benchr
Google's July 21 release note calls Gemini 3.6 Flash stable and production-ready, with an emphasis on token efficiency, code, and agentic planning. It is useful positioning, not a workload result: Google does not publish a dedicated benchmark table for this model. Routing, tool behavior, and the cost of an accepted result should decide whether it goes into production.
The model page confirms a broad input surface—text, image, video, audio, and PDF—with text output. It lists caching, code execution, file search, function calling, Google Search and Maps grounding, structured outputs, thinking, URL context, and Computer Use in preview. It does not list image generation, audio generation, or the Live API as supported. That boundary should be part of the architecture diagram before migration begins.
| Mode | Input / 1M | Output / 1M | Use it when | Operational caution |
|---|---|---|---|---|
| Standard | $0.75 | $3.75 | A user or service is waiting for the response | Do not pay synchronous rates for queueable work |
| Batch | $0.375 | $1.875 | Jobs can be grouped and completed asynchronously | Design idempotency and result reconciliation first |
| Flex | $0.375 | $1.875 | Latency can vary without harming the product | Measure tail latency, not only the average |
| Priority | $1.35 | $6.75 | Interactive service requirements justify the premium | Keep it on an explicit route, not as a global default |
The production question is lane selection
A Flash deployment should not send every request through one consumption mode. Keep ordinary interactive traffic on standard. Move scheduled enrichment, document indexing, and replayable evaluation runs to batch or flex. Reserve priority for the small route where response-time variance materially changes the user experience. The model can be the same while the service economics change substantially.
Cached input costs $0.075 per million tokens and cache storage is $0.50 per million token-hours, both doubling on January 1, 2027. Caching is not automatically a saving: a prefix must be reused enough, and soon enough, to repay its storage. Track cache hit volume, storage age, and the tokens that would otherwise have been resent. A stable but rarely reused prefix should not be cached merely because the feature exists.
Workload test: replay the agent loop
Take a production-shaped trace and replay it with the exact model ID, tools, schemas, safety settings, and retry policy proposed for launch. Keep the old route as a control. Do not score only the final prose.
- Tool correctness: inspect function selection, arguments, retries, and whether a failed tool response causes a safe recovery or an invented completion.
- Output economy: record billed input, cached input, thinking and output tokens, then calculate cost per accepted result. Google's release note mentions lower verbosity; your trace must show whether that appears in your prompts.
- Contract stability: validate every structured output and verify that downstream code handles refusals, empty fields, and partial tool results.
- Multimodal fidelity: isolate PDF, image, audio, and video cases. A single blended score can hide a failure in the least common but most expensive input type.
- Service behavior: compare median and tail latency by standard, flex, and priority routes. Averages alone cannot select a production lane.
Routing matrix
| Request class | Recommended starting lane | Pass condition | Do not ship when |
|---|---|---|---|
| Interactive coding or tool loop | Standard; priority only after latency evidence | Correct tool calls, valid patch or artifact, bounded retries | The loop claims success after a tool error |
| Scheduled extraction or enrichment | Batch or flex | Schema-valid output and reliable reconciliation | Duplicate or late jobs cannot be safely replayed |
| Large repeated source pack | Standard with measured caching | Reuse exceeds storage and invalidation cost | The prefix changes frequently or contains stale policy |
| Computer Use | Preview-only sandbox | Reversible actions, visible trace, confirmation boundaries | The workflow can create an irreversible external effect |
What not to overclaim
Google's release note says 3.6 Flash improves token efficiency and code or agentic planning. The current model page does not attach a benchmark table to those claims. Any capability score shown elsewhere on benchr is editorial and should not be converted into a provider result. The hard record is the stable model ID, release status, token limits, price modes, and documented features. Quality approval still belongs to the replay test.
FAQ
Which Gemini 3.6 Flash pricing mode should a production service use?
Use standard for ordinary synchronous requests, batch or flex for delay-tolerant queues, and priority only when the higher rate is justified by interactive service requirements. Google lists standard at $0.75 input and $3.75 output per million tokens, batch and flex at $0.375 and $1.875, and priority at $1.35 and $6.75. All of them double on January 1, 2027.
Does Gemini 3.6 Flash have official benchmark scores?
Google's current model page and release note do not provide a dedicated benchmark table for Gemini 3.6 Flash. Treat claims about improved token efficiency and code or agentic planning as provider positioning until they are validated on your own workload.
What should a Gemini 3.6 Flash migration test measure?
Replay production-shaped requests and compare accepted-result cost, output-token volume, tool-call correctness, schema validity, safety behavior, and tail latency. Test multimodal and long-context inputs separately instead of averaging them into one score.
Sources
- Google AI for Developers, Gemini API release notes, re-verified July 29, 2026.
- Google AI for Developers, Gemini API pricing, re-verified July 29, 2026.
- Google AI for Developers, Gemini 3.6 Flash model page, re-verified July 29, 2026.
Changelog
August 24, 2026: Google cut the standard rate to $0.75 / $3.75 through December 31, 2026, matching the Gemini 3.7 Flash introductory price. Every price on this page now reflects the promotional rate and names the January 1, 2027 return to $1.50 / $7.50.
July 29, 2026: Expanded the launch note into a production routing guide with a consumption-mode matrix, cache decision, trace-replay workload test, and explicit limits on benchmark claims.
July 22, 2026: Published after Google listed Gemini 3.6 Flash and Gemini 3.5 Flash-Lite as generally available in the Gemini API release notes.