Claude API Pricing in 2026: Opus 4.8, Sonnet 4.6 & Haiku 4.5 Token Costs Explained

Claude API billing is simple in principle — you pay per token — but the details (input vs output rates, cache pricing, and model choice) are where real money is won or lost. Here’s a practical 2026 guide to what Claude costs and how to spend less without giving up quality.

How Claude API billing works

Claude uses pay-as-you-go token billing in USD. There’s no subscription and no monthly minimum on the API itself: you’re charged for the tokens each request consumes, split into:

• Input tokens — everything you send: system prompt, conversation history, documents, tool definitions.
• Output tokens — everything the model generates back. Output is always priced higher than input.
• Cache tokens — discounted reads (and a small write premium) when you reuse a stable prompt prefix.

Tokens are billed per million tokens (MTok). A rough rule of thumb: 1 token ≈ 0.75 English words, so 1M tokens is roughly 750,000 words in or out.

The model tiers and what drives their price

The three core tiers trade off intelligence against cost:

• Claude Opus 4.8 — the flagship. Highest input/output rates, justified only when you need its reasoning depth (hard debugging, planning, agentic tasks).
• Claude Sonnet 4.6 — the workhorse. Far cheaper than Opus while handling the large majority of coding and production traffic. The best intelligence-per-dollar for most teams.
• Claude Haiku 4.5 — the budget tier. Cheapest and fastest, ideal for classification, extraction, routing, and high-volume simple calls.

The single biggest lever on your bill isn’t a discount — it’s using the cheapest model that still passes your evals for each task. Routing bulk work to Sonnet or Haiku and reserving Opus for the hard 10% can cut spend by more than half.

Prompt caching: the discount most teams underuse

If you send the same large prefix repeatedly — a long system prompt, a document, a tool schema — prompt caching stores it so subsequent requests read it at a steep discount instead of paying full input price each time.

Caching pays off when:

• You have a large, stable prefix reused across many requests (RAG context, long instructions, big tool definitions).
• You run multi-turn conversations where history accumulates.
• You run agents that replay the same context every step.

There’s a small write premium the first time you cache, so it’s worth it when the prefix is reused at least a few times. For long agents, cache hits routinely cut the input portion of the bill dramatically.

Other cost levers

• Batch processing — for non-urgent bulk jobs, batch APIs typically run at a discount versus real-time calls.
• Cap max_tokens — output is the expensive half; don’t let the model ramble when you only need a short answer.
• Trim context — prune stale tool results and old turns from long agent loops (see our context-management guide) so you’re not paying input on tokens that no longer matter.
• Pick the right context window — only pay for 1M-context models when you actually need them.

Cutting the rate itself

Beyond optimization, you can reduce the rate you pay by accessing Claude through a pay-as-you-go gateway. AI Prime Tech routes requests to the same Claude models — Opus 4.8, Sonnet 4.6, Haiku 4.5 — at up to 80% below official list prices, with one API key that also covers GPT and Gemini. Because it’s a drop-in for the Anthropic SDK, you change one base URL and your existing billing math simply gets smaller. For teams running continuous agents or high request volume, the combination of model routing + caching + a discounted gateway is what keeps the bill sane.

A quick cost-estimation method

To sanity-check a feature before you ship it:

1. Estimate average input tokens (prompt + context) and output tokens per request.
2. Multiply by the per-MTok input and output rates for your chosen model.
3. Multiply by expected requests per day.
4. Subtract realistic cache savings on the stable prefix.

Run that for Opus vs Sonnet vs Haiku and the right model usually becomes obvious. The cheapest model that passes your quality bar, plus caching, plus a discounted gateway, is the winning formula for 2026.