Updated July 2026. Every Claude model's input and output rate, the caching and batch discounts that actually move the bill, and a worked example so you can estimate your own Claude spend.
Anthropic bills the Claude API per token, quoted per million tokens, with input and output priced separately. This page lists the current per-million rate for every Claude model, explains the discounts, and walks through a real cost example. For an exact figure on your own prompt, use the calculator and pick your Claude model.
Current 2026 rates. Each model also has its own page with a built-in calculator — linked below.
Across the Claude line, output costs about 5× input — Opus 4.8 is $5 in / $25 out, Haiku 4.5 is $1 in / $5 out. Output is the expensive direction because the model runs a forward pass for every token it generates. Practical takeaway: the single fastest way to lower a Claude bill is to cap max_tokens and prompt for concise answers.
Anthropic's headline rates are not the whole story. Two mechanisms can cut a real bill in half or more:
Say a support-assistant call sends a 3,000-token prompt (system + context + question) and returns a 400-token answer on Opus 4.8 ($5 / $25):
Now cache the 2,500-token static system prefix. That input is billed at ~10%, so input drops to roughly $0.004 per call and the monthly bill falls toward $700 — for the same output. Switch non-urgent runs to Batch and it halves again. The calculator lets you test each of these scenarios against your own numbers.
Rule of thumb: start on Haiku, escalate only when quality demands it. Haiku 4.5 at $1 / $5 handles classification, extraction, routing and simple chat at a fraction of Opus. Reserve Opus 4.8 for genuine reasoning, and Fable 5 for autonomous coding agents where the per-run cost is small next to the engineer-hours saved. To see two models side by side on price and positioning, use the compare pages or the full model list.
It depends on the model. Claude Opus 4.8 is $5 input / $25 output per million; Haiku 4.5 is $1 / $5; Fable 5 and Opus 4.8 Fast are $10 / $50.
Claude Haiku is the cheapest per token — roughly one fifth the price of Opus — and the right default for high-volume, latency-sensitive work.
Yes. Prompt caching bills cached input at about 10% of the standard rate, and the Batch API gives a flat 50% discount for jobs that can wait up to 24 hours. They stack.
Multiply input tokens by the input rate and output tokens by the output rate, both per million, then add them. The gpt-cost.com calculator does it for any Claude model.
Prices synced from Anthropic's official pricing page, July 2026. For exact billing, use Anthropic's official token counting.