Anthropic API Pricing (2026)
Anthropic Claude API pricing for 2026: per-million-token rates for Opus, Sonnet, and Haiku, plus prompt caching and the batch discount.
Anthropic’s 2026 Claude API runs from Haiku 4.5 at $1 per million input tokens, to Sonnet 4.6 at $3, to Opus 4.8 at $5. Cache reads cost 10% of input, and the Batch API takes 50% off. Sonnet is the default for production agents and coding; Haiku handles high-volume extraction; Opus is for the hardest reasoning. Rates below are updated 2026 from Anthropic’s pricing docs.
Per-million-token rates (updated 2026)
| Model | Input | Cache read | Output | Use it when |
|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $0.50 | $25.00 | Hardest reasoning, complex agents |
| Claude Sonnet 4.6 | $3.00 | $0.30 | $15.00 | Most production agent and coding work |
| Claude Haiku 4.5 | $1.00 | $0.10 | $5.00 | High-volume support, extraction, routing |
Opus 4.5 through 4.8 share the same $5 / $25 rate. Sonnet 4.5 and 4.6 share $3 / $15. The latest Opus, Sonnet, and Fable models include the full 1M-token context window at standard pricing.
How caching and batching work
- Prompt caching: a 5-minute cache write costs 1.25x base input, a 1-hour write costs 2x, and a cache hit costs just 0.1x. Caching pays off after a single read on the 5-minute window.
- Batch API: 50% off input and output for asynchronous jobs. Sonnet 4.6 batch is $1.50 / $7.50.
- The two discounts stack.
Honest read
Claude is not the cheapest on the input column, but Sonnet is the model many teams settle on for agents and coding because of its reliability per dollar. The big lever here is prompt caching: agent workloads resend large system prompts constantly, and a 0.1x cache-read price changes the economics more than switching providers would. If your work is mostly simple and price-driven, compare against Gemini and OpenAI.
Related
See the cheapest LLM API comparison for the full table and cost control for caching and routing in practice. Background: what an LLM is and how the KV cache relates to prompt caching. For implementation help, see AI integration.