWorld-2.0, while other models lead some knowledge-work and desktop-agent evaluations. The more significant development may therefore be the combination of competitive performance and Google’s aggressive introductory pricing, which could make sophisticated AI agents cheaper to deploy at scale.
Looking Ahead
Google’s next challenge will be proving that Gemini 3.7 Flash’s benchmark gains translate into reliable performance in real-world agent deployments. Developers will be looking closely at how consistently the model can plan, use tools, recover from errors and complete long-running tasks without human intervention. Its large context window and strong coding results could make it particularly attractive for software-development agents and enterprise automation.
The economics of agentic AI will also become increasingly important. Google’s introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens gives developers an incentive to experiment through the end of 2026, but those prices are scheduled to double in 2027. If Gemini 3.7 Flash can maintain strong real-world performance while keeping inference costs competitive, Google could strengthen its position in the rapidly expanding market for AI agents
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
