Claude API pricing in 2026: Opus, Sonnet, Haiku and Fable — table and cost examples

On Kumo the Claude models are 29–49% below Anthropic's list price. For the current generation the difference is 45% for Opus 5.5, Sonnet 5.5 and Haiku 5.5 and 29% for Fable 5.2.

Haiku 5.5 has two prices at Anthropic. If a request carries more than 100K input tokens, the whole request costs $0.5 per million on input and $2.5 on output. The Kumo price list has one rate for Haiku 5.5, with no length threshold. The other models from Claude 4.6 on handle a context of up to 1M tokens at the standard price: a 900K-token request costs the same per token as a 9K one.

Claude Mythos is available only to organisations verified by Anthropic, and Kumo does not carry it. The retired Opus 4.1, Opus 4 and Sonnet 4 are left out of the table. Exact model ids for the API are in the article Claude model names in the API; the live catalogue is on the models page.

A request is billed for input tokens (the prompt, the history, files, tool definitions) and output tokens (the model's answer). The model's thinking counts as output too: Anthropic's documentation calls it part of the billed output. Output costs five times as much as input for every model in the table, so long answers and long thinking show most on the bill.

A token is a piece of a word that the model works with. Only a model's own tokenizer knows the exact count, but for an estimate there is a rule from the Kumo documentation: about 3 characters of Russian text or about 4 characters of English per token. So 1M tokens is roughly 2,222 pages of English text or 1,667 pages of Russian, at 1,800 characters a page. Anthropic adds that models from Claude 4.7 on split the same text into about 30% more tokens than earlier ones.

Prompt caching pays off when every request repeats the same beginning: a system prompt, documents, the conversation history. The first write to the cache costs more than plain input, and every later read costs many times less. Claude Code uses the cache all the time, so for it these rows matter more than the headline input price.

At Anthropic a 5-minute cache write costs 1.25 times the input price and a 1-hour write 2 times. A read usually costs 0.1 of the input price, but less on the new models: 0.05 on Opus 5.5 and Sonnet 5.5 and 0.025 on Fable 5.2. The Kumo price list has one cache-write rate and one cache-read rate per model, and those are what the table shows.

Batch processing (the Batch API) at Anthropic takes 50% off input and output for jobs that can wait: Opus 5.5 in that mode costs $2 and $10 per million. The Kumo price list has no separate batch rate.

This is an example, not a promise: real usage depends on the project, the session length and the settings. Every sum is computed from the prices in the tables above.

Assumptions for Claude Code. A day takes 10M input tokens. 85% of them, 8.5M, are cache reads: the Kumo price list calls that share typical. Another 500K are 5-minute cache writes, and the remaining 1M is plain input. Output is 300K tokens, thinking included.

Assumptions for a chat on Haiku 5.5: 1,000 dialogs a day, each with 2,000 input and 300 output tokens, no cache. A support bot, for example, whose every request stays under the 100K-token threshold.

From the Kumo wallet such a day costs 45% less than at Anthropic's list price. Over 22 working days that is about $85.9 on Sonnet 5.5 and $171.8 on Opus 5.5, against $156.2 and $312.4 at Anthropic's list price.

Packages are less simple. A package counts every token at the model's multiplier, cache reads included, while the wallet charges a cache read only a small share of the input price. So on Opus 5.5 the Pro package comes out cheaper than the wallet, but its 60M tokens last about 3 such days. On Sonnet 5.5 Pro costs more than the wallet and Ultra less. For a chat on Haiku 5.5 the wallet beats every package. Check your own usage pattern in the console before you choose.

A subscription is another way to pay: a fixed sum a month and usage limits instead of paying per token. The prices and contents of the plans below come from Anthropic's plans page as of 12 November 2026.

Subscription limits are not stated in tokens, so an exact comparison with paying per token is impossible. The monthly sums in the example above are a guide.

For readers in Russia something else matters more. Russia is not on Anthropic's list of supported regions, and subscription accounts of users from Russia are being banned. What to do if that has already happened is in the article Claude account banned. Working through the Kumo API needs no Anthropic account: requests go out with a Kumo key.

There are two ways. The wallet (pay as you go): you top up the balance, every request is charged at the rates in the table, and the balance does not expire. The minimum top-up is $3. You pay in roubles at the rate shown before payment, and the wallet keeps dollars. The payment methods depend on the checkout page: card, SBP or cryptocurrency.

A package: tokens for a term, paid in advance. A package's volume is stated on ×1 models, and a model spends it at its multiplier: Opus 5.5 has the multiplier ×1.9, so a million Opus 5.5 tokens spends 1.9M package tokens. A package payment is not refundable, and the unspent balance expires at the end of the term (documentation).

You can also try Claude without paying. Kumo has 7 free models now, Claude Opus 5.5 and Claude Fable 5.2 among them: a request to such a model costs nothing while it fits the free limit. A request that does not fit goes at the normal price. Details are in the article about free models.

Current prices of the Claude models on Kumo: