Gemini 3.7 Flash
Google has released Gemini 3.7 Flash, an AI model specifically optimized for complex software engineering, autonomous agent workflows, and web development. It arrives as a successor to Gemini 3.6 Flash with significant gains in reasoning, code generation, and cost efficiency.
Key Technical Improvements
Coding & Development: Gemini 3.7 Flash shows substantial performance leaps in software engineering tasks. It has higher "first-pass" code accuracy and generates more production-ready code.
FrontierCode 1.1: Scored 43.6% (up from 34.4% in 3.6 Flash).
DeepSWE v1.1: Scored 65.3% (up from 49.0%).
Web Development: Achieved an Elo score of 1588 on Arena. Ai WebDev Arena, demonstrating improved layout functionality and reference-matching (design-to-code) accuracy.
Reasoning & Agentic Workflows: The model is designed to "think more diligently" during multi-step planning and tool-calling tasks. It is more capable of handling roadblocks, clarifying intent, and following complex instructions, which reduces the need for manual developer oversight and retries.
Flexible Reasoning Control: Developers can now adjust the model’s "thinking level" via the API:
Low: Optimized for latency-critical tasks like incident response and fast data analysis.
Medium (Default): Balanced for most complex code and agentic use cases.
High: Maximizes reasoning effort for the most difficult math, coding, and multi-step agent tasks.
Infrastructure & Availability
Context Window: Supports a 1M token context window and 64k maximum output tokens.
Introductory Pricing: Through December 31, 2026, the model is available at $0.75 per million input tokens and $3.75 per million output tokens (half the price of previous-generation Flash models). Standard pricing ($1.50 input / $7.50 output) will take effect starting January 1, 2027.
Integration: The model is being integrated immediately into Google’s "Antigravity" agent platform and is available for developers via the Gemini API.
Migration Notes
Developers migrating to gemini-3.7-flash should note:
Parameter Changes: Deprecated sampling parameters like temperature, top_p, and top_k should be removed in favor of thinking_level.
Validation: Multi-turn conversations should be standardized using previous_interaction_id, and prefilled model turns are no longer supported.
Function Calling: The model requires strict adherence to call_id and name in FunctionResponse objects.