Gemini 3 Flash

Why this oneWhat it is actually for

Gemini 3 Flash is the one to reach for when the job is high-volume and the context is still large. It carries the same 1M-token window and the same 64K output ceiling as Gemini 3.1 Pro, and it takes the same four input modalities, but text input lists at 0.50 dollars per million against Pro at 2.00 — and, the part people miss, it has no long-context surcharge. Pro doubles its rate on every token in a request once the prompt passes 200K; Flash charges the same 0.50 whether the prompt is 5K or 900K. If your workload is document extraction, transcript triage or repo-wide search over big inputs, Flash is often cheaper than Pro by considerably more than the headline 4x.

What it isArchitecture, lineage, training

Gemini 3 Flash is a closed-weight, natively multimodal reasoning model. Its model card, published December 2025, is explicit about lineage in a way that is unusually useful: Gemini 3 Flash is based on Gemini 3 Pro, and it is described as built off the Gemini 3 Pro reasoning foundation with thinking levels to control the mix of quality, cost and latency. Architecture, training dataset, training data processing and known limitations are all delegated to the Gemini 3 Pro model card rather than restated.

Following that pointer: Gemini 3 Pro is a sparse mixture-of-experts transformer with native multimodal support for text, vision and audio inputs, trained on Google TPUs over a large-scale pre-training set of publicly available web documents, text, code, images, audio and video. Sparse MoE routes each input token to a subset of experts, which is what lets a family decouple model capacity from serving cost per token — and is the most plausible reading of how a Flash tier sits under a Pro tier at a quarter the price. Note what the card does not say. It does not say Flash is a distillation of Pro, and it does not disclose any parameter count, expert count, active-parameter figure or layer count for either model. Flash was separately trained on TPUs per its own card.

The API model string is gemini-3-flash-preview. Context is 1M input and 64K output, and the knowledge cutoff is January 2025, both from the Gemini 3 developer guide's model table and the Flash model card. Google positions it for agentic workflows, everyday coding, reasoning and planning, and multimodal analysis, and claims it significantly outperforms Gemini 2.5 Pro on the December 2025 benchmark set — a comparison against the previous generation's flagship, not against Gemini 3 Pro.

At a glanceSee it

Gemini 3 Flash prices input by modality, not by prompt length, and batch halves it.

CapacityContext, output and what fits

The parametersEvery knob, and what moving it does

On Flash, thinkingLevel is the parameter that decides whether this model is cheap or not. It defaults to high, which is a frontier-model setting on an efficiency-tier price, and Flash accepts minimal, which Gemini 3.1 Pro does not. For classification, extraction and routing, minimal or low is usually correct and cuts both latency and billed tokens sharply. mediaResolution is the other one that matters, because at 0.50 per million input tokens a default-resolution image at 1120 tokens is a meaningful share of a small request. Leave the sampling knobs alone — Google now advises removing topP and topK from Gemini 3.x requests rather than tuning them.

SamplingShaping the output distribution

The sampling guidance is identical across the Gemini 3 family and it is a hard recommendation, not a hint: keep temperature at its default of 1.0. The developer guide warns that lowering it may cause looping or degraded performance, especially on complex mathematical or reasoning tasks. This surprises people migrating classification jobs from older models, where temperature 0 was standard practice for anything deterministic. The fix is to move determinism out of the sampler — set thinkingLevel to minimal or low, and constrain the output with responseSchema. Both topP and topK are accepted; the reference gives topP a 0.0 to 1.0 range and no default, and gives topK neither. A reasonable starting point for high-volume work is temperature 1.0, topP and topK unset, thinkingLevel at low, and a schema on the response. Then raise thinking only for the items that fail validation.

ReasoningThinking, effort and budgets

Yes, and unlike Gemini 3.1 Pro it can be effectively switched off. thinkingLevel accepts minimal, low, medium and high on Gemini 3 Flash, and the default is high — which means an untuned Flash call is doing frontier-depth reasoning on every request and billing you for it. The docs describe minimal as matching the no-thinking setting for most queries, with the caveat that the model may still think a little on complex coding tasks. The level is a ceiling on depth, not a fixed spend. Billing is unambiguous: response pricing is the sum of output tokens and thinking tokens, and you pay for the full thought tokens even though only a summary is ever returned to you. Summaries are opt-in. The legacy thinkingBudget integer still exists but cannot be combined with thinkingLevel in one request.

ToolsFunction calling and server tools

Function calling works the same way as on Gemini 3.1 Pro, including parallel calls for independent functions and compositional calls that chain on earlier results. Tools are declared with type: "function", name, description and parameters on the Interactions API; tool_choice takes auto, any, none and a preview validated mode. On legacy generateContent the equivalent is toolConfig.functionCallingConfig.mode with AUTO, ANY or NONE. Structured output uses response_format with mime_type and a schema drawn from a JSON Schema subset; very large or deeply nested schemas may be rejected. Server-side Google Search grounding is available and billed separately from tokens. The failure mode specific to Flash is worth naming: thinking is what makes tool selection reliable, so if you set thinkingLevel to minimal to save money on an agentic loop, expect tool-choice accuracy to drop. Image segmentation is explicitly not supported.

