Gemini API Pricing Explained: When the Cheapest Model Actually Wins

Google’s Gemini API pricing page now has more than one answer to “What does this model cost?” A request can run at Standard, Priority, Flex or Batch rates; part of the input may be cached; output includes thinking tokens; Google Search grounding can add a separate fee; and some 2026 prices are promotional rates that change on January 1, 2027.

That complexity creates an opportunity. For work that does not need the strongest model or an immediate answer, a cheaper Flash-Lite model plus Batch or Flex can cut the token bill by far more than prompt trimming alone. But the cheapest model wins only after it passes the same task-level evaluation as the expensive one. A failed classification retried twice is not cheap.

This guide prices three production-shaped workloads on current Gemini models, shows when caching or Batch matters, and gives a decision rule that includes quality instead of treating every token as interchangeable.

Four metersWhat the Gemini API actually bills

Google’s official billing guide lists four token-based inputs to the bill:

Google says failed requests returning a 400 or 500 response are not billed for their tokens, although they still count against quota. GetTokens requests are not billed and do not count against inference quota. Those are useful operational details, but they do not replace request logging: the response usage fields are what let you reconcile estimated and actual tokens.

Our AI Token Cost Calculator prices the text you actually send rather than assuming every request is one million tokens. The broader four-number AI bill guide adds requests per user and users per month—the multipliers that turn a tiny request into a budget.

Current rate cardThree models and four delivery modes

The focused comparison below uses one current general Flash model and two cheaper Flash-Lite models. It avoids voice, image generation and specialist media models because their units differ.

Three observations matter immediately.

1. 01Output is the expensive side. On Gemini 3.8 Flash, one million output tokens cost five times one million fresh input tokens at the promotional Standard rate.
2. 02Batch and Flex halve token rates. Google describes Batch as asynchronous with a target of completing within 24 hours; Flex is best-effort synchronous processing for work that can tolerate variable latency and availability.
3. 03Older can be cheaper. The 3.1 Flash-Lite row is below 3.5 Flash-Lite on both input and output. Model numbering is not a price ladder and price is not a quality score.

Priority goes the other direction: it buys non-sheddable, user-facing reliability at a premium. The official optimization guide frames these modes by latency and reliability, not only cost. Use Standard as the default, Batch for offline volume, Flex for non-urgent sequential work and Priority only where the service requirement justifies it.

Three workloadsWhat the rate card becomes in production

All three workloads below use text input and output only. They exclude tax, Google Search or Maps grounding, cache storage and retries. That makes them a clean comparison of model and delivery mode, not a full invoice forecast.

The classification job is where “the cheapest model wins” has the best chance to be true. The output is constrained, the task can be labeled objectively and offline batches are acceptable. A representative test set can decide whether 3.1 Flash-Lite clears the quality threshold before one million requests are routed to it.

The support job is different. It can still be evaluated—policy accuracy, required fields, tone, escalation—but latency may rule out Batch and a wrong answer may cost far more than the token saving. The model choice belongs behind a router or evaluation gate, not in a rate-card spreadsheet.

The long-document workload is input-heavy. It shows why cache design can matter as much as model choice when the same source material is queried repeatedly.

CachingA discount with a storage clock

Google supports implicit caching on Gemini 2.5 and newer models and says it automatically passes on savings when a request hits an eligible cache. Explicit caching lets an application create a cache object with a time-to-live; the default TTL is one hour when none is provided. The bill then includes both reduced cache-read tokens and token-hours of storage.

For the document workload above—500 million input tokens and 10 million output tokens—suppose 80% of input hits Gemini 3.8 Flash’s 2026 cached-input price:

The saving is real only when the prefix is reused enough times before expiry. A huge one-hour cache used once can cost more than simply resending it. Log cached token count, cache-hit rate and storage hours; do not assume enabling a feature guarantees a saving.

This principle is provider-independent. The OpenAI API pricing guide and Claude API pricing guide show different cache mechanics but the same operational truth: stable repeated prefixes are valuable, long unique conversations are not.

The quality gateCheap tokens are not cheap tasks

Use a labeled evaluation set drawn from the real workload. Then price a successful task, not merely an attempt.

Our cross-provider ChatGPT vs Claude vs Gemini cost-per-task comparison shows that model tier often changes cost more than provider. This article adds the missing condition: a cheaper tier is a saving only when it produces an accepted task often enough.

Free tierFree is a testing environment, not a zero-cost architecture

Google’s pricing page says the free tier offers limited access and free input/output tokens; paid accounts receive higher rate limits, paid features and content is not used to improve Google’s products. The same page states free-tier content may be used to improve products. Google’s billing FAQ also says that, starting March 2026, the $300 Google Cloud free-trial credit does not apply to Gemini API usage.

That makes the free tier useful for prototyping and evaluation, but it should not be treated as the production unit economics. Record the paid rate that applies after the prototype succeeds, and review the data-handling terms before sending customer content.

Choose the modeA practical decision table

Where these numbers come from

Rates and service behavior come from Google’s official Gemini API pricing, billing, Batch API, context caching and optimization pages, read September 23, 2026.

drafts/gemini-api-pricing-numbers.mjs multiplies each workload’s input and output tokens by the published per-million rates. The 80% cache case charges 20% at fresh-input rates and 80% at cached-input rates, then adds output; it deliberately excludes storage because storage depends on token-hours. All examples exclude taxes, grounding/tool charges, retries and human review, which biases them downward. The quality-gate section explains how to add those costs back.

FAQFrequently asked questions

Is the Gemini API free? Google offers a limited free tier with free input and output tokens for eligible models. Production workloads that need higher limits and paid features use paid billing; free-tier and paid data-handling terms differ.

What is the cheapest Gemini model in this comparison? Gemini 3.1 Flash-Lite has the lowest listed Standard and Batch/Flex text rates among the three models compared here. It is cheapest only when it passes the application’s quality and latency requirements.

Does Gemini Batch API really cost half? Google’s official documentation says Batch is priced at 50% of the equivalent Standard interactive API and is designed for asynchronous jobs with up to a 24-hour turnaround target.

Are thinking tokens billed? Yes. The pricing tables describe output prices as including thinking tokens, so a short visible answer can still use more billable output than its displayed text suggests.

When does context caching save money? When a large stable prefix is reused often enough before the cache expires. Compare reduced cache-read cost plus token-hour storage with the cost of sending fresh input each time.

Why can the invoice be higher than the token estimate? Search or Maps grounding, media tokens, cache storage, Priority processing, retries, taxes and a changed model/rate can sit outside a simple input-plus-output estimate.

Next stepPrice the request you actually send

Paste a real prompt into the AI Token Cost Calculator, choose the Gemini models you are evaluating and record input, output and cached shares. Then put the winning candidate into the AI Cost Calculator with monthly volume. The decision is complete only when the cheapest row also passes the task.