Google Introduces Workhorse Model Gemini 3.7 Flash
Google officially released Gemini 3.7 Flash on August 13, 2026, focusing on long-horizon coding, complex reasoning, and agentic workflows.
Arriving just three weeks after Gemini 3.6 Flash, this release highlights Google's rapid monthly cadence for its flagship Flash series. While maintaining high inference throughput, the new model is specifically tuned for multi-step autonomous planning and external tool interactions.
Starting today, developers can access the model via Google AI Studio, Vertex AI, and the Gemini API. In addition, Gemini 3.7 Flash is integrated into GitHub Copilot and serves as the underlying engine powering Google's personal AI agent platform, Gemini Spark.
Architecture Features and Tunable Thinking Modes
Gemini 3.7 Flash features a standard 1M-token context window and supports up to 65,536 output tokens in a single request.
The model maintains a native multimodal architecture, supporting direct input of text, code, images, audio, video, and dense PDF documents. At the system level, Google implemented architectural optimizations for Function Calling, structured JSON outputs, and Context Caching, substantially reducing latency and cost in multi-turn interactions.
Furthermore, the model natively integrates a tunable thinking mechanism with Low, Medium, and High budget configurations. Developers can flexibly allocate compute resources between ultra-low latency responses and deep multi-step chain-of-thought reasoning depending on their specific workload requirements.
Major Gains in Software Engineering and Agent Benchmarks
Across key engineering benchmarks such as DeepSWE and FrontierCode, Gemini 3.7 Flash achieves major performance gains over 3.6 Flash.
Official benchmark results demonstrate significant progress in multi-file refactoring, full-stack web development, and dense document comprehension:
| Benchmark | Evaluation Dimension / Scenario | Gemini 3.6 Flash | Gemini 3.7 Flash | Improvement |
|---|---|---|---|---|
| DeepSWE v1.1 | Long-horizon software engineering issue resolution | 49.0% | 65.3% | +16.3% |
| FrontierCode 1.1 Main | Production code generation and bug fixing | 34.4% | 43.6% | +9.2% |
| WebDev Arena (Elo) | End-to-end full-stack web application development | 1538 | 1588 | +50 Elo |
| GDP.pdf | Dense financial, legal, and bioscience document parsing | 22.0% | 34.0% | +12.0% |
| AutomationBench | Enterprise multi-step workflow automation execution | 17.0% | 30.4% | +13.4% |
On the DeepSWE v1.1 benchmark for multi-file collaborative editing and complex bug fixing, Gemini 3.7 Flash scores 65.3%, significantly reducing infinite loops and hallucinations during agent execution; on GDP.pdf for long document comprehension, accuracy rises from 22.0% to 34.0%.
Pricing Strategy and Ecosystem Integration
To accelerate enterprise and developer adoption, Google launched a 50% introductory discount, lowering API pricing significantly.
The official pricing schedule defines the introductory promotional rates and subsequent standard rates:
| Pricing Tier | Effective Period | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|---|
| Introductory Period | Through December 31, 2026 | $0.75 | $3.75 |
| Standard Pricing | Starting January 1, 2027 | $1.50 | $7.50 |
Developers can test tunable thinking configurations in Google AI Studio today or deploy production workloads via Vertex AI with enterprise-grade concurrency guarantees.