Google API pricing: every Gemini model, per 1M tokens

Budget a Gemini API workload from the model and media it actually processes. Separate text charges, generated media and optional features before comparing free and paid usage.

How Google pricing works

Gemini’s pricing reference separates model entries, input and output types, free and paid tiers, and supported processing modes. Official documentation.

Create separate workload estimates for text answers and generated media. A mixed-media application needs an explicit record of which branch is called and how frequently, rather than a single average text prompt.

Sketch the complete Gemini workflow before entering a token estimate. A document assistant might receive extracted text, a scanned page or a mixture of text and images. An image tool might generate a new asset or edit a supplied one. Keep those paths separate because their input preparation, output settings and charge categories need different evidence.

For each path, record the model identifier, API interface and expected output. Add any optional grounding or other hosted feature explicitly. This prevents a plain-text estimate from quietly becoming the budget for a much broader media application. If a feature’s charge is not represented in the calculator, keep it as an additional unresolved budget item rather than forcing it into an invented token conversion.

Full price list

Read the model entry and the unit together. Compare like with like: a media-output row and a text-output row do not describe equivalent work. Inspect the qualifying mode and the free or paid column before carrying a value into a planning document. Missing data means the relevant fact has not been verified here; it should not be interpreted as a free service.

Follow the source link when a workload depends on a footnote or a particular output setting. Keep the date of the comparison with your saved assumptions. If the application changes its input type or generation mode, revisit the appropriate row instead of continuing to use the original text-only estimate. A useful budget remains connected to the request that actually runs.

Estimate your cost

Use the Gemini workload calculator for token-based estimates, and inspect the table’s unit for non-token charges.

Begin with a small representative text request and inspect its returned usage. Keep the selected identifier attached to those observations. Then build separate scenarios for a concise answer, a detailed answer and the relevant conversation context. Label any unmeasured input as an assumption so a colleague can see where the forecast still needs evidence.

For a mixed-media product, estimate each branch independently and combine them only after you understand their units. Keep the branch frequency explicit. A rarely used image feature can still matter to the application budget, while an average across unrelated outputs can hide the behavior you need to control. Preserve the scenarios as a decision record rather than only copying their final totals.

Three worked scenarios

• Chatbot: compare concise and detailed replies while keeping conversation history explicit.
• Summarization: identify whether the input is plain text or media that requires separate processing assumptions.
• Coding assistant: include the code context and generated patch, then measure accepted results.

For a Gemini document assistant, compare a clean text document with a scan that conveys the same information. Evaluate whether the result identifies the requested evidence and whether preprocessing changes the answer. Do not count a fluent description of the document as successful extraction when the application needs specific fields.

For a chatbot, test the same question with a short and a context-rich conversation. Keep relevance and correctness as the acceptance bar while comparing usage. For coding, use a saved repository task and verify the generated change with project checks. If the application needs a repair request, include that work in the evaluation rather than treating the first generated patch as a completed task.

How to reduce cost on Google

The Interactions API supports implicit caching; explicit cache objects belong to the generateContent workflow. Official documentation.

The documented Batch API is an asynchronous generateContent feature. Check interface eligibility before assuming a batch mode can be added to an Interactions request. Official documentation.

Identify which interface the application uses before adding an optimization. The current Gemini guides distinguish Interactions from generateContent for explicit cache objects and batch processing. Make a small proof of the intended request path and result handling before assuming a feature can be enabled with a single option on the existing request.

For repeated material, measure actual cache-related usage and keep the repeated prefix stable during the experiment. Do not infer a cache hit from a faster response alone. For offline batches, assign input identifiers and reconcile individual results before marking the job complete. The lower-cost path is useful only if its lifecycle fits the application and its results still meet the acceptance criteria.

Billing, credits and limits notes

Review billing in the project that owns the application, including the applicable tier and balance behavior. Official documentation.

Read free-tier conditions and project capacity before committing to a deployment budget.

Keep the Cloud project visible in the budget record. Document who owns its billing arrangement, which application uses it and how the workload can be paused. Inspect actual project controls rather than assuming a familiar Google Cloud notification behaves like an enforced stop for every Gemini request.

Separate a prototype’s temporary funding from its ongoing economics. Before enabling a paid continuation, compare current model rates with the measured request mix and decide which optional features are authorized. After representative usage, reconcile the assumptions with the account view. Explain differences in terms of request shape, feature use or account conditions instead of silently replacing the forecast.

Price change history

Use a recorded change as a reason to review the affected application paths. Keep the previous model and workload assumptions, then recalculate the same scenarios under the updated source values. A media-model change should not be described as a general improvement in the cost of every Gemini text request.

If a change makes another model attractive, repeat the quality evaluation before switching. Preserve output checks, interface compatibility and data-handling requirements alongside the price comparison. When the official page and a displayed row appear inconsistent, submit the exact model, unit and source link for correction and keep the disputed figure out of an external quote until it is resolved.

Last verified 12 Sep 2026 · Source ↗

Frequently asked questions

Sources

• Gemini API getting started ↗
• Gemini models ↗
• Developer API pricing ↗
• Rate limits ↗
• API-key management ↗
• API errors ↗
• Troubleshooting ↗
• Billing ↗
• Context caching ↗
• Batch API ↗
• Python SDK reference ↗

Last verified 12 Sep 2026 · Source ↗