project
Gemini 3.5 Flash - Google's next-generation AI large-scale model
Gemini 3.5 Flash is Google's next-generation AI model, focusing on cutting-edge intelligence and action capabilities. The model achieved a 76.2% score on the Terminal-Bench 2.1 coding benchmark and a GDPval-AA agent task score of 16...
What is Gemini 3.5 Flash?
Gemini 3.5 Flash is Google's next-generation AI model, positioned with cutting-edge intelligence and action capabilities as its core. The model achieves a 76.2% score on the Terminal-Bench 2.1 coding benchmark, a 1656 Elo score on the GDPval-AA agent task, and 83.6% on the MCP Atlas, all surpassing Gemini 3.1 Pro. It boasts output speed four times faster than leading-edge models, at less than half the cost, supports 1 million tokens in multimodal context, and is globally available on the Gemini App and Search AI Mode.
Main functions of Gemini 3.5 Flash
- Rapid Reasoning and GenerationOutput speed reaches 289 tokens/second, which can be improved by 12 times after joint optimization within Antigravity, significantly reducing the latency of Agent multi-round calls.
- Top-level coding and agent capabilitiesIt outperforms Gemini 3.1 Pro across all benchmarks, including Terminal-Bench 2.1 (76.2%), SWE-Bench Pro (55.1%), MCP Atlas (83.6%), and Toolathlon (56.5%).
- Multimodal long context understandingIt natively supports 1 million token input windows, scores 84.2% on CharXiv Reasoning and 83.6% on MMMU-Pro, and can uniformly process text, images, videos, audio, and PDFs.
- Sub-agent collaborative orchestrationThe official demo demonstrates the deployment of parallel subagents using Antigravity 2.0, showcasing 93 subagents working in parallel to complete a full operating system development within 12 hours for less than $1,000.
- Dynamic inference level control:support
thinking_levelParameters (standard/extended) allow users to switch the depth of thought as needed; code and tool calls use the low-level optimization speed by default.
Technical principles of Gemini 3.5 Flash
- Model distillation and compression architectureBased on the new architecture of the Gemini 3.5 series, it compresses the model size through advanced distillation technology, achieving extremely low latency while maintaining near-Pro-level capabilities.
- Dynamic reasoning budget allocation: Introduction
thinking_levelThe mechanism replaces the fixed budget and automatically allocates computing resources based on task type. Code generation and tool invocation use the low setting by default. - Native Multimodal Unified CodingIt employs a unified multimodal encoder to embed text, images, audio, and video into the same semantic space, optimizing attention computation for million-level long sequences.
- Advanced Security Alignment FrameworkFollowing the Frontier Safety Framework, we enhance network and CBRN security protection, and use interpretability tools to examine internal inference, reducing harmful outputs and false rejection rates.
How to use Gemini 3.5 Flash
- Select access platform Access the model via the Gemini App, Google Search AI Mode, Google AI Studio, and the Antigravity 2.0 desktop application.
- Update API model identifierReplace the model name in the code with the official version identifier.
gemini-3.5-flash. - Configure inference level: Set according to task type
thinking_levelParameters: Select low for code and tool calls, medium for general tasks, and high for complex reasoning. - Remove sampling parameter overwriteRemove the manually set default values for temperature, top_p, and top_k, allowing the model to use its internal optimized configuration.
- Note the exceptions to function migration.If the workflow relies on the Computer Use browser's control functions, you need to continue using it.
gemini-3-flash-previewThis feature is not currently supported in Flash 3.5.
The core advantages of Gemini 3.5 Flash
- Speed LeadingOutput speed reaches 289 tokens/second, which is 4 times that of other cutting-edge models. After internal optimization by Antigravity, it can be improved by 12 times. Agent multi-round calls have almost no perceptible latency.
- Cost advantageThe API is priced at $1.50/$9.00 per million tokens (input/output), which is more than half the price of frontier competitors such as Claude Opus 4.7 and GPT-5.5. Google estimates that leading cloud customers can save more than $1 billion annually by migrating 80% of their frontier workloads to 3.5 Flash.
- Agent performance is outstandingMCP Atlas scored 83.6% and Toolathlon 56.5%, supports 93 sub-agents working in parallel, and completed the development of the entire operating system in 12 hours for less than $1,000. It is optimized for complex, long-cycle agent tasks.
- Extra Long Context WindowIt supports 1 million tokens of multimodal input and can process an entire novel, a hundred-page legal contract, or long video content at once, making long document and video analysis effortless.
- Full platform ecosystem coverageIt has been integrated with all Google products, including Gemini App, Search AI Mode, Antigravity 2.0, AI Studio, and Android Studio, and is available to consumers, developers, and enterprises simultaneously.
Project address for Gemini 3.5 Flash
- 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 competing products
| Comparison Dimensions | 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 AssistantIts low latency feature makes it suitable for IDE auto-completion, code review, and interactive debugging; Antigravity 2.0 has set it as the default model.
- Enterprise-level Agent AutomationDeploy parallel sub-agents to process multi-week workflows, such as Xero's automated management of 1099 tax form vendor identification and information collection.
- Intelligent review of financial documentsMacquarie Bank is used to reason about complex documents of 100 pages or more, accelerating client onboarding and compliance reviews.
- Multimodal Invoices and OCR ProcessingRamp combines multimodal understanding with historical pattern reasoning to achieve intelligent recognition and classification of complex invoices.