What is Gemini 3.5 Flash
Gemini 3.5 Flash is Google's new generation AI large model, positioned around frontier intelligence + action capabilities as its core. The model scores 76.2% on the Terminal-Bench 2.1 coding benchmark, 1656 Elo on GDPval-AA agent tasks, and 83.6% on MCP Atlas, all surpassing Gemini 3.1 Pro, with output speed reaching 4 times that of frontier models at less than half the cost, supporting a 1 million Token multimodal context, and has been launched globally on Gemini App and Search AI Mode.
Main Features of Gemini 3.5 Flash
- Ultra-fast Reasoning and Generation: Output speed reaches 289 tokens/second, and after internal joint optimization via Antigravity it can achieve up to a 12x improvement, significantly reducing Agent multi-turn call latency.
- Top-tier Coding and Agent Capabilities: Comprehensively surpasses Gemini 3.1 Pro on benchmarks including Terminal-Bench 2.1 (76.2%), SWE-Bench Pro (55.1%), MCP Atlas (83.6%), and Toolathlon (56.5%).
- Multimodal Long-context Understanding: Natively supports a 1 million Token input window, scores 84.2% on CharXiv Reasoning and 83.6% on MMMU-Pro, and can uniformly process text, images, video, audio, and PDF.
- Sub-agent Collaborative Orchestration: Deploys parallel sub-agents via Antigravity 2.0; the official demo had 93 subagents working in parallel to complete full operating system development in under 12 hours for less than $1000.
- Dynamic Reasoning Level Control: Supports the
thinking_levelparameter (standard/extended); users can switch thinking depth as needed, and code and tool calls use the low level by default to optimize speed.
Technical Principles of Gemini 3.5 Flash
- Model Distillation and Compression Architecture: Based on the all-new architecture of the Gemini 3.5 series, it compresses model size through advanced distillation techniques, achieving extremely low latency while maintaining capabilities close to the Pro level.
- Dynamic Reasoning Budget Allocation: Introduces the
thinking_levelmechanism to replace fixed budgets, automatically allocating computational resources based on task type, with code generation and tool calls using the low level by default. - Native Multimodal Unified Encoding: Uses a unified multimodal encoder to embed text, images, audio, and video into the same semantic space, optimizing attention computation for million-level long sequences.
- Frontier Safety Alignment Framework: Follows the Frontier Safety Framework, strengthens cyber and CBRN safety protection, inspects internal reasoning through interpretability tools, and reduces harmful output and false refusal rates.
How to Use Gemini 3.5 Flash
- Choose an Access Platform Access the model through the Gemini App、Google Search AI Mode、Google AI Studio、Antigravity 2.0 desktop application.
- Update the API Model Identifier: Replace the model name in the code from the old preview identifier with the official version
gemini-3.5-flash。 - Configure the Reasoning Level: Set the
thinking_levelparameter according to task type: choose low for code and tool calls, medium for general tasks, and high for complex reasoning. - Remove Sampling Parameter Overrides: Delete manually set default values for temperature, top_p, and top_k so the model uses its internally optimized configuration.
- Note Feature Migration Exceptions: If the workflow depends on the Computer Use browser control feature, you need to continue using
gemini-3-flash-preview; 3.5 Flash does not currently support this feature.
Core Advantages of Gemini 3.5 Flash
- Leading Speed: Output speed reaches 289 tokens/second, 4 times that of other frontier models, and up to a 12-fold improvement after Antigravity's internal optimization, with virtually imperceptible latency in multi-turn Agent calls.
- Cost advantage: API pricing is $1.50/$9.00 per million tokens (input/output), more than half the price of frontier competitors such as Claude Opus 4.7 and GPT-5.5. Google estimates that top Cloud customers migrating 80% of frontier workloads to 3.5 Flash can save more than $1 billion per year.
- Outstanding Agent performance: MCP Atlas score of 83.6%, Toolathlon 56.5%, support for 93 sub-agents collaborating in parallel, and completion of full operating system development in under 12 hours for less than $1,000, optimized specifically for complex long-horizon agent tasks.
- Ultra-long context window: Supports 1 million tokens of multimodal input, can process an entire long novel, a hundred-page legal contract, or long video content in one go, with no pressure in long-document and video analysis.
- Full-platform ecosystem coverage: Already integrated into the full range of Google products including the Gemini App, Search AI Mode, Antigravity 2.0, AI Studio, and Android Studio, and simultaneously available to consumers, developers, and enterprises.
Gemini 3.5 Flash project address
- Project official website：https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/#gemini-3-5-flash
Comparison of Gemini 3.5 Flash with similar competitors
| Comparison dimension | Gemini 3.5 Flash | Claude Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Output speed | ~289 t/s（4x） | ~67 t/s | ~71 t/s | ~135 t/s |
| Terminal-Bench 2.1 | 76.2% | – | 78.2% | 70.3% |
| SWE-Bench Pro | 55.1% | – | 58.6% | 54.2% |
| MCP Atlas | 83.6% (leading) | 79.1% | 75.3% | 78.2% |
| GDPval-AA Elo | 1656 | 1753 | 1769 | 1314 |
| CharXiv Reasoning | 84.2% | 82.1% | 84.1% | 83.3% |
| Context window | 1M Token | 200K | 128K | 1M |
Application scenarios of Gemini 3.5 Flash
- Real-time AI programming assistant: Its low-latency characteristics make it suitable for IDE autocompletion, code review, and interactive debugging, and Antigravity 2.0 has set it as the default model.
- Enterprise-grade Agent automation: Deploys parallel sub-agents to handle multi-week workflows, such as Xero automatically managing 1099 tax form vendor identification and information collection.
- Intelligent review of financial documents: Macquarie Bank uses it to reason over complex documents of more than 100 pages, accelerating customer onboarding and compliance review.
- Multimodal invoice and OCR processing: Ramp combines multimodal understanding with historical pattern reasoning to achieve intelligent recognition and classification of complex invoices.