← सभी गाइड पर लौटें

Claude API Pricing 2026: Decode Every Token Charge

Understanding Claude API pricing starts with separating ordinary input, generated output, cache writes and cache reads. These categories have different rates; a single average token price can hide where your budget goes. This 2026 guide explains the billing mechanics and shows how to compare direct API costs with a low-cost relay.

Input, output and cache tokens explained

Input tokens represent material Claude processes: system instructions, conversation messages, documents and tool definitions. Count the complete request, not just the latest user question. Anthropic provides a token-counting endpoint for supported request content, which is more reliable than estimating from word counts. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/token-counting))

Output tokens are generated by the model, including text and tool-call content. They have a separate rate: Sonnet 4.6, for example, costs $3 per million ordinary input tokens and $15 per million output tokens at standard direct-API rates. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing))

Cache tokens are input tokens billed differently because reusable prompt content is being stored or retrieved. In native Claude Messages responses, input_tokens excludes cache writes and reads; those appear separately as cache_creation_input_tokens and cache_read_input_tokens. Do not charge cached tokens again at the ordinary input rate. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/prompt-caching))

Cache writes cost more; cache reads cost less

Anthropic charges 1.25 times the ordinary input rate for five-minute cache writes and 2 times for one-hour writes. Reads are model-specific: many models charge 10% of ordinary input pricing, but Opus 5.5 and Sonnet 5.5 charge 5%, while Fable 5.1 charges 2.5%. Never assume one cache discount applies across every Claude version. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing))

Caching does not mean receiving a saved answer. Claude still generates and bills new output; the saving applies to reused input context. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/prompt-caching))

A worked Claude API cost example

Use this calculation with prices quoted per million tokens:

Request cost = (ordinary input × input rate + cache writes × write rate + cache reads × read rate + output × output rate) ÷ 1,000,000. Split five-minute and one-hour writes when both occur. ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing))

Consider Sonnet 4.6 with a reusable 10,000-token document, 1,000 fresh input tokens and 500 output tokens. Its five-minute write rate is $3.75 per million tokens; reads cost $0.30. Calculated from those published rates:

• First request: $0.0375 cache write + $0.003 input + $0.0075 output = $0.048.
• Matching follow-up: $0.003 cache read + $0.003 input + $0.0075 output = $0.0135.
• Without caching: each equivalent request costs $0.0405.

The initial write costs extra, but repeated reads recover that premium. These figures illustrate direct Anthropic pricing, not relay charges. ([platform.claude.com](https://platform.claude.com/docs/en/models/sonnet-4-6/overview))

Make prompt caching work in production

• Stabilize the prefix. Put reusable instructions and documents first; keep timestamps and changing questions afterward.
• Enable caching explicitly. Set cache_control with type ephemeral on the last stable content block. Add ttl with value 1h when longer retention is justified.
• Check minimum length. Sonnet 4.6 requires 1,024 cacheable tokens; other models have different thresholds.
• Verify a real hit. Send an identical prefix again within its lifetime and inspect cache_read_input_tokens. Exact prefix matching matters. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/prompt-caching))

Before deployment, count representative prompts and record actual usage afterward. Compare cost per successfully completed task rather than token price alone. ([platform.claude.com](https://platform.claude.com/docs/en/build-with-claude/token-counting))

Where 59API lowers the bill

59API offers pay-as-you-go Claude and GPT access, with Claude Code, Codex and OpenAI SDK compatibility. It describes its Opus, Sonnet, Haiku and Fable access as native, official-quality models without downgrading. ([59api.com](https://59api.com/blog/your-first-claude-api-call-with-59api-in-5-minutes-2879))

Its published Sonnet 4.6 rates are approximately $0.6944 input and $3.61 output per million tokens, making it a compelling low-cost option. Prices are displayed in approximate USD but billed in CNY; check cache rates separately rather than applying Anthropic’s multipliers automatically. A referral rebate is also available. ([59api.com](https://59api.com/pricing?lang=en))

If that fits your workflow, consider signing up for 59API and benchmarking a small, repeatable workload before scaling.

मूल्य · दस्तावेज़