Gemini 3.5 Flash
Gemini 3.5 Flash is Google's stable Flash model for sustained frontier performance on agentic coding, long-horizon workflows, and multimodal tasks. The Gemini API model ID is gemini-3.5-flash, with 1,048,576 input tokens and 65,536 output tokens.
GoogleWhat Gemini 3.5 Flash is built for
Google positions Gemini 3.5 Flash as a speed-focused model that keeps high reasoning quality for real-world agent tasks, especially coding loops and tool-heavy workflows.
Agentic coding loops
Use Gemini 3.5 Flash when an agent needs to plan, edit, run tools, inspect results, and iterate quickly without jumping immediately to a Pro-class model.
Long-horizon execution
The 1M input window makes it practical to include large repositories, PDFs, logs, recordings, screenshots, and task history in a single model session.
Multimodal work at scale
The model accepts text, images, video, audio, and PDF inputs, then returns text that can drive structured outputs, function calls, search-grounded answers, and code execution.
Official Gemini 3.5 Flash benchmark results
Google DeepMind reports broad gains over Gemini 3 Flash across agentic coding, MCP workflows, UI control, expert tasks, multimodal understanding, long context, and reasoning.
| Benchmark | What it measures | Gemini 3.5 Flash | Gemini 3 Flash | Selection signal |
|---|---|---|---|---|
| Terminal-bench 2.1 | Agentic terminal coding with the Terminus-2 harness | 76.2% | 58.0% | +18.2 points for terminal coding agents |
| SWE-Bench Pro (Public) | Diverse single-attempt agentic coding tasks | 55.1% | 49.6% | +5.5 points on harder software tasks |
| MCP Atlas | Multi-step workflows using Model Context Protocol tools | 83.6% | 62.0% | +21.6 points for tool workflows |
| Toolathlon | Real-world general tool use | 56.5% | 49.4% | +7.1 points on general tool use |
| OSWorld-Verified | Agentic computer and UI control | 78.4% | 65.1% | +13.3 points for UI-control tasks |
| Finance Agent v2 | Financial analysis and decision-making | 57.9% | 42.6% | +15.3 points on expert finance workflows |
| MMMU-Pro | Multimodal understanding and reasoning with no tools | 83.6% | 81.2% | +2.4 points on multimodal reasoning |
| MRCR v2 (8-needle, 128K average) | Long-context retrieval and recall | 77.3% | 67.2% | +10.1 points at 128K average context |
| Humanity's Last Exam | Academic reasoning across text and multimodal tasks | 40.2% | 33.7% | +6.5 points on difficult reasoning |
| ARC-AGI-2 | Abstract reasoning puzzles | 72.1% | 33.6% | +38.5 points in the official table |
Scores follow Google DeepMind's Gemini 3.5 Flash performance table and methodology link checked on June 17, 2026. Vendor benchmark harnesses differ, so validate with your own prompts, tools, latency limits, and budget before production use.
Why the benchmark gains matter for agents
The official DeepMind page also includes customer evaluation signals that are useful when deciding whether Flash is enough for production agent work.
Cyber defense token efficiency
Armadin reported Gemini 3.5 Flash performing 42% better than Flash 3 on a long-range, multi-turn cyber benchmark while improving token efficiency by 68%.
Enterprise work evaluation
Box reported Gemini 3.5 Flash beating Gemini 3 Flash by 19.6% on its enterprise work evaluation set for real-world, multi-step customer tasks.
Domain-specific accuracy
Box also reported 96.4% greater accuracy for Life Sciences data extraction and calculation workflows, plus 46.7% greater accuracy for building financial reports from structured data.
Official Gemini API pricing for Gemini 3.5 Flash
Google lists Gemini 3.5 Flash pricing per 1M tokens in USD. Output prices include thinking tokens. Search or Maps grounding can add separate query charges after the included monthly allowance.
| Tier | Paid input price | Paid output price | Context caching | Best fit |
|---|---|---|---|---|
| Standard | $0.25 text / image / video; $0.50 audio | $1.50 | $0.025 text / image / video; $0.05 audio | Default production calls where price and latency are both important. |
| Batch | $0.125 text / image / video; $0.25 audio | $0.75 | $0.0125 text / image / video; $0.025 audio; $0.50 per 1M tokens per hour storage | Offline jobs, evaluations, bulk analysis, and lower-cost asynchronous processing. |
| Flex | $0.125 text / image / video; $0.25 audio | $0.75 | $0.0125 text / image / video; $0.025 audio; $0.50 per 1M tokens per hour storage | Cost-optimized workloads that can tolerate flexible serving behavior. |
| Priority | $0.45 text / image / video; $0.90 audio | $2.70 | $0.045 text / image / video; $0.09 audio; $1.80 per 1M tokens per hour storage | Latency-sensitive workflows where priority serving is worth the higher token price. |
Google lists 5,000 Grounding with Google Search or Google Maps prompts per month free, shared across Gemini 3, then $14 per 1,000 search queries. Prices can change, so verify the official Gemini API pricing page before committing production spend.
Is there a Gemini 3.5 Flash Lite?
As of the official Google docs checked on June 17, 2026, the stable lightweight sibling is Gemini 3.1 Flash-Lite, not a model named gemini-3.5-flash-lite. Use it as the cheaper router or extraction model, then escalate hard tasks to Gemini 3.5 Flash.
| Decision point | Gemini 3.5 Flash | Gemini 3.1 Flash-Lite |
|---|---|---|
| Official model ID | gemini-3.5-flash | gemini-3.1-flash-lite |
| Release status | Generally available on May 19, 2026; stable in the Gemini API model docs. | Preview launched March 3, 2026; generally available on May 7, 2026. |
| Best use | Agentic coding, long-horizon execution, complex tool workflows, and higher reasoning quality at Flash latency. | High-frequency translation, transcription, classification, extraction, document triage, and model routing. |
| Token limits | 1,048,576 input tokens and 65,536 output tokens. | 1,048,576 input tokens and 65,536 output tokens. |
| Model routing pattern | Use as the main worker for complex tasks that need stronger agentic reasoning. | Use as the low-latency classifier or first-pass extraction model before escalating to Flash or Pro. |
The keyword “gemini 3.5 flash lite” usually maps to a comparison question. Google's current developer docs list Gemini 3.1 Flash-Lite as the stable Flash-Lite model in the Gemini 3 family.
Where Gemini 3.5 Flash fits in MyClaw agents
Start with Gemini 3.5 Flash when you need more intelligence than a routing model, but still care about cost, speed, and scale.
Repository implementation work
Give an agent files, logs, issues, and commands, then use Gemini 3.5 Flash for fast code-edit-test loops.
Grounded research and synthesis
Combine search grounding, file search, URL context, and code execution for research tasks that need evidence and structured output.
Multimodal operations
Analyze PDFs, screenshots, audio, video, and text in the same session before producing text that can drive tools.
Cost-aware routing
Route simple work to Flash-Lite, keep most agent work on Gemini 3.5 Flash, and reserve Pro-class models for the hardest prompts.
Gemini 3.5 Flash FAQs
Gemini 3.5 Flash
Evaluate Gemini 3.5 Flash in a hosted MyClaw agent
Use MyClaw to test Gemini 3.5 Flash with files, tools, long-context prompts, grounded research, and coding workflows before choosing the production model mix.
Run Gemini 3.5 Flash in MyClaw