Claude API cost depends on the selected Claude model, normal input tokens, eligible cache-read tokens, and output tokens. Add each token group at its published per-million rate. The calculator below models standard direct API text rates; cache writes, regional multipliers, partner platforms, and provider-side token differences require separate verification.
This estimate uses published text-token rates. It excludes tool calls, media, storage, taxes, regional pricing, platform fees, and provider-side token differences.
Use identical normal input, eligible cache-read, and output counts for each model. The lower estimate applies only to that workload and the verified standard direct-API rates; response quality, latency, retries, context needs, and cache-write charges can change the practical decision.
| Provider / model | Input / 1M | Cached / 1M | Output / 1M | Context | Accuracy | Verified |
|---|---|---|---|---|---|---|
| Claude Fable 5Anthropic | $10 | $1 | $50 | 1,000,000 | Estimate | 2026-08-09Verified |
| Claude Haiku 4.5Anthropic | $1 | $0.1 | $5 | 200,000 | Estimate | 2026-08-09Verified |
| Claude Opus 5Anthropic | $5 | $0.5 | $25 | 1,000,000 | Estimate | 2026-08-09Verified |
| Claude Sonnet 4.6Anthropic | $3 | $0.3 | $15 | 1,000,000 | Estimate | 2026-08-09Verified |
| Claude Sonnet 5Anthropic | $2 | $0.2 | $10 | 1,000,000 | Estimate | 2026-08-09VerifiedValid through 2026-08-31 |
The 17 model rows use USD list prices per one million text tokens, checked against the linked official provider pages. OpenAI and Gemini long-context tiers are applied automatically from the entered input count. Batch, audio, image, video, tools, storage, regional, and platform fees may differ.
Normal input, cache reads, and output are separated by the Claude API cost calculator and priced at their Anthropic rates. Output is commonly the larger line item, so a short prompt with a long answer can cost more than an input-only comparison suggests.
The browser uses a compatible encoding for a Claude planning estimate, not Anthropic's billing tokenizer. Published rates are applied by the Claude API cost calculator to your scenario; the local token count is not presented as provider usage.
Only input tokens entered as cached receive Anthropic's published cache-hit or refresh rate in the Claude API cost calculator. Cache creation, five-minute and one-hour write multipliers, duration, minimum size, prefix rules, invalidation, and actual hit rate must come from current Anthropic documentation and application logs.
Anthropic models are the scope of the Claude API cost calculator. Compare both providers with the GPT vs Claude cost calculator and review the wider LLM API pricing table.
Across normal input, cache-read input, output, zero-token, 1-token, 1 million-token, 10 million-token, and high-volume cases, the Claude API cost calculator measured one centralized set of Anthropic rates. Its limitation is billing scope: cache writes, regional multipliers, tools, media, retries, taxes, and Anthropic-side tokenization require separate evidence. Anthropic official API pricing. Last verified: .
| Evidence | Observed result or boundary |
|---|---|
| Measured Claude cases | 0, 1, 1,000, 1 million, and 10 million tokens with cache and output variations. |
| Comparison boundary | Anthropic Claude results can be compared with OpenAI, Google Gemini, and DeepSeek rates. |
Claude API cost questions
Pricing is calculated by the Claude API cost calculator from normal input, eligible cache-read input, and output using the selected Claude model's published per-million rates.
No. Cache-hit or read rates are modeled by the Claude API cost calculator, while Anthropic cache writes and retention tiers can use separate multipliers.
No. A local Estimate can feed the Claude API cost calculator, but Anthropic token-counting and API usage records provide the authoritative request total.
No. The documented US-only inference multiplier and partner-cloud terms are not automatically applied by the Claude API cost calculator.
Use the Claude API cost calculator with representative provider token counts when possible. It is not for invoicing; apply current prices and real cache behavior to the workload, then reconcile against provider usage.