Skip to content
sinc-LLMBook a 30-minute call

Engineering instrument

LLM Cost Calculator

Compare published text-token rates and calculate a reference estimate from your token counts. Checked October 21, 2026; confirm current pricing and model availability with the provider.

Enter token counts rather than prompt text. This calculator runs locally in your browser and does not call an AI provider.

TokenCalc Logo

Estimate standard text-token charges from the sourced rate table below. Enter input tokens, billed output tokens and the number of requests. This is a reference estimate, not a provider bill.

Calculate token costs

Calculate a token-cost estimate

Use counts from your provider usage record or the tokenizer for the exact model. No prompt text is required.

Enable JavaScript to calculate the estimate.Formula: (input tokens × input rate + billed output tokens × output rate) ÷ 1,000,000 × requests. Includes repeated identical requests, not a forecast of future usage. Excludes caching, batch discounts, taxes, tools, review and hosting.

Provider pricing source · Include review and operating costs

USD per 1 million tokens, using standard uncached text input and text output. This is a selected reference set, not a complete model catalogue or availability guarantee. Batch, caching, tools, media, hosting and human review can change the total.

ModelProviderInput / 1MOutput / 1MSource
GPT-4o mini — legacy / deprecatedOpenAI$0.15$0.60Provider source
Gemini 3 Flash Preview — legacy previewGoogle$0.50$3.00Provider source
Claude Haiku 4.5Anthropic$1.00$5.00Provider source
o3-mini — legacy referenceOpenAI$1.10$4.40Provider source
GPT-5 — legacy / deprecatedOpenAI$1.25$10.00Provider source
Claude Sonnet 4.6Anthropic$3.00$15.00Provider source
Claude Opus 4.6Anthropic$5.00$25.00Provider source

Download the dated rate reference (JSON)

Choose a model for your workload

Compare candidates on the same representative inputs, acceptance criteria and latency budget. A lower token rate does not establish lower cost per useful result. Check lifecycle and account availability before building around a model; legacy entries remain here only as cost references.

For a complete operating-cost example, use the worked cost-per-query worksheet to include retries, review time and infrastructure.

#1: Right-Size Your Model

Evaluate smaller available models on representative cases. Compare accepted-output quality, latency and full task cost before changing the workload assignment.

#2: Use Prompt Caching

Check the exact provider and model rules for eligible prompt caching. Cache reads, cache writes, expiry, misses and output charges affect the total; an input discount is not a whole-request saving.

#3: Optimize Prompts

Remove redundant context only when the same task still meets its acceptance criteria. Preserve the information and constraints the model needs, and measure the effect before assuming a saving.

Frequently Asked Questions

Which model has the lowest cost for my workload?

There is no universal cheapest suitable model. Compare input and output rates, cache categories, billed reasoning tokens and acceptance rates on the same workload. The selected rate table is not a market-wide ranking.

How do I reduce LLM API costs?

Measure a baseline, then test smaller models, eligible caching or batch processing while checking accepted-result quality and latency. Include retries, tools, human review and infrastructure before claiming a saving.

How should I compare GPT and Claude costs?

Use the exact model IDs and current provider rates. Multiply input and billed output tokens by the matching rates, then add separately billed features. Check model lifecycle and evaluate task quality rather than choosing on price alone.

How does this calculator handle token counts and privacy?

Enter provider-reported or model-tokenizer counts. The arithmetic runs locally and does not send these inputs to an AI provider. It does not tokenize text, enforce model context limits, apply cache discounts or call a model. Account availability and current pricing must be checked with the provider.