Google publishes the cheapest flagship rate of the big three, then buries three billing mechanics that can double or triple the bill you budgeted from the sticker: a long-context surcharge that kicks in at 200K prompt tokens on the Pro models, an audio-input premium on the Flash models, and a context-caching fee that bills by the hour — the only per-hour line item any major LLM provider charges. This page lays out every Gemini rate as of July 12, 2026, verified against Google's official Gemini Developer API pricing page and cross-checked against the Vertex AI pricing page, and prices the traps alongside the discounts. For the cross-provider view, the same figures sit in our live LLM API price tracker , verified against the July 10, 2026 snapshot.
Disclosure: we have no affiliate or paid relationship with Google, OpenAI, Anthropic, xAI, or DeepSeek as of publication. Nobody pays us for placement in these tables. If that changes, this paragraph will say so.
Gemini API pricing at a glance
Gemini API pricing is pay-as-you-go, metered per million tokens. As of July 12, 2026, rates run from $0.10 input on Gemini 2.5 Flash-Lite to $12.00 output on Gemini 3.1 Pro — and to $18.00 once a prompt passes 200K tokens. Batch jobs get a flat 50% off, and caching adds an hourly storage fee.
| Model | Input /1M (≤200K prompt) | Output /1M (≤200K) | Input /1M (>200K) | Output /1M (>200K) | Context |
|---|---|---|---|---|---|
| Gemini 3.1 Pro (Preview) | $2.00 | $12.00 | $4.00 | $18.00 | 1.05M |
| Gemini 2.5 Pro | $1.25 | $10.00 | $2.50 | $15.00 | 1.05M |
| Gemini 2.5 Flash | $0.30 text/image/video, $1.00 audio | $2.50 | no surcharge | no surcharge | 1M |
| Gemini 2.5 Flash-Lite | $0.10 text/image/video, $0.30 audio | $0.40 | no surcharge | no surcharge | 1M |
Source: Google's Gemini Developer API pricing page, verified July 12, 2026. The ≤200K rates match our canonical cross-provider table (snapshot July 10, 2026) exactly; the >200K columns are the nuance that table's at-a-glance grid does not carry. Rates move — treat any figure older than the model's last release week as suspect.
Three things to read off this grid before you budget anything. First, the two Pro models are two different products above and below 200K prompt tokens; the right column is not a rounding error, it is input doubling. Second, the Flash models are flat across the full 1M window — the only rates in the table you can multiply naively. Third, audio input is a separate, higher meter on both Flash models: $1.00 per 1M on Flash against its $0.30 text rate, $0.30 on Flash-Lite against $0.10. A transcription-adjacent workload priced off the text sticker comes in roughly 3x over budget.
The tier ladder: which Gemini model for which job
Google's lineup slots into the same four tiers as everyone else's, and at each rung the honest question is whether a rival does the same job for less. Here is the ladder with the swap math, all rates per 1M tokens as of the July 2026 verification.
Flagship — Gemini 3.1 Pro, $2.00/$12.00. This is the aggressive price in the July 2026 market. GPT-5.6 Sol lists $5.00/$30.00 and Claude Opus 4.8 lists $5.00/$25.00 (both per our July 10, 2026 tracker snapshot), which makes Gemini 's flagship 2.5x cheaper than OpenAI's on both sides of the meter. The only flagship that undercuts it on output is Grok 4.5 at $2.00/$6.00 — with the trade-off of a 500K context window, half of Gemini 's 1.05M. Two caveats before you standardize on it: 3.1 Pro is still labeled Preview, meaning Google can reprice it at general availability, and it is the one Gemini model with no free tier for prototyping.
Workhorse — Gemini 2.5 Pro, $1.25/$10.00. A dead tie with GPT-5.1 at $1.25/$10.00; pick by ecosystem and context ( Gemini carries 1.05M against GPT-5.1's 400K). Claude Sonnet 5 sits at $2.00/$10.00 on introductory pricing. If the job tolerates a non-US provider, DeepSeek V4 Pro undercuts the whole tier at $0.44/$0.87 per the same snapshot — roughly 11x cheaper per output token than 2.5 Pro, with the jurisdiction and data-terms questions that discount buys.
