Gemini 2.5 Flash: Full Spec Sheet (June 2026)

Gemini 2.5 Flash is the smaller, faster sibling of Gemini 2.5 Pro built on the same architecture and training pipeline. Same 1M context window, same native multimodal capability, same built-in tools (code execution, Search grounding), same thinking mode. Same feature surface — different model size, dramatically different price.

Flash at $0.30/$2.50 per 1M is approximately 4× cheaper than Gemini 2.5 Pro (≤200K tier) on input and 4× cheaper on output. The trade-off is quality on hard reasoning tasks: Flash is calibrated for production volume — chat, classification, extraction, structured-data pipelines, multimodal Q&A on shorter inputs. For complex code synthesis, multi-step reasoning, or long-form analysis where Pro's extra quality pencils, escalate to Pro.

Critically, Flash does NOT have Pro's >200K input tier bump. The flat $0.30/M input rate holds from 1 token to 1M tokens. For workloads that occasionally cross 200K (full PDF processing, long video summarization), Flash is dramatically cheaper than Pro's >200K tier ($0.30/M vs $2.50/M).

Standard text-only call: `cost = (input_tokens / 1M) × $0.30 + (output_tokens / 1M) × $2.50`. The representative 1,000-in / 500-out call: `0.001 × $0.30 + 0.0005 × $2.50 = $0.0003 + $0.00125 = $0.00155`. About 0.155¢ per call — cheapest of any model in this guide.

Audio input gotcha: pass audio data and the input portion of that audio bills at $1/M, not $0.30/M. A 10-minute audio recording is ~19,200 tokens at the audio rate: `0.0192 × $1 = $0.0192` per audio input. Roughly 3× what text-only billing would suggest. Always check whether your input mix includes audio when modeling per-call cost.

Thinking mode toggles output pricing. With thinking enabled (default for many use cases): output bills at $3.50/M instead of $2.50/M, plus the thinking tokens themselves bill at the same $3.50/M. Disable thinking entirely with `thinking_config: {thinking_budget: 0}` to lock output at $2.50/M for fastest, cheapest responses on mechanical tasks.

Caching: explicit, via the same `cachedContents` API endpoint as Gemini 2.5 Pro. Pre-create a cached block, reference by ID. Cached portion reads at $0.075/M (text/image/video) or $0.25/M (audio). Largest cost lever on workloads with stable prefixes.

Flash supports the same native multimodal as Pro: pass text, images, audio, video, or PDF in the `contents` array in any combination. The token accounting matches Pro: video bills at ~258 tokens per second; audio at ~32 tokens per second; images at fixed token cost per resolution tier.

Practical: a 5-minute video bills `300 × 258 = 77,400` input tokens. On Flash text/image/video rate, that's `0.0774 × $0.30 = $0.0232` for the video input. A 30-minute meeting video is ~465K input tokens = $0.1395. Tractable cost for analytical workloads that would be expensive on Pro.

Pure audio is more expensive per token (the $1/M rate), but audio token density is lower (32 tokens/sec vs video's 258 tokens/sec). A 30-minute audio-only meeting is ~57,600 audio tokens = $0.0576. Still cheap, just be explicit about the audio rate.

For files over 20MB, use the Google File API: upload once, reference across multiple analytical calls. Cuts re-upload time and bandwidth, doesn't change per-token billing.

Flash supports the full Gemini function-calling surface: declare functions in `tools`, model picks one (or several in parallel) and returns the arguments. Parallel function calling is on by default.

Structured outputs via `responseSchema` (JSON Schema subset). Pass in `generationConfig`, output is guaranteed to validate. Identical mechanism to Gemini 2.5 Pro.

Built-in tools: code execution (sandboxed Python), Google Search grounding (cites results with attribution), URL context (fetches and reads URLs). Same as Pro — same orchestration savings. For agentic workflows where a tool's main cost is calling out to your own infrastructure, Flash + built-in tools often replaces a custom orchestrator that would have cost more to build than the LLM bill saves.

Thinking mode is configurable on Flash via `thinking_config: {thinking_budget: N}`. Default behavior varies by deployment context but tends to enable a small thinking budget by default.

On Flash specifically, thinking mode is often a net loss. Flash is calibrated for fast, cheap responses on production-volume tasks (classification, extraction, chat). Adding thinking adds output cost (+$1/M on output) AND adds latency AND adds thinking tokens (billed at $3.50/M). For most Flash workloads, explicit `thinking_budget: 0` is the right call — it locks the cheaper output rate and removes a hidden cost lever.

When to enable thinking on Flash: when you've measured a specific quality gap on a known task and a small budget (500-1,500 tokens) closes it. For anything harder, escalate to Gemini 2.5 Pro rather than turning up thinking on Flash.

**Pick Gemini 2.5 Flash** when you need 1M context at the lowest price, when native multimodal (especially video) matters for the task, when built-in tools replace custom orchestration, or when you're in the Google ecosystem and Vertex AI billing simplifies procurement. The cheapest frontier model with this feature surface.

**Pick Gemini 2.5 Pro** over Flash when: the task is hard reasoning / code synthesis / complex multi-step planning where Flash's quality isn't enough. Pay the 4× premium only when the quality gap is measurable on your eval.

**Pick gpt-5-mini** over Flash when: structured outputs with the most-mature JSON Schema enforcement matter, you're in the OpenAI tooling ecosystem, or you don't need video/audio input and don't need 1M context. Comparable price ($0.25/$2 vs $0.30/$2.50). Pick by ecosystem fit.

**Pick Claude Sonnet 4.6** over Flash when: long-form writing voice, refusal-calibration discipline, or extended thinking matters more than 1M context. Sonnet is 10× more expensive than Flash on input ($3 vs $0.30) — the premium is justified only when Anthropic's quality differential is the bottleneck.

Every number on this page was verified against Google's live documentation on 2026-06-20. Sources: ai.google.dev/gemini-api/docs/models/gemini for feature support; ai.google.dev/pricing for input/output/cached prices and the audio-input tier; cloud.google.com/vertex-ai/generative-ai/pricing for Vertex AI pricing (currently identical to AI Studio direct).

Flash pricing has moved twice since launch (both downward). Google publishes price changes via the Vertex AI release notes; subscribe if your monthly Flash bill exceeds $500.

Methodology: when a number could not be cross-confirmed against an official Google page on the verification date, it was omitted from this card rather than guessed.