Gemini 3 Flash Preview: Pro-Grade Reasoning at a Flash Price

The usual assumption about a "Flash" model is that it is the cheap, fast, slightly dim sibling you route easy traffic to. Gemini 3 Flash broke that pattern when it arrived: Google pitched it as Gemini 3 Pro-grade reasoning at Flash latency and cost, and its own benchmark table had it beating Gemini 3 Pro on at least one coding evaluation.

Ten months on, the model is still served under a -preview identifier, and Google has shipped several newer Flash generations since. That makes it a useful case study in two things at once: what a well-built mid-tier reasoning model looks like, and how to run production traffic on something that was never formally declared finished.

What Google shipped in December 2025

Gemini 3 Flash was released on 17 December 2025 as a preview in the Gemini API, Google AI Studio, Vertex AI and Gemini Enterprise. The API model id is gemini-3-flash-preview. On the consumer side it replaced Gemini 2.5 Flash as the default model in the Gemini app and became the default for AI Mode in Search.

Google's Vertex AI documentation lists a 1,048,576-token input window and 65,536 tokens of maximum output. Input covers text, images, audio, video and PDF; output is text only. The knowledge cutoff is January 2025, the same as Gemini 3 Pro and the 2.5 series. The platform features you would expect from a current Gemini are there: function calling, structured output, context caching, Google Search grounding, code execution and batch inference.

The headline numbers from Google's launch post:

• GPQA Diamond — 90.4%
• Humanity's Last Exam — 33.7% without tools
• MMMU Pro — 81.2%, which Google describes as comparable to Gemini 3 Pro
• SWE-bench Verified — 78%, which Google says beats both the 2.5 series and Gemini 3 Pro

Google also cites Artificial Analysis benchmarking showing it outperforming Gemini 2.5 Pro while running about three times faster. These are vendor-selected results on vendor-chosen settings, so treat them as a reason to test rather than as a conclusion.

Four thinking levels are the real price control

Gemini 3 replaced the older numeric thinking_budget with a thinking_level parameter. Gemini 3 Flash accepts four values — minimal, low, medium and high — where Gemini 3 Pro launched with only low and high. In the preview the default is high, and at that level the model adjusts how much it thinks to the difficulty of the request.

This matters more than the list price, because reasoning tokens are billed as output. A classification call that thinks at high can cost several times what the same call costs at minimal, for no gain in accuracy. The token cost of reasoning models walks through why hidden output dominates bills on models like this.

A practical default: set the level explicitly on every call rather than inheriting high. Start extraction, routing and short-answer work at minimal or low, and keep high for multi-step coding or analysis where you have measured that it helps. If your client speaks the OpenAI protocol, check how it maps reasoning_effort onto the Gemini setting — gateways differ.

The price went up, and why that may not matter

Google priced the model at $0.50 per million input tokens and $3 per million output, with audio input at $1 per million. Gemini 2.5 Flash had been $0.30 and $2.50, so on paper this was a price increase for the Flash tier. OpenRouter listed it at the same $0.50 and $3.00 at the time of writing, with cached input reads at $0.05 per million.

Two things offset that. Google claims the model uses around 30% fewer tokens than Gemini 2.5 Pro on typical traffic, and unlike Gemini 3 Pro it has no higher price tier above 200K tokens of context. If you are feeding long documents, the flat rate across the whole 1M window is the more important number. Input vs output token pricing covers how to work out which side of the bill your workload actually lands on.

What preview means when successors already exist

A preview model carries fewer guarantees than a stable one: behaviour can change, quotas can be tighter, and the retirement notice period may be shorter. Google's Gemini API deprecations page currently lists no shutdown date for gemini-3-flash-preview, but it names gemini-3.6-flash as the recommended replacement, and Google's own migration notes describe Gemini 3.8 Flash as the latest Flash model.

That is a clear signal of direction. Nothing forces you off today, but new work should not be built on the assumption that this identifier will be around indefinitely. The guides to Gemini 3.6 Flash and Gemini 3.8 Flash cover what the successors change.

If you do keep it in production, put the model id in configuration rather than code, keep a pinned evaluation set, and have the replacement already tested before a retirement date appears. Pinning model versions explains why aliases and previews are where silent regressions come from.

Variants: the batch endpoint

The catalogue lists one variant besides the main endpoint: google/gemini-3-flash-preview:batch. It is the same model behind an asynchronous, discounted endpoint — OpenRouter listed it at $0.25 input and $1.50 output per million at the time of writing, half the interactive rate. You give up immediate responses in exchange; results arrive when the batch completes.

For nightly document processing, evaluation runs and backfills, that trade is usually worth taking. Batch API savings covers when the latency cost is real and when it is imaginary.

Where it still fits

The model's shape suits a few jobs well:

• Long multimodal inputs — a 1M window with flat pricing across it, accepting video, audio and PDF in the same request.
• Mid-difficulty coding assistance — the SWE-bench Verified claim puts it in a range where it is worth testing against more expensive models on your own tasks.
• Mixed traffic behind one id — the thinking levels let one endpoint serve both trivial and hard requests, provided you set the level per call.

It is a weaker choice for anything that needs a long support horizon, and Google notes that pixel-level image segmentation is not supported in Gemini 3 Flash; it points to Gemini 2.5 Flash with thinking off for that.

A decision rule

If you are already running it and it passes your evaluations, keep it, but schedule the comparison against its named successor now rather than when a deprecation email arrives. If you are starting fresh, test the newer Flash first and only fall back to this one if it wins on your own tasks.

Either way, compare at matched thinking levels and record cost per completed task, not price per token. A model that thinks less to reach the same answer can be cheaper even at a higher rate.