CostPrice, caching, batching, what drives the bill

Prices are from the Gemini Developer API pricing page, read on 2026-07-26, and agree with the model table in the Gemini 3 Developer Guide, which lists this model flat at 0.50 input and 3.00 output. The Gemini Enterprise Agent Platform (Vertex) pricing page could not be retrieved to cross-check. US dollars per million tokens. The structurally important fact is what is missing: Gemini 3 Flash has no long-context tier. Its pricing is quoted once, with no separate rate for prompts over 200K tokens, where Gemini 3.1 Pro doubles input to 4.00 and lifts output to 18.00. Audio is the modality that costs extra, at double the text rate on input.

Cached input is one tenth of standard input — a 90 percent saving computed from the published rates, not a figure Google states, since the caching page says only that cost savings are passed on automatically. Caching is implicit: on by default for Gemini 2.5 and newer, nothing to enable, with usage.total_cached_tokens confirming a hit. Batch and flex halve everything. The one that catches people is thinking: it defaults to high and thinking tokens bill at the 3.00 output rate, so an untuned Flash job can cost several times what the input price implied.

Where it runsSurfaces and availability

StrengthsWhat it is good at

• Flat input pricing across the whole 1M window — 0.50 per million whether the prompt is 5K or 900K tokens, per the Vertex pricing table. Gemini 3.1 Pro doubles above 200K.
• Full frontier-family modality set: text, images, audio and video in, per its own model card, at efficiency-tier prices.
• Accepts thinkingLevel: minimal, which Gemini 3.1 Pro does not — the cleanest way to buy Gemini 3 quality on a latency budget.
• Batch and flex tier halves both input and output, taking text input to 0.25 and output to 1.50 per million.
• Implicit caching at a 90 percent discount with no code change, tracked through usage.total_cached_tokens.

LimitsWhere it falls down

• Thinking defaults to high. Leave it alone and you pay frontier-depth thinking tokens at 3.00 per million on a model you chose because it was cheap.
• Audio input costs double text at 1.00 per million, so speech-heavy pipelines are not the bargain the headline rate implies.
• Google's own benchmark claim on the model card is against Gemini 2.5 Pro, not against Gemini 3 Pro — there is no published primary-source margin between Flash and the current Pro tier.
• Architecture is disclosed only as based on Gemini 3 Pro. No parameter count, no expert count, and no statement of whether it is a distillation.
• Knowledge cutoff of January 2025 and a -preview model string, with image segmentation explicitly unsupported.

Against its neighboursHow it compares

Against Gemini 3.1 Pro the split is cleaner than the tier names suggest. Same window, same output cap, same modalities, same knowledge cutoff, same parameter surface — the differences are price, the absence of a long-context surcharge here, and Flash's extra minimal thinking level. Reach for Pro when the task is genuinely hard; reach for Flash when it is merely long. Against Claude Haiku 4.5 and GPT-5.6 Luna, Flash's argument is the input side: a million-token window and native audio and video at efficiency-tier pricing is not something the other efficiency models offer. Its argument against DeepSeek Flash is enterprise distribution through Vertex AI rather than raw price per token.

Getting startedThe smallest call that works

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-flash-preview:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "contents": [{
      "parts": [{"text": "How does AI work?"}]
    }],
    "generationConfig": {
      "thinkingConfig": {
        "thinkingLevel": "minimal"
      }
    }
  }'

Set thinkingLevel deliberately on the very first call — the default is high and that is where an unexpected bill comes from. For extraction work add responseMimeType and responseSchema next, and set mediaResolution to low if you are sending images. Google's newer surface is POST /v1beta/interactions with model, input and snake_case generation_config.

SourcesWhere every claim above came from

• Gemini 3 Developer Guide — Gemini API
• Gemini 3 Flash Model Card, PDF
• Gemini 3 Pro Model Card, PDF
• Gemini Developer API pricing
• Agent Platform pricing, Google Cloud
• Thinking — Gemini API
• Context caching — Gemini API
• Function calling — Gemini API
• Structured output — Gemini API
• Media resolution — Gemini API
• GenerateContent API reference — GenerationConfig
• Models — Gemini API
• Could not confirm from a primary source as of 2026-07-25: any parameter count or expert count; whether Flash is a distillation of Gemini 3 Pro rather than a separately trained sibling; supervised fine-tuning eligibility on Vertex AI; and the ranges or defaults for topP, topK, presencePenalty and frequencyPenalty, which the API reference lists without them. The media-resolution token counts quoted are documented for Gemini 3 generally, not broken out per Flash.

RelatedGo deeper