Claude Cost Calculator

Anthropic sells capability in tiers, and the price gaps between them are large enough to decide your architecture. Haiku is built for high-volume, well-defined tasks — classification, extraction, routing, short answers — where its per-token rate is a small fraction of the larger models'. Sonnet is the workhorse tier for reasoning-heavy work like coding, analysis and long-form writing. The practical question is rarely "which model is best" but "what is the cheapest tier that clears my quality bar", and the honest way to answer it is to run your own test set through each tier and compare failures, not vibes.

A useful mental model: if a task can be described completely in the prompt — fixed format, clear rules, no judgment calls — it probably belongs on Haiku. The moment a task needs the model to weigh trade-offs or hold a long chain of reasoning, budget for Sonnet rates in your estimate above.

Prompt caching is Anthropic's most underused discount. You pay a small premium to write a prompt prefix into the cache, and subsequent requests that reuse that exact prefix read it at a large discount — roughly an order of magnitude cheaper than regular input. The catch is the cache lifetime: entries expire after a few minutes of inactivity, so the discount only materializes when requests arrive close together.

That shape fits three workloads almost perfectly: chatbots with a large system prompt serving steady traffic, RAG applications that pin the same reference context across many questions, and agents that repeatedly re-read the same documents in one session. If your traffic is sparse — a request every ten minutes — cache entries expire between calls and the write premium becomes pure overhead. Estimate with the full-price number from the calculator above, then treat caching as an optimization to validate with real traffic.

Three habits keep Anthropic invoices predictable. Set max_tokens deliberately — it is a hard cost ceiling per request, and generous defaults quietly permit expensive answers. Watch conversation length: Claude's large context window makes it easy to carry an entire session's history in every request, and all of it is billed as input each turn. And route asynchronous work through the Batch API, which cuts the price of jobs that can tolerate delay roughly in half.

To sanity-check whether Claude is the right home for your workload at all, run the same numbers through the OpenAI cost calculator and the GPT vs Claude comparison — the winner often changes with the input/output ratio. For a full head-to-head with tier mapping and caching differences, read Claude Sonnet 5 vs GPT-5.6: a fair cost comparison.

Anthropic's current line-up has four tiers, all with a 1M-token context window, per million tokens (input / output): Claude Fable 5.1 at $10.00 / $50.00, Opus 5.5 at $4.00 / $20.00, Sonnet 5.5 at $2.00 / $10.00 and Haiku 5.5 at $0.10 / $0.50. Opus 5.5 is 20% cheaper than Opus 5 ($5.00 / $25.00), Sonnet 5.5 keeps the price of Sonnet 5 but halves cache reads to $0.10, and Haiku 5.5 costs a tenth of Haiku 4.5 ($1.00 / $5.00, 200k window). The previous models stay in the calculator so you can compare. Haiku 5.5 has one catch: prompts over 100K tokens are billed at $0.50 / $2.50.

OpenAI now matches Claude almost tier for tier: GPT-6 Astra at $10.00 / $50.00 equals Fable, GPT-6.1 Sol at $2.00 / $10.00 equals Sonnet and GPT-6 Luna at $0.10 / $0.50 equals Haiku, while the older GPT-5.6 Sol ($4.00 / $20.00) costs the same as Opus 5.5. With identical list prices, what separates the two is long prompts: Fable, Opus and Sonnet charge a flat rate across the full window, whereas GPT-6 doubles the input rate above 272K tokens and Haiku 5.5 jumps above 100K. Outside the two, Kimi K3 ($3.00 / $15.00) now costs more than Sonnet 5.5, and GLM-5.2 ($1.40 / $4.40) undercuts Sonnet but not Haiku 5.5. The tier-by-tier guide goes deeper on picking between them.

A customer support assistant. A 1,200-token system prompt plus 300 tokens of user message, answering in about 250 tokens, run 20,000 times a month — roughly 30M input and 5M output tokens. On Haiku 5.5 that is about $3 in and $2.50 out: $5.50 a month, against $55 on Haiku 4.5. The same volume on Opus 5.5 costs about $120 in and $100 out: $220, forty times more for a task Haiku usually handles fine.

A document summarizer. A 40k-token report in, a 1,500-token summary out, 200 documents a month. Sonnet 5.5: 8M input tokens ($16) and 300k output ($3), so about $19 — the same as Sonnet 5, since the list price did not change. On Haiku 5.5 the same job costs under $1, because each 40k-token prompt stays below the 100K threshold. Our PDF summary calculator models this shape in more detail.

A coding agent. These are output-heavy and iterate: 15k in, 4k out, 500 runs a month. On Sonnet 5.5 that is 7.5M input ($15) and 2M output ($20) — about $35; on Opus 5.5, $70. Cache the parts of the prompt that repeat and the input side drops by 90–95%: cache reads cost 5% of the input rate on Opus 5.5 and Sonnet 5.5, and 10% on Haiku 5.5.