Budget — Gemini 2.5 Flash, $0.30/$2.50. GPT-5-mini is slightly cheaper at $0.25/$2.00, Claude Haiku 4.5 markedly dearer at $1.00/$5.00. Flash's case is the window: 1M tokens of flat-rate context where GPT-5-mini caps at 400K. For summarize-this-pile workloads, that is the whole decision.
Floor — Gemini 2.5 Flash-Lite, $0.10/$0.40. GPT-5-nano beats it on input, $0.05 against $0.10, and ties it on output at $0.40. Our canonical table lists DeepSeek V4 Flash at $0.09/$0.18 (July 10, 2026 snapshot; DeepSeek 's rates have been the most volatile line in the tracker, so re-check before committing volume — rates move). What Flash-Lite owns at this tier is context per dollar: a 1M-token window at floor pricing, with no surcharge anywhere in it. Nano gives you 400K.
The pattern across the ladder: Gemini is priced to win the flagship tier outright, ties the workhorse tier, and competes at the floor on context rather than rate. Nothing in the lineup is the single cheapest option for every job — the cheapest-LLM breakdown runs that split in full.
The 200K surcharge: where Pro prompts double in price
Past 200K prompt tokens, Gemini 3.1 Pro input goes from $2.00 to $4.00 per 1M and output from $12.00 to $18.00; Gemini 2.5 Pro goes from $1.25 to $2.50 input and $10.00 to $15.00 output. Google's pricing page states this plainly and almost every third-party comparison ignores it, because comparison tables — including the at-a-glance grid on our own tracker — quote the ≤200K rate. Those quotes are correct; they are just not the rate a long-context job pays. Note the asymmetry with the Flash models, where the full 1M window really is folded into the flat rate.
The math bites exactly where Gemini 's 1M window is the selling point. Feed 3.1 Pro an 800K-token codebase and the input side of that single call costs 0.8 × $4.00 = $3.20 — double what the sticker rate suggests — and every output token of the answer bills at $18.00 per million, not $12.00. A daily job that re-sends that codebase 50 times runs about $160 per day on input alone before caching. The same prompt shape on 2.5 Flash bills flat: 0.8 × $0.30 = $0.24 per call. If the task survives the quality drop, the Flash swap on long-context work is a 13x cut that no discount program matches.
The practical rule: below 200K tokens, price the Pros off the sticker; above it, double the input rate in your forecast and add 50% to output. And remember the threshold is per prompt, not per month — a workload that averages 150K but spikes to 300K pays the surcharge on the spikes.
Batch API: a flat 50% off everything, surcharge included
Google's Batch API halves every rate in the table for asynchronous jobs — the same roughly-50% deal OpenAI and Anthropic document, and Gemini 's version applies uniformly. Per Google's pricing page, verified July 12, 2026: Gemini 3.1 Pro drops to $1.00/$6.00 (≤200K prompts) and $2.00/$9.00 above 200K; 2.5 Pro to $0.625/$5.00; Flash to $0.15/$1.25; Flash-Lite to $0.05/$0.20.
Two readings worth pausing on. Batch Flash-Lite at $0.05/$0.20 matches GPT-5-nano's real-time input rate while halving its output rate — for overnight classification or extraction backlogs, that is currently one of the cheapest metered rates from any US provider. And the long-context surcharge survives into batch on the Pro models: a >200K batch prompt on 3.1 Pro still pays double input at $2.00. The discount is a multiplier on whatever tier your prompt lands in, not an escape from the tiering.
The condition is the standard one: submit asynchronously, target turnaround within 24 hours, no interactive use. Real-time chatbots get nothing from it. Document pipelines, evals, embedding-adjacent batch scoring, and refresh jobs get half off for tolerating a delay they mostly tolerate anyway.
Context caching: 90% off reads, billed by the hour to exist
Cached input tokens cost roughly 10% of the standard input rate: $0.20 per 1M on 3.1 Pro (≤200K; $0.40 above), $0.125 on 2.5 Pro, $0.03 on Flash, $0.01 on Flash-Lite — all per Google's pricing page, July 12, 2026, and consistent with the ~10% cache-read ratio our tracker shows across the US big three. So far, standard.
