Cost calculator

Estimate text workloads from input/output tokens, request volume, cache reuse, and processing-tier pricing.

Usage input

Estimate a repeated workload from tokens per request, request volume, cache reuse, and processing-tier pricing.

More presets (10)

Choose a preset or scroll for custom input.

Explore benchmark scores to see which models perform best on specific tasks.

Input

Used to estimate cost per active user; not included in shared links.

Use average revenue across the same active users, including free users. Not your subscription price; stays on this page.

High-usage customer check (optional)

Test one user's API cost using this workload's effective cost per base request. Does not change the monthly total or shared link.

Estimate how many base requests fit your token budget; not an enforced spending limit.

For measured reasoning usage, provider counters differ: OpenAI Responses and Anthropic Messages already include reasoning/thinking inside output_tokens; Gemini Generate Content reports candidatesTokenCount and thoughtsTokenCount separately, so add them. For planning, make Output tokens the total billed generated output you expect. See reasoning-token accounting.

Estimate input tokens from text

Text is processed only in your browser and is not saved or sent anywhere. This is a reference count for o200k_base, not the exact billable token count for the selected model or provider. It also excludes API message, tool, schema, image, and other request overhead.

Advanced options

Share of input tokens billed at the cached rate, not the share of requests with a cache hit.

Include only billable retries, fallbacks, or orchestration calls. For agent workflows, compare cost per successful task using measured outcomes.

Region-specific processing surcharges are excluded from this estimate. OpenAI charges 10% extra for eligible data-residency models released on or after March 5, 2026; support varies by model, endpoint, and region. Check eligibility.

Estimate cache shares from reuse

A warm cache entry is one reusable prefix across a group of requests. Assume one write/miss, then reuse until the prefix changes or goes cold.

TTL is a guide, not a guaranteed cache miss. Reuse may refresh retention, and behavior varies by model and route.

Planned cache rates are assumptions, not guaranteed discounts. Short prompts or changing prefixes may be ineligible for caching or fail to match reusable content. Check your model's minimum cacheable length and measured cached-token usage before relying on cache savings in a budget. See the prompt caching guide for eligibility and reuse considerations.

Cost estimates use the generated model database last built on Nov 16, 2026. Pricing, lifecycle, and capability fields can be incomplete or provider-specific, so verify production decisions with the official provider.

How to use this page

Start with a preset for a forecast, then change tokens and request volume to match your product. If you already have token totals from a completed request or session, choose Use actual token usage and enter that breakdown directly.

How pricing is calculated

Costs are calculated as (tokens ÷ 1,000,000) × price per 1M tokens. For planned workloads, cache hits and cache writes partition the total input-token budget into standard input, cached input, and cache-write input. The formula is total = ((standardInputTokens × inputPrice) + (cachedInputTokens × cachedInputPrice) + (cacheWriteTokens × cacheWritePrice) + (outputTokens × outputPrice)) × requests. Billable extra-call overhead scales the request volume after the per-call price is determined; when it is above zero, the per-request result is the effective cost per planned base request. If no separate cached-input rate is stored, cached input uses the selected input rate. A positive cache-write rate requires an explicit applicable cache-write price; the calculator does not infer one for another tier or pricing period.

Need concrete examples first?

Read the calculator examples guide for chatbot, RAG, summarization, and coding-agent inputs before changing the fields.

Read examples
Not sure if you need API pricing?

Check when a subscription is enough and when usage-based API pricing matters for product work.

API vs subscription
Which provider is cheapest overall?

Compare OpenAI, Anthropic, Google, Mistral, and DeepSeek across cheapest chat, mid-range, and reasoning model pricing.

Provider comparison
What do benchmark scores mean?

See MMLU, GPQA, HumanEval, and other benchmark scores across providers to understand model quality beyond pricing.

Benchmark scores
Want to compare models side by side?

Put 2-3 models next to each other to compare pricing, context windows, modalities, and capabilities.

Compare models
Observed cost per accepted task

Divide actual billable spend for the same cohort and period—including charged retries, fallback and tool calls—by accepted tasks. This is independent of the model estimate above.

Enter both observed values to calculate.

Use one consistent acceptance rule, and include billable failed attempts. Inputs are calculated locally, never stored or sent. Taxes, credits and unbilled attempts are not inferred.