Gemini 3.1 Flash-Lite Guide: Thinking Levels at High Volume

The interesting thing about Gemini 3.1 Flash-Lite is not that it is cheap. It is that it is cheaper than the model many teams were using as their default mid-tier. Google lists it below Gemini 2.5 Flash on both input and output price, and claims it answers faster. If that holds on your workload, it is not a budget option so much as a replacement for a whole tier.

This guide covers the release history, what Google claims, how the thinking controls behave, and where it belongs in a routing setup.

From preview to stable in nine weeks

Google launched gemini-3.1-flash-lite-preview on 3 March 2026, describing it as the first Flash-Lite model in the Gemini 3 series. The generally available gemini-3.1-flash-lite followed on 7 May 2026. The preview identifier was deprecated within days and shut down on 25 May 2026.

That short overlap is the operational lesson. Anything still pointing at the preview string stopped working two and a half weeks after the stable model appeared. Some aggregators still list a preview entry, but Google no longer serves it. Target the stable identifier, and treat preview names in config files as debt with a short fuse.

The specification is familiar: a 1,048,576-token input limit, 65,536 output tokens, text, image, video, audio and PDF input, and text output. It supports function calling, structured outputs, code execution, context caching, URL context and grounding with Google Search, and is available on the Batch API. It does not do computer use or the Live API.

What Google claims

At the preview launch Google reported an Arena Elo of 1432, 86.9% on GPQA Diamond and 76.8% on MMMU Pro. On speed, it cited Artificial Analysis measurements showing 2.5 times faster time to first answer token and 45% higher output speed than Gemini 2.5 Flash, at similar or better quality.

Those are vendor-selected figures from the preview build, published before the stable release. They are a reasonable signal that the model is competitive for its tier. They are not a reason to skip your own measurement, particularly the latency figure, which depends heavily on region, load and prompt length. Best model for low latency covers how to measure time to first token properly.

Google pitches it for two kinds of work: high-volume, cost-sensitive jobs like translation and content moderation, and heavier tasks like generating UI, following complex instructions and sorting large image sets.

Four thinking levels, and why minimal is not zero

Gemini 3 models replace the 2.5-era token budget with a thinking_level setting. Flash-Lite supports all four levels: minimal, low, medium and high. Gemini 3.1 Pro, for comparison, does not offer minimal.

Three details from Google's documentation are easy to miss:

• Levels are relative allowances. Google describes them as guidance for how much to think, not strict token caps. Measure actual reasoning tokens per level on your prompts.
• Minimal does not guarantee no thinking. If you need fully deterministic latency, test it rather than assuming.
• Do not mix controls. Sending both thinking_level and the legacy thinking_budget returns an error, which is the most common breakage when porting 2.5 code.

Set the level explicitly on every request rather than relying on a default. Output prices include thinking tokens, so the level you pick is a direct line on the bill. Google also recommends leaving temperature at its default of 1.0 on Gemini 3 models.

The price, placed properly

Google lists 3.1 Flash-Lite at $0.25 per million input tokens and $1.50 per million output tokens on its paid tier, with audio input at $0.50. The Batch API halves that, and aggregators expose it as google/gemini-3.1-flash-lite:batch, the same model on an asynchronous endpoint.

Placed against the 2.5 family, it lands in an unusual spot:

• Against 2.5 Flash-Lite ($0.10 / $0.40) it costs two and a half times as much on input and nearly four times on output. The 2.5 Flash-Lite guide covers when the older model is still enough.
• Against 2.5 Flash ($0.30 / $2.50) it is cheaper on both, and claimed to be faster.

So the real question for most teams is not whether to upgrade from 2.5 Flash-Lite. It is whether 3.1 Flash-Lite can take over work currently running on 2.5 Flash, which is being retired on Google Cloud and limited on the Gemini API.

Where it fits in a router

A cheap, fast model with adjustable reasoning is the natural first tier of a router. Send every request to Flash-Lite at a low thinking level; escalate to a higher level, then to a larger model, only when a check fails — invalid JSON, a failed test, a low-confidence classification. Model routing and fallbacks covers the mechanics.

The thinking levels make this unusually cheap to build, because the first escalation step is a parameter change on the same model rather than a second integration.

One caveat on longevity: Google has already shipped a newer Flash-Lite generation. The Gemini 3.5 Flash-Lite guide covers it. 3.1 Flash-Lite is stable and supported, but compare both before committing a new pipeline.

A quick evaluation plan

1. Take two hundred real inputs from the workload you want to move, with known good outputs.
2. Run them at minimal, low and medium, and record accuracy, reasoning tokens and latency at each level.
3. Run the same set on your current model.
4. Pick the lowest level that matches your current accuracy, and price it per correct answer.