Claude API token calculator
Estimate your Claude API bill model by model. Enter the input and output size of a typical request and how many you run per month, and see the cost for Fable 5, Opus 4.8, Sonnet 4.6, and Haiku 4.5 side by side, with prompt caching and Batch API savings applied. June 2026 pricing, nothing sent until you ask.
Which Claude model?
Pricing is per million tokens, input and output billed separately. Your pick drives the headline; all models are compared below.
Size a typical request
Enter the input and output of one representative request and your monthly volume. Defaults model a typical assistant workload.
Rough guide: 1 token ≈ 4 characters ≈ 0.75 words. For an exact count, use Anthropic's token-counting endpoint with your model.
Savings levers (prompt caching & batch)
June 2026 list pricing from the Anthropic pricing docs. Details in Assumptions & sources.
Every model, same workload
Your exact request size and volume, priced across the full Claude lineup. The cheapest option isn't always the right one, but you should know the gap.
Right-sizing the model is usually the biggest cost lever: moving a workload that doesn't need flagship reasoning from Opus to Sonnet or Haiku can cut spend 60 to 80% before any caching. See how iServ deploys Claude in production.
Send me this estimate
We'll email your breakdown and a short, practical review of how to cut Claude API cost without losing quality, model routing, caching, batching, and prompt design.
Tokens in, tokens out, times volume
No black box. Every number traces back to per-million-token rates you can verify in Anthropic's docs.
1. Input + output
Each request bills for the tokens you send (prompt, context, tools) and the tokens the model writes back, at different per-million rates.
2. Times volume
Multiply the per-request cost by requests per month. This is where small per-call differences become real budget lines.
3. Caching
A reused prefix (system prompt, document) can be cached: reads cost about 10% of input. The calculator applies your cache-hit share.
4. Batch
Non-urgent work sent through the Batch API is 50% off input and output, processed asynchronously within 24 hours.
The rates behind the numbers
June 2026 standard-API list pricing, per million tokens. The lineup moves quickly; verify against the live docs before committing a budget.
| Model | Input / Output (per 1M) | Notes |
|---|---|---|
| Claude Fable 5 | $10 / $50 | Most capable; 1M context, 128K max output. |
| Claude Opus 4.8 | $5 / $25 | Flagship Opus; 1M context, 128K max output. |
| Claude Sonnet 4.6 | $3 / $15 | Best balance of speed and intelligence; 1M context. |
| Claude Haiku 4.5 | $1 / $5 | Fastest, lowest cost; 200K context. |
| Prompt caching | read 0.1× input | 5-minute cache write 1.25× input, 1-hour write 2× input. |
| Batch API | −50% | Half price on input and output, async within 24 hours. |
Source: Anthropic pricing documentation. This tool models standard global API pricing; it does not include Opus Fast mode, the 1.1× US data-residency multiplier, or per-model cache-write minimums. Token estimates use ~4 characters per token; for billing-grade accuracy use Anthropic's token-counting endpoint.
Answered directly
How much does the Claude API cost?
Claude API pricing is per million tokens, with input and output priced separately. As of June 2026: Claude Fable 5 is $10 input / $50 output, Claude Opus 4.8 is $5 / $25, Claude Sonnet 4.6 is $3 / $15, and Claude Haiku 4.5 is $1 / $5. Your bill is those rates applied to the tokens each request uses, times your volume.
What is the difference between Opus, Sonnet, and Haiku pricing?
They trade capability for cost. Opus 4.8 ($5/$25) is the flagship for the hardest reasoning and long-horizon agentic work; Sonnet 4.6 ($3/$15) is the best balance for most production workloads; Haiku 4.5 ($1/$5) is fastest and cheapest for high-volume, simpler tasks. Fable 5 ($10/$50) sits above Opus for the most demanding work. Routing each workload to the smallest model that meets the quality bar is usually the single biggest cost lever.
How does prompt caching change Claude API pricing?
Prompt caching reuses a stable prefix across requests. Cache reads cost about 10% of the base input price, a 90% saving on the cached portion, while writing the cache costs 1.25× base input for a 5-minute window or 2× for a 1-hour window. For workloads that resend a long system prompt or document on every call, caching can cut input cost substantially. Set the cache-hit share in the calculator to see the effect.
How do I estimate how many tokens my requests use?
A rough rule is about 4 characters or 0.75 words per token, so 1,000 words is roughly 1,300 tokens. Code and non-English text tokenize differently, so for billing-grade accuracy use Anthropic's token-counting endpoint with the exact model you will call rather than a generic tokenizer. This calculator lets you enter request sizes in tokens or words.
Build it cost-aware from day one
iServ deploys Claude in production for businesses: model routing, caching, batching, and governance that keep quality high and the bill predictable.