As of August 25, 2026, gemini-3.7-flash is the most defensible starting point for a new general-purpose Gemini Developer API integration. It is the newest stable Flash model and Google’s current workhorse for coding, agentic workflows, and reliable multi-step execution. Use gemini-3.5-flash-lite when latency and throughput dominate and the result is easy to validate. Treat gemini-3.1-pro-preview as an experiment for the hardest reasoning and tool-use cases, not as an automatic production default.
This guide compares general Gemini models that accept multimodal input and return text. Image generation, embeddings, Live audio, TTS, managed agents, and Gemini app subscriptions are different products or API surfaces.
A practical shortlist
| Situation | Start with | Decision logic |
|---|---|---|
| New app, coding assistant, multi-step tool workflow | gemini-3.7-flash | Newest stable Flash; balanced capability, speed, tools, and lifecycle |
| Stable workload already on 3.6 | Keep gemini-3.6-flash, test 3.7 beside it | Both are GA and share the same 2026 promotional standard price |
| Classification, extraction, translation, document triage | gemini-3.5-flash-lite | Built for low-latency, high-volume bounded work |
| Existing cost-sensitive 3.1 Lite system | gemini-3.1-flash-lite, with a migration plan | Cheaper today, but older and already points toward 3.5 Lite as its successor |
| Difficult reasoning, complex code, custom tool orchestration | Compare gemini-3.1-pro-preview with 3.7 | Pay for Pro only where it produces a measurable gain |
| Gemini 2.5 in production | Begin migration now | Google lists October 16, 2026 as the earliest shutdown date for the main 2.5 text models |
One model does not need to serve every request. A small router can send well-defined, machine-checkable work to Flash-Lite and escalate ambiguous, failed, or tool-heavy requests to 3.7 Flash. That often controls cost better than sending every prompt to the most expensive endpoint.
What the current names actually mean
Google’s model overview lists 3.7 Flash, 3.6 Flash, 3.5 Flash, 3.5 Flash-Lite, and 3.1 Flash-Lite as stable. Gemini 3.1 Pro and Gemini 3 Flash remain Preview. The status label matters more than the marketing tier: the exact model ID is your API contract.
Gemini 3.7 Flash: the default candidate
The 3.7 Flash model page specifies 1,048,576 input tokens and 65,536 output tokens. It accepts text, image, video, audio, and PDF input and returns text. Supported features include caching, code execution, function calling, structured outputs, search grounding, URL context, and computer use in Preview.
Google positions it for coding, agents, and reliable multi-step work. That is provider guidance, not proof that it wins on your private workload. A safe rollout still measures first-pass success, tool-call correctness, P95 latency, and total tokens per accepted result.
Gemini 3.6 Flash: a useful GA fallback
gemini-3.6-flash remains stable and uses the same temporary 2026 standard price as 3.7. Existing systems do not need an emergency migration simply because 3.7 exists. Keep 3.6 available as a fallback, then compare both on the same prompts.
If 3.7 reduces retries or verbose output, it may cost less per successful task even when the listed token price is identical. If it regresses on a particular tool chain, a configuration-driven 3.6 fallback is safer than a code change.
Gemini 3.5 Flash: stable, but no longer the first stop
gemini-3.5-flash has the same published 1M input and 65k output limits, but the current overview describes it as a legacy Flash model. Its standard list price is also higher than the temporary 3.7/3.6 rate. It is best treated as an existing-system baseline or compatibility fallback.
Flash-Lite: work with a clear definition of done
The 3.5 Flash-Lite page targets low-latency, cost-effective, high-throughput execution: extraction, document parsing, simple automation, and subagent tasks. It also publishes 1,048,576 input and 65,536 output token limits. Equal context limits do not imply equal reasoning quality.
Flash-Lite works best when you can validate the output: a fixed label set, a JSON Schema, required extracted fields, or sampled translation review. For architecture, long-horizon debugging, or complex tool planning, one failed attempt can erase the savings from a lower token rate.
gemini-3.1-flash-lite is cheaper at current standard rates. It can remain sensible in an existing measured pipeline, but it is older. Google’s deprecation table identifies 3.5 Flash-Lite as its recommended replacement, so a new long-lived system should include migration cost.
Gemini 3.1 Pro Preview must earn its risk
Google’s current model overview describes 3.1 Pro Preview as an advanced option for complex problem solving, software engineering, and precise tool use. It also offers gemini-3.1-pro-preview-customtools for workflows that mix bash and custom tools.
The final word is the constraint: Preview. Behavior, limits, or contracts can change before a stable release. Select tasks where Flash misses are expensive, run a blinded comparison against 3.7, and route to Pro only if the quality gain covers its higher price, latency, and change risk.
Standard prices on August 25, 2026
These figures come from the official pricing page. They are USD per 1 million tokens for standard paid inference. Output includes thinking tokens. Batch, Flex, and Priority have separate rates.
| Model | Standard input | Standard output, including thinking | Status |
|---|---|---|---|
| Gemini 3.7 Flash | $0.75 | $3.75 | GA; promotional through Dec. 31, 2026, then scheduled for $1.50 / $7.50 |
| Gemini 3.6 Flash | $0.75 | $3.75 | GA; same promotional period |
| Gemini 3.5 Flash | $1.50 | $9.00 | GA, legacy |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | GA |
| Gemini 3.1 Flash-Lite | $0.25 for text/image/video input | $1.50 | GA, older generation |
| Gemini 3.1 Pro Preview | $2.00 at ≤200k prompt; $4.00 above 200k | $12.00; $18.00 | Preview |
The useful unit is:
cost per accepted task = input + thinking/output + tool-related usage + retries + cache storage
Track success rate, P95 latency, input tokens, thinking/output tokens, and retries together. A $0.30 model that retries or calls the wrong tool can cost more than a $0.75 model that succeeds once.
Eligible free-tier models may provide free tokens, but availability and limits vary by model, project, and region; submitted free-tier content may be used to improve Google products. The paid tier says submitted content is not used for that purpose. AI Studio availability is not a promise of unlimited free production use.
Thinking level changes the comparison
Google’s thinking documentation shows different defaults. Gemini 3.7 Flash defaults to medium and supports low, medium, and high. Gemini 3.6 Flash and 3.5 Flash-Lite also support minimal. Gemini 3.1 Pro Preview defaults to high.
Fix the thinking level when comparing models. Start low or minimal for latency-critical extraction; raise it for hard code, math, and multi-tool work while watching billed output and timeout rate. The 3.7 migration guide requires removing deprecated temperature, top_p, and top_k, replacing thinking_budget with thinking_level, and reviewing multi-turn and function-response rules.
Migrate from Gemini 2.5 before it becomes an incident
Google’s deprecation schedule says Gemini 2.0 Flash and Flash-Lite shut down on June 1, 2026. It lists October 16, 2026 as the earliest shutdown date for 2.5 Pro, 2.5 Flash, and 2.5 Flash-Lite. Google notes that these are earliest possible dates and may issue exact timing later.
A production migration should cover more than the model string:
- Inventory model IDs in code, environment variables, queues, and configuration services.
- Build a fixed evaluation set with normal inputs, long context, tool failures, structured output, and refusal cases.
- Update 3.x parameters, thought-signature handling, function responses, and conversation state.
- Put old and new models behind configuration, canary the change, and keep a rollback path.
- Compare quality, latency, and cost per accepted task after rollout.
For a new system, start with 3.7 Flash. Add 3.5 Flash-Lite for repetitive, verifiable work. Test 3.1 Pro Preview only for the tasks where stable Flash is demonstrably insufficient. That creates a model policy you can operate, rather than a model name you hope will stay best.