The Gemini -specific trap is storage. Google bills for keeping a cache alive: $4.50 per 1M tokens per hour on the Pro models, $1.00 per 1M tokens per hour on Flash and Flash-Lite. Nobody else in our cross-provider table meters caching by the clock — Anthropic charges a one-time 1.25x write premium, OpenAI charges nothing to hold a cache. On Gemini , a cached 500K-token corpus on 2.5 Pro costs 0.5 × $4.50 = $2.25 per hour, about $54 per day, before you read a single token from it.
Whether that is a trap or a bargain is pure arithmetic on your query rate. Each read of that 500K cache saves you 0.5 × ($1.25 − $0.125) ≈ $0.56 against re-sending the corpus at full price. At four reads an hour the storage fee is covered; at forty, caching is cutting your input bill by roughly 85%. At two reads a day, you are paying $54 daily to save about a dollar — delete the cache and resend. The failure mode to audit for: caches created by a pipeline and never explicitly expired, billing $4.50/1M/hour for content nothing reads. Set TTLs at creation time; do not rely on remembering.
Free tier, paid tiers, and AI Studio vs Vertex AI
The Gemini Developer API has a genuinely free tier on Gemini 2.5 Pro, 2.5 Flash, and 2.5 Flash-Lite — free input and output tokens through an AI Studio API key. Gemini 3.1 Pro has no free tier; prototyping the flagship costs money from the first call. We are deliberately not printing free-tier request-per-minute or per-day numbers: Google's rate-limits documentation no longer publishes a per-model table and points instead to the live dashboard in AI Studio, and the third-party trackers that do print numbers contradict each other. The honest statement is that limits are per-project, visible live in AI Studio, and revised without notice. Linking a billing account (Tier 1) removes most of the daily caps. One line from Google's terms worth the whole section: free-tier prompts may be used to improve Google's products; paid-tier prompts are not. If the input is client data, the free tier is not a discount, it is a data decision.
The other fork is where to call the API from, and the pricing answer is simple: per-token rates are identical on the Gemini Developer API (AI Studio) and on Vertex AI — we verified the same $2.00/$12.00, $1.25/$10.00, $0.30/$2.50, and $0.10/$0.40, plus the same batch and caching rates, on both Google pricing pages on July 12, 2026. What actually differs: the free tier exists only on the AI Studio side; Vertex adds the enterprise layer — provisioned throughput, regional endpoints and data residency, CMEK and VPC Service Controls; and billing routes through your Google Cloud account on Vertex versus the API-key billing account on AI Studio. Pick by governance requirements, not by rate card, because the rate cards are the same.
Where this pricing breakdown falls short
Honest limits of this page. Gemini 3.1 Pro is still marked Preview on Google's pricing page, and preview pricing has been repriced at general availability before — the $2.00/$12.00 rate is verified for July 12, 2026, not guaranteed for August. We could not verify the free tier's exact request limits from any official source, so this page refuses to print numbers that third-party trackers disagree on; you will have to read your own AI Studio dashboard. This is a pricing page, not a quality page — nothing here tells you whether 3.1 Pro's output on your task justifies choosing it over a cheaper rival, and we have run no benchmarks of our own. The rates cover text-model token metering only: grounding with Google Search, Imagen and Veo generation, and embeddings bill on separate meters this page does not cover. And every dollar figure above is a dated snapshot of the fastest-moving price list in software — the live tracker is re-pulled on each snapshot and is the page to trust when this one and it disagree by a release cycle.
All guides in this topic
- LLM API Pricing: Live Cross-Provider Cost Per Token — the canonical live tracker every figure on this page reconciles against.
- OpenAI API Pricing: Live Per-Token Rates and What They Cost on Real Work — the same tier-ladder treatment for the GPT lineup.
- Claude API Pricing: Live Cost Per Token for Opus, Sonnet, and Haiku — Anthropic's rates, including the 1.25x cache-write premium Gemini 's hourly fee replaces.
- Cheapest LLM API — the floor-tier shootout: GPT-5-nano vs Flash-Lite vs DeepSeek , split by token mix.
- Gemini vs ChatGPT — the consumer-plan and capability comparison behind these API rates.
- Claude vs ChatGPT — the task-by-task verdict on the other flagship pair.
- Grok vs ChatGPT — where the cheapest flagship output rate ($6.00/1M) fits.
- DeepSeek vs ChatGPT — the discount that is partly a jurisdiction discount.
- Best LLM for Coding — price joined to coding benchmarks, where Gemini 's rates meet its scores.