On August 13, 2026, Google launched Gemini 3.7 Flash, a new version of its model designed for intensive everyday use, with a clear focus on coding, agent-based workflows, and enterprise knowledge tasks. The release comes just three weeks after Gemini 3.6 Flash, and Google said the rapid update cycle is linked to developer feedback and algorithmic improvements.
The company is temporarily cutting the model’s API usage prices by 50%; through December 31, 2026, the cost will be approximately $0.75 per million input tokens and $3.75 per million output tokens. Starting January 1, 2027, prices will rise to $1.50 for input and $7.50 for output, while context caching will cost $0.075 per million tokens during the introductory period before rising to $0.15.
Focus on Executing Multistep Tasks
Google describes Gemini 3.7 Flash as its “smartest working model” yet for coding and agents. According to the company, the model is now better at handling obstacles, asking for clarification when necessary, and following instructions more precisely. It also says the model “thinks more carefully” when planning multistep tasks and executing tool calls.
These improvements are intended to reduce repeated attempts and human intervention in coding and business agents. In an enterprise development environment, the model could help reduce unnecessary modifications, recover from errors, and execute complex plans more reliably. The same applies to agents working across multiple documents and applications, where an incorrect tool call or inaccurate interpretation of instructions can halt a workflow.
Google DeepMind said the new release makes progress in debugging and problem-solving, creating more complete web layouts and applications with fewer prompts, and improving reasoning and accuracy in real-world business workflows.
Strong Coding Improvement Without an Overall Lead
According to Google’s tests, Gemini 3.7 Flash scored 43.6% on the FrontierCode 1.1 Main benchmark, which measures production-code quality, compared with 34.4% for version 3.6. It also slightly exceeded the scores of Claude Sonnet 5, at 42.7%, and GPT-5.6 Terra, at 41.3%, according to Google’s table.
In the DeepSWE v1.1 benchmark for long-horizon coding tasks, the model scored 65.3% versus 49.0% for its predecessor, but remained behind GPT-5.6 Terra, which scored 69.6%. In web development, it achieved an Elo rating of 1588 in Code Arena, compared with 1538 for Gemini 3.6 Flash, 1541 for Claude Sonnet 5, and 1523 for GPT-5.6 Terra.
The results do not show an overall lead in every area. The model scored 85.8% on Terminal-bench 2.1, compared with 87.4% for GPT-5.6 Terra, while Terra also outperformed it on Terminal-bench 3.0 and OSWorld-2.0. In Agent's Last Exam, which covers office and multimodal tasks, Claude Sonnet 5 led with a success rate of 33.3%, compared with 26.3% for Gemini 3.7 Flash.
Enterprise Uses and Competitive Pricing
The announced improvements extend to business automation and document understanding. Gemini 3.7 Flash scored 30.4% on AutomationBench, compared with 17.0% for the previous version, 23.6% for GPT-5.6 Terra, and 10.7% for Claude Sonnet 5, according to Google’s data. It also achieved 34.0% on the GDP.PDF benchmark for understanding complex PDF files, compared with 22.0% for Gemini 3.6 Flash, 28.0% for Claude Sonnet 5, and 24.7% for GPT-5.6 Terra.
Google is making the model available to Google AI Pro and Ultra subscribers inside Spark, its personal artificial-intelligence agent, with improvements for knowledge tasks and the use of Google Workspace tools, such as gathering files, drafting emails, and updating status documents. It is also available to enterprises through the Gemini Enterprise Agent Platform and the Gemini Enterprise application.
The introductory prices stand out in the competition to run autonomous agents because a single request may generate a long sequence of model calls, reasoning tokens, and tool interactions. The article explains that a lower token price does not necessarily mean a lower final cost if the model requires repeated attempts; the practical metric will therefore be the cost of successfully completing a task, not the price per million tokens alone.
Availability and the Model’s Launch Context
Developers can access Gemini 3.7 Flash through the Gemini API in Google AI Studio and Android Studio, as well as Google’s Antigravity environment. The model is available to enterprises through the Gemini Enterprise Agent Platform and Gemini Enterprise, and to consumers subscribed to Google AI Pro or Ultra through Spark in supported countries. Google said it has also launched updated safeguards covering chemical, biological, radiological, and nuclear risks, in addition to misuse of the model in cyberattacks.
The release comes at a time when Google has not announced a date for making Gemini 3.5 Pro available, which remained under partner testing according to the article, while Gemini 3.1 Pro remains the latest generally available Pro model the company has introduced. Gemini 3.7 Flash therefore strengthens Google’s fast-model lineup, but does not settle the competition over the leading model.
For development teams, the performance increase and introductory pricing provide an incentive to test the model on their own repositories, prompts, tool schemas, and failure cases before moving any production application to it. How long the economic advantage will last after prices return to their standard levels in January 2027 will depend on the model’s ability to complete real-world tasks consistently and with fewer attempts.