Claude API Pricing (August 2026)

Start in the Claude Console, create an API key, then send a request to https://api.anthropic.com/v1/messages. Anthropic's official quickstart includes cURL, Python, and TypeScript examples. The API key has no subscription fee; the selected model, token counts, tools, and pricing modifiers determine the bill.

Use claude-opus-5 for the current Opus tier, claude-sonnet-5 for Sonnet, and claude-fable-5 for Fable. Haiku's pinned ID is claude-haiku-4-5-20251001; claude-haiku-4-5 is its convenience alias. Anthropic says dateless IDs from the 4.6 generation onward are pinned snapshots, not pointers that silently move to a later model. The Models API returns the models available to your account.

Prompt caching separates writes from reads. A 5-minute cache write costs 1.25x the model's base input rate, a 1-hour write costs 2x, and each cache hit costs 0.1x. On Opus 5 that means $6.25/M for a 5-minute write, $10/M for a 1-hour write, and $0.50/M for reused input. Cache the stable system prompt, tool definitions, or document prefix when the same content will be read again.

Batch processing halves input and output for asynchronous jobs. Anthropic states that the Batch discount stacks with prompt-cache pricing, so a batched Opus 5 cache hit costs $0.25/M and batched output costs $12.50/M. Fast mode does not work with Batch. The calculator above supports cache and Batch together; compare providers with the LLM API pricing calculator.

Claude Free, Pro, Max, Team, and Enterprise are subscription products. They do not create a shared pool of standard Claude API tokens. API usage is metered through a developer account and billed from input, cached-input, output, and any paid server-side tool usage.

Claude Code can run under an eligible subscription for interactive use or with API credentials for usage-based automation. When Claude Code uses an API key, the selected model's token rates apply. Use the table and calculator for API-key workloads; use the consumer plan page for a seat or subscription decision.

Haiku 4.5 ($1/$5) is the low-cost tier for classification, extraction, routing, and short summaries. Sonnet 5 is the faster general-purpose tier at $2/$10 through August 31, then $3/$15. Opus 5($5/$25) is Anthropic's recommended starting point for complex agentic coding and enterprise work. Fable 5 ($10/$50) is the highest-cost self-serve tier for workloads where an evaluation proves the extra capability pays back.

The limit matters: the newest expensive model is not automatically the cheapest system. A routed workload can keep routine traffic on Haiku or Sonnet and escalate only failures to Opus or Fable. Fable loses when its quality gain does not offset a 2x input and output premium over Opus.

At the flagship tier, Claude Opus 5 costs $5/$25 per million input/output tokens. That matches GPT-5.6 Sol on input and is cheaper than its $30/M output rate. Sonnet 5's temporary $2/$10 rate undercuts GPT-5.6 Terra at $2/$12 on output, while Gemini 3.1 Pro lists at $2/$12 on prompts up to 200K tokens.

List price does not settle model quality, context behavior, or tokenization. Claude 4.7 and later use a tokenizer that Anthropic says produces about 30% more tokens for the same text, depending on the workload. Compare the full OpenAI and Gemini price tables, then rerun cost math with tokens measured from the model you plan to deploy.