The Gemini 3.6 Flash API gives developers programmatic access to Google’s generally available Flash model for coding agents, multimodal analysis, long-context knowledge work, and tool-driven automation. Google released the stable gemini-3.6-flash model on July 21, 2026. Unlike the earlier gemini-3-flash-preview, this is a stable model ID intended for ongoing application use.
Gemini 3.6 Flash accepts text, images, video, audio, and PDF files, but it returns text only. Its 1,048,576-token input window can hold large repositories, long document sets, or extended media inputs, while the 65,536-token output limit supports detailed reports, code, and structured responses. Google lists caching, structured output, function calling, code execution, file search, URL context, Search grounding, Maps grounding, and configurable thinking among the underlying model’s capabilities.
The capability boundaries matter. Gemini 3.6 Flash does not generate images or audio, and it does not support the Live API. Computer use is available as a preview feature rather than a fully stable capability. When an integration depends on a Google-native tool, confirm that the GPTProto compatibility layer exposes it before migration. If your product requires speech-to-speech interaction, native image generation, or strict stability for desktop automation, route that part of the workload to another model.
| Specification | Gemini 3.6 Flash |
|---|---|
| Provider | |
| Status | Generally available |
| Stable model ID | gemini-3.6-flash |
| Input types | Text, image, video, audio, PDF |
| Output type | Text |
| Input token limit | 1,048,576 |
| Output token limit | 65,536 |
| Thinking levels | minimal, low, medium, high |
| Default thinking level | medium |
| Google-listed tools | Function calling, code execution, file search, Search, Maps, URL context |
| Computer use | Supported in preview |
| Not supported | Image generation, audio generation, Live API |
Where Gemini 3.6 Flash Fits Best
Gemini 3.6 Flash is most useful when a task needs both a large working context and repeated tool calls. It is not automatically the cheapest choice for every short prompt, and its default medium thinking level can spend more tokens than a non-reasoning model on simple classification. Match the thinking level and model route to the actual job.
| Workload | Why it fits | Implementation note |
|---|---|---|
| Repository coding agents | Fewer unwanted edits and execution loops than Gemini 3.5 Flash in Google’s testing | Use medium or high thinking and validate changes with tests |
| Multimodal document analysis | Reads PDFs, images, charts, audio, and video in the same request | Ask for structured output when downstream systems need stable fields |
| Long-context research | 1M input tokens and supported context caching | Cache repeated corpora instead of resending unchanged files |
| Browser or desktop automation | Native computer-use tooling and 83.0% OSWorld-Verified in Google’s evaluation | Treat computer use as preview and keep approval gates for consequential actions |
| Grounded assistants | The underlying model supports Search, Maps, file search, and URL context | Enable only the tools each request needs and track resulting token use |
Gemini 3.6 Flash vs Gemini 3.5 Flash
Gemini 3.6 Flash is a direct efficiency upgrade over Gemini 3.5 Flash rather than a larger-context replacement. Both models keep the same 1M context class and default medium thinking level. The difference is task efficiency: Google reports 17% fewer output tokens on Artificial Analysis workloads, fewer reasoning turns and tool calls, and stronger results in coding, machine-learning engineering, computer use, and long-context retrieval.
The official input rate remains $1.50 per 1M tokens, while the official output rate falls from $9.00 to $7.50. On GPTProto, Gemini 3.6 Flash costs $0.90 per 1M input tokens and $4.50 per 1M output tokens; Gemini 3.5 Flash is currently $0.90 and $5.40. For new agentic or multimodal systems, 3.6 Flash is the better default. Keep 3.5 Flash only when you have already validated its behavior and need time to regression-test the migration.
| Metric | Gemini 3.6 Flash | Gemini 3.5 Flash |
|---|---|---|
| Context window | 1,048,576 | 1,048,576 |
| Max output | 65,536 | 65,536 |
| Default thinking | Medium | Medium |
| Google standard input / output per 1M | $1.50 / $7.50 | $1.50 / $9.00 |
| GPTProto input / output per 1M | $0.90 / $4.50 | $0.90 / $5.40 |
| DeepSWE v1.1 | 49.0% | 37.0% |
| MLE-Bench | 63.9% | 49.7% |
| OSWorld-Verified | 83.0% | 78.4% |
| GDM-MRCR v2 at 1M | 54.0% | 26.6% |
Benchmark figures above are provider-reported. Use them to choose candidates, then run both models on your own prompts, tools, and success criteria before moving production traffic.
API Changes to Check Before You Migrate
Do not treat migration to 3.6 Flash as only a model-name replacement if your application uses Google’s native Interactions API. Google changed several generation and conversation fields for the latest models:
- Change the model ID to
gemini-3.6-flash. - Remove
temperature,top_p, andtop_kfrom generation configuration. - Replace
thinking_budgetwiththinking_level; supported values areminimal,low,medium, andhigh. - Remove
candidate_count, which Gemini 3.x does not support. - Do not send prefilled model turns. For multi-turn Interactions API sessions, use the server-side
previous_interaction_id. - Regression-test tool schemas, structured outputs, and file handling before shifting all traffic.
If you use GPTProto’s OpenAI-compatible request format, follow the Quick Start and GPTProto documentation on this page rather than copying Google-native fields directly. The practical advantage is that one GPTProto API key and balance can also call Gemini 3.5 Flash, GPT-5.6 Luna, Claude Opus 4.8, Kimi K3, and 200+ other models, so you can run the same evaluation set without opening separate provider accounts.
