Claude API pricing: $0.35 to $35 per 1,000 messages

How much does the Claude API cost per model?

The Claude API costs $0.10 to $10 per million input tokens and $0.50 to $50 per million output tokens, depending on the model. We read every row below from Anthropic's pricing docs on 8 November 2026. Prices are US dollars per million tokens, on the default global route. The list price isn't the whole bill: Claude 4.7 and later models count about 30% more tokens for the same text, so Opus 5.5 ends up about 4% dearer per job than Opus 4.6.

Output costs five times as much as input on every model in the table. Claude Mythos 5.1 and Mythos 5 carry the same prices as Fable, but only the organizations that pass one of Anthropic's verification programs can use them. Opus 4, Opus 4.1 and Sonnet 4 are retired on the Claude API. Sonnet 5's low price is now permanent: Anthropic's footnote says "The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price."

Haiku 5.5 is the only model that's priced by the length of the prompt. If a request goes over 100,000 tokens, it pays five times more on every line, even when part of it is a cache hit. Every other model from Claude 4.6 on bills all of its 1 million token window at the same rate. Anthropic's example: "A 900k-token request is billed at the same per-token rate as a 9k-token request."

What do 1,000 Claude API messages cost?

1,000 short Claude messages cost $0.35 on Haiku 5.5, $7 on Sonnet 5.5, $14 on Opus 5.5 and $35 on Fable 5.1. We priced a short message at 1,000 tokens in and 500 out, at the list prices above:

Here's how we did the math for Sonnet 5.5. 1,000 input tokens at $2 per million is $0.002. 500 output tokens at $10 per million is $0.005. That's $0.007 a message, or $7 per 1,000. A longer job scales the same way, so we priced it in three of the shapes that you're most likely to send:

A chat also sends its whole history again on each turn, so turn ten of a thread carries the nine before it as input. That's why a real chat bill runs above the short column. The AI model cost index prices the same standard answer across 100 models from every provider.

Is Opus 5.5 cheaper than Opus 4.6 per job?

No. Opus 5.5 costs about 4% more than Opus 4.6 for the same text, even though its list price is 20% lower. Anthropic's docs explain why: "This tokenizer produces approximately 30% more tokens for the same text." The newer tokenizer covers Claude 4.7 and later, and Sonnet 4.6 and earlier use the old one. We took a short message that measures 1,000 tokens in and 500 out on the old tokenizer, added 30% for the new one, and priced 1,000 of them:

So the price cut on Sonnet 5.5 is real, but it's smaller than it looks. Haiku 5.5 is the only one that drops by a lot. Anthropic says the exact increase "depends on the content and workload shape," so count your own prompts with its token counting endpoint before you budget a migration. This table is about cost. It doesn't say that the models are as good as each other at the job.

How much do prompt caching and the Batch API save?

Prompt caching and the Batch API together cut the cost of a repeated Claude job by about 80%. A cache hit costs 10% of the input price on most models and 5% on Opus 5.5 and Sonnet 5.5. Batch halves input and output. Anthropic's docs say the cache multipliers "stack with other pricing modifiers, including the Batch API discount and data residency."

Here's a worked example. You classify 1,000 support tickets on Sonnet 5.5 with the same 5,000-token instructions, 500 tokens of ticket and 500 tokens out:

The same job on Opus 5.5 drops from $32 to $13 with caching, and to $6.50 with both. Two costs sit outside the table. The first request writes the cache at 1.25x the input price, or $0.0125 for those 5,000 tokens on Sonnet 5.5. A 5-minute cache also expires when nobody uses it, and the 1-hour cache costs 2x to write. Batch results aren't instant. Keep it for work that can wait. Anthropic's Batch processing docs cover the limits.

What else adds to a Claude API bill?

Web search is the add-on that most teams notice first: it's $10 per 1,000 searches, plus the tokens that the results add. Most of the other extras are just tokens in disguise. Here's every charge outside the per-model table, read on 8 November 2026:

Fast mode doubles the price of Opus 5.5, and you can't use it with Batch. US-only inference is the one that's easy to miss: it adds a flat 10% to all of the input, output and cache lines. Anthropic gives new accounts "a small amount of free credits to test the API," and bills the rest monthly in US dollars.

Is the Claude API cheaper than Claude Pro?

The Claude API is cheaper than Claude Pro for light use: $20 buys about 2,857 short Sonnet 5.5 messages at $0.007 each, or 95 a day. Pro costs $20 a month, or $17 a month on a $200 yearly plan, per Anthropic's pricing page read on 8 November 2026.

Pro includes no API credit. Max at $100 and Ultra at $200 include $100 and $200 a month of it. Long threads, Opus and Claude Code all move the line, so our Claude API vs subscription breakdown works it out plan by plan.

What does the Claude API cost next to the rest of your AI stack?

A small Claude-powered app plus two chat subscriptions runs about $98.50 a month. Take a side project that sends 300 Sonnet 5.5 requests a day, shaped like the cached ticket job above, at $0.0065 each. Over 30 days, that's 9,000 requests and $58.50. Most developers also keep Claude Pro at $20 and ChatGPT Plus at $20, per OpenAI's plan page read the same day:

Whizi, which publishes this page, can stand in for the two chat plans, not for the API. Whizi Pro costs $29.99 a month for 2,500 credits and puts Claude Opus, Claude Sonnet, ChatGPT and Gemini in one chat, with credits per message varying by model. That's less than $40 if you'd otherwise keep both of the chat plans. It's more than Claude Pro on its own. Whizi has no API, no Claude Code and no coding agent, so keep your Anthropic key for the app.