Gemini 3.6 Flash: pricing, API specs, and production evaluation
Gemini 3.6 Flash combines a 1M-token context window, multimodal input, and agent tools. Here is how to evaluate it against newer models on your own workload.
Gemini 3.6 Flash is Google's stable July 2026 model for fast, tool-driven, multimodal workloads. It accepts text, images, video, audio, and PDFs, supports function calling and structured output, and provides a 1,048,576-token input window. It is still available, although Gemini 3.7 Flash is now the newer Flash model.
That makes 3.6 Flash a production candidate, not an automatic default. Existing deployments should evaluate the cost and regression risk of moving. New builds should test 3.6 and 3.7 on the same application traces.
Gemini 3.6 Flash specs at a glance
| Specification | Gemini 3.6 Flash |
|---|---|
| Stable API model ID | gemini-3.6-flash |
| Input context | 1,048,576 tokens |
| Maximum output | 65,536 tokens |
| Inputs | Text, image, video, audio, PDF |
| Output | Text |
| Agent features | Function calling, structured output, code execution |
| Grounding | Google Search and Maps |
| Other tools | File search, URL context, context caching |
| Preview capability | Computer use |
| Not supported | Live API, native image generation, native audio generation |
Google's current promotional Standard price, through December 31, 2026, is $0.75 per million input tokens and $3.75 per million output tokens. Output pricing includes thinking tokens. Google says the rate becomes $1.50 input and $7.50 output on January 1, 2027, so capacity planning should use the post-promo price as well as today's bill.
Where Gemini 3.6 Flash fits
The model is attractive when one workflow mixes long documents, screenshots, audio, video, code, and tools. Likely use cases include:
- Reviewing a codebase and iterating through tool-assisted fixes
- Extracting evidence from long PDFs and mixed-media records
- Grounded research with Search or Maps
- Agents that must return validated JSON or call several functions
- High-volume tasks where a Pro-class model is too expensive
A million-token window does not make retrieval design optional. Sending all available context can increase cost, latency, and distraction. Compare full context, retrieved context, and cached-prefix strategies on identical cases.
What changed for API integrations
Gemini 3.x migrations require more than swapping the model name. Google's
current guidance deprecates custom temperature, top_p, and top_k values.
It also replaces older thinking-budget controls with thinking levels and does
not support prefilled model turns.
Before moving production traffic:
- Change the model ID and remove unsupported sampling parameters.
- Replay representative multi-turn conversations.
- Validate every function schema and tool result.
- Compare thinking levels for quality, latency, and output-token cost.
- Test timeouts, refusals, malformed tool arguments, and long-context recall.
Do not choose a model from benchmark averages
Vendor benchmarks can identify promising candidates, but they cannot reproduce your prompts, retrieval, tools, permissions, users, or definition of success. A model that improves a coding benchmark may still regress on your support policy, tool ordering, or regional language mix.
Build a fixed evaluation set from production traces. Score task completion, tool accuracy, groundedness, policy compliance, latency, and cost per successful outcome. Keep the harness constant while changing only the model and thinking level.
Monitor the model after deployment
Offline evals cover known cases. Production exposes new phrasing, stale knowledge, provider errors, and tool edge cases. Record the actual model version, context size, thinking configuration, tool sequence, token usage, latency, and final outcome for every run.
Currai connects those traces to evaluations, violations, errors, and user outcomes. Use the Currai integration skill to compare Gemini 3.6 Flash with 3.7 Flash or another provider on the workload that matters.