Gemini API Pricing: Token Rates, Caching and Batch Costs

Gemini API pricing depends on the exact model, request mode, input modality, cached tokens and output including thinking. Google's direct API rates are not necessarily a gateway's rates. This guide gives a selected, dated price table and worked calculations for an interactive request, Batch processing and cache reuse. Prices were checked against Google's official pricing page on October 4, 2026; the examples are calculations, not invoices or new model tests.

October 4 update: replaced the mixed June price table with dated Google-direct rates, separated expiring discounts and long-context prices, and added explicit Batch and cache-storage calculations. Key creation and request setup now live in the Gemini API key guide.

Selected Google-direct Gemini API token rates

The table covers selected text-output models in Standard mode, in USD per million tokens. It is not the full catalog and does not quote image generation, video, Live audio, tool fees, Flex or Priority inference. The source is Google's official Gemini API pricing page, not an aggregator or our gateway catalog.

Two boundaries to retain with any copied number: Google lists Gemini 3.8 Flash Standard rates rising to $1.50 input, $0.15 cached input and $7.50 output on January 1, 2027. For Gemini 3.1 Flash-Lite audio input, the current input/cache rates are $0.50/$0.05 rather than the text/image/video rates above. A model family name without the modality and date is not a complete quote.

The Pro threshold also changes the output rate. Do not price a prompt over 200K at $4 input while retaining the short-prompt $12 output rate. Use the exact model ID and the table for your request mode.

Calculate an interactive Gemini request

For a request without cached input or extra tools:

cost_usd = (input_tokens * input_rate
          + billed_output_tokens * output_rate) / 1000000

Illustrative token counts: 10,000 input tokens and 2,000 total billed output tokens on Gemini 3.8 Flash at the current Standard discount cost (10,000 * 0.75 + 2,000 * 3.75) / 1,000,000 = $0.015. One thousand requests with exactly that profile would cost $15 before other charges. This is arithmetic, not a measured production workload.

If the visible answer has 2,000 tokens but the model also uses billable thinking tokens, the $0.015 example does not cover the full output. Inspect the documented usage breakdown, and include retries, abandoned requests and multi-call agent tasks when comparing monthly spending.

On the Pro Preview short-prompt tier, 100,000 input tokens and 2,000 billed output tokens cost $0.224. With 250,000 input tokens, the long-prompt rates make the corresponding calculation 250,000 * $4 / 1M + 2,000 * $18 / 1M = $1.036. The threshold is a billing condition, not merely a larger context allowance.

Batch costs less when asynchronous work fits

Google's Batch API processes queued requests asynchronously at 50% of the equivalent Standard cost. Its documented turnaround target is 24 hours, not a guaranteed live-response time. Use it for work that can wait, and check support for the selected model and feature.

The Gemini 3.8 Flash Batch tab currently lists $0.375 input and $1.875 output per million tokens through December 31, 2026. The same hypothetical 10K-input/2K-output request profile costs $0.0075, or $7.50 for 1,000 identical requests, before extras. From January 1, the published Batch rates change to $0.75/$3.75.

Sending ordinary interactive requests in a loop is not the Batch API. Check completed job results and token usage before reporting savings. A 50% rate discount does not prove a 50% lower application bill if request size or retry behavior changes.

Caching has a read price and a storage price

For explicit context caching, include both the cached-input charge and storage duration. Google's Standard Gemini 3.8 Flash table currently lists cached input at $0.075 per million tokens and storage at $0.50 per million tokens per hour, through December 31. Its published January rates are $0.15 and $1.00 respectively.

Worked reuse scenario: retain a 100K-token context for one hour and read it ten times. Assume every read is billable at the stated cache rate. Storage costs $0.05; the ten reads total one million cached tokens and cost $0.075. Together that is $0.125, excluding cache-creation input and any new input, output or tool use.

Sending the same 100K context ten times without a cache would cost $0.75 in Standard input at the current Flash rate, before output. If creating the cache also bills one 100K input pass at that rate, add $0.075 to the cached scenario, making the modeled total $0.20. Use your actual cache-creation usage record rather than assuming setup is free.

A cache kept for a long time but seldom read can be a poor trade. Our calculation is useful because it exposes storage and reuse separately; it is not a promised saving for every prompt.

Free access and paid setup are model-specific

The free column in Google's price table applies only to the listed model and capability, subject to eligibility and quota. AI Studio's interactive access is not identical to every model's API free tier. Review the current limits in your dashboard rather than treating a price of zero as unlimited capacity.

The billing documentation describes both Prepay and eligible Postpay arrangements. Linking billing can involve an initial prepaid credit purchase; do not equate it with creating another key. For the actual project and key steps, use our Gemini API key tutorial.

Google's price table also distinguishes free and paid data-use treatment. Review the applicable service terms before sending sensitive information; a token-rate comparison is not a privacy or compliance assessment.

Google direct, Cloud and gateways have different billing boundaries

For DataLLM Lab access, check our gateway catalog separately. A gateway model ID and price are not evidence of Google's direct billing. Likewise, our historical Gemini cost-per-task tests retain their recorded route and list-price date; they are not current Google invoices.

Log enough information to reproduce the cost comparison

• Exact requested and returned model, provider route and request mode.
• Uncached input, cached input and billed output, including thinking where charged.
• Cache-creation usage, retained token volume and storage duration.
• Tool usage, retries and the number of completed application tasks.
• Rate source and effective date, then reconcile the estimate with billing records.

Choose a model only after checking your task's quality and latency requirements. The lowest token rate alone does not establish the lowest cost per successful task. This page supplies pricing arithmetic, not a model-quality ranking.

FAQ