benchr
Anthropic's Claude lineup is renowned for its high quality in coding, reasoning, and writing. The pricing structure is segmented cleanly across three model size families. Below is the live pricing table for the Anthropic Claude API family, sourced dynamically from our index.
| Model ID | Input / 1M | Output / 1M | Cached Input / 1M | Max Tokens |
|---|---|---|---|---|
| claude-opus-4-8 | $5.00 | $25.00 | $0.500 | 1,000,000 |
| claude-opus-4-7 | $5.00 | $25.00 | $0.500 | 1,000,000 |
| claude-sonnet-4-6 | $3.00 | $15.00 | $0.300 | 1,000,000 |
| claude-haiku-4-5 | $1.00 | $5.00 | $0.100 | 200,000 |
Unlocking Massive Savings with Prompt Caching
Anthropic was a pioneer in offering Prompt Caching on public APIs. If you have long, static system instructions, documents, or context history, you can tag them to be cached. This yields significant savings:
- 90% Off Input Cost: Cached input reads cost only 10% of the standard input rate (e.g. $0.50/1M for Opus 4.8 instead of $5.00/1M).
- Fast Retrieval: Caching drastically reduces first-token latency (time-to-first-token) for long documents, which is ideal for real-time applications.
- Simple API Integration: Simply add
"cache_control": {"type": "ephemeral"}to block headers in your messages list.
For cross-provider comparison, read the global AI Model Pricing Comparison, or calculate custom usage via the Cost Calculator.
The Claude cost lever: cached context
Claude pricing is easiest to misunderstand when you look only at standard input. The practical question is how much of your prompt is repeated. Long system prompts, policy manuals, codebase summaries, and retrieval packets can become dramatically cheaper once cached.
For agents, measure cold-start calls separately from warm-cache calls. A model that looks expensive on the first request can become competitive when most of the repeated context is billed at the cached rate. If every request is unique, use the standard input and output rates instead.
Choosing between Haiku, Sonnet, and Opus
Use Haiku when the task is narrow and verifiable: routing, extraction, tagging, short summaries, and first-pass support triage. Use Sonnet when the task needs stronger writing, coding, or reasoning but still has a human or automated check. Use Opus when failure is expensive enough that a higher success rate pays for itself.
The cleanest architecture is tiered: Haiku tries the easy work, Sonnet handles the default path, and Opus receives only the difficult or high-risk calls. That structure usually beats sending every request to one model, because it protects quality without paying flagship prices for trivial calls.
Claude's economics are especially sensitive to document workflows. If users repeatedly ask questions over the same policy, contract, or codebase, cache-aware design can make a stronger model affordable. If every request brings unrelated context, the same model may need stricter routing to stay within budget.
For procurement, document why a request goes to each Claude tier. That one note makes future cost reviews easier: if a call goes to Opus, there should be a measurable reason such as lower review burden, fewer failed tool calls, or better long-document consistency.
Keep those routing decisions visible in product analytics. When a feature becomes expensive, you want to know whether the cause is more users, longer prompts, more failed attempts, or too many calls being sent to Opus. Without that split, optimization turns into guesswork.