How much does the Claude API cost?

Anthropic prices the Claude API per model, per million tokens, split into an input rate and a higher output rate. There is no seat fee and no minimum spend: you pay for the tokens a request actually uses, billed in arrears against a prepaid or invoiced balance. The confusing part isn't the pricing model, it's that Anthropic ships several tiers at once, each with a different price, and the tier names alone (Haiku, Sonnet, Opus) don't tell you the multiple between them.

The three tiers, and what separates them

Haiku is the fast, cheap tier: classification, extraction, short-turn chat, anything latency-sensitive where the task doesn't need deep reasoning. Sonnet is the general-purpose default most production apps build on — strong enough for agentic tool use and long documents, priced well under Opus. Opus is the reasoning-first tier: harder multi-step problems, longer autonomous runs, code changes across a large repo. The gap between Haiku and Opus input pricing is roughly an order of magnitude, and the gap compounds on output tokens, which are always priced higher than input.

Output tokens are where the bill actually grows

Every Claude model's output rate is a fixed multiple of its input rate. That's deliberate: generation is the expensive half of a request, and it's also the half developers underestimate, because a short prompt can still produce a long completion. If your application does anything agentic — multi-step tool calls, code generation, extended thinking — output tokens will dominate the bill even though the prompt looks small in your logs. Before shipping, run your actual prompt-and-response shapes through the calculator rather than eyeballing the per-token rate.

Cached input tokens cut the input side sharply

Anthropic's prompt caching reuses the same system prompt, tool definitions, or long context block across requests at a fraction of the standard input rate, with a small one-time write cost the first time a block is cached. Any workload that repeats a large static prefix — a coding agent re-sending the same file tree, a support bot re-sending the same knowledge base excerpt — should be caching that prefix. It is usually the single largest lever on a Claude bill, larger than picking a cheaper model tier.

Batch processing halves the rate, at the cost of latency

The Batch API trades synchronous response time for a flat discount on both input and output tokens, applied uniformly across the current model lineup. It's built for the workloads that don't need a live answer: nightly re-classification of a support queue, bulk document summarization, offline evals. If a job can tolerate finishing within hours instead of seconds, routing it through batch is close to free money.

Estimating a real bill

• Separate your workload into input tokens (prompt, context, tool schemas) and output tokens (the completion) — they're priced differently and output is usually the bigger line item.
• Identify anything that repeats across requests (system prompts, tool definitions, RAG context) — that's your caching candidate.
• Identify anything that can run asynchronously — that's your batch candidate.
• Multiply by requests per day and 30.44 days for a monthly estimate, not requests per month divided evenly.

Meterlark's calculator does this math for you, including cached and batch rates per model — no spreadsheet required.

Choosing between Sonnet and Opus specifically

This is the decision most teams actually agonize over, since Haiku is an obvious fit for simple tasks and the Sonnet-versus-Opus line is blurrier. As a starting heuristic: if the task has a single clear right answer and a bounded number of steps, Sonnet is almost always sufficient and meaningfully cheaper. Reach for Opus when the task is genuinely open-ended — ambiguous requirements, long autonomous runs, code changes that touch many interdependent files — where the extra reasoning depth measurably improves the result rather than just adding cost.

Why the number on this page might differ from what you remember

Anthropic revises pricing when it ships new model generations, and older tiers get marked legacy or deprecated rather than silently repriced. That's exactly why this page renders live from a dataset with a verified-as-of date instead of a number written once and left to rot — see why AI API prices keep changing for the pattern across every provider, not just Anthropic.