Get started
All posts

Market note7 min read

Claude API pricing: how Opus fell from $75 to $20

A million tokens of Opus output cost $75 in May 2025. Opus 5.5 launched at $20 in September. Every Claude model's price today, every change since, and what one coding session costs.

A million tokens of Opus output cost $75 in May 2025.
Opus 5.5 costs $20.

The output price of each Claude line on Anthropic's own API, at every change.

  • Opus
  • The other Claude lines
  • USD per million output tokens, log scale
$0.20$0.50$1$2$5$10$20$50$100Jul 2025Jan 2026Jul 2026Sonnet 4 $15Sonnet 5 $10SonnetHaiku 4.5 $5Haiku 5.5 $0.50HaikuFable 5 $50FableOpus 4 $75Opus 4.5 $25Opus 5.5 $20Opus
Anthropic's announcements and price list, each change footnoted belowVitna · vitna.ai
Haiku is its price for prompts up to 100,000 tokens. Each line holds its price until the next change. Newer models count about 30% more tokens for the same text, so a price per token is not quite a price per word.

Our interest. A member of the team behind Vitna holds an economic interest in Anthropic through a special-purpose vehicle and a fund, and stands to gain if Anthropic's valuation rises. We are not neutral on Anthropic or its rivals, and this page is about Anthropic's prices.

Claude, Anthropic's model, helped draft this page. The person with the interest holds no other position in a company named here.

When Anthropic launched Claude 4 in May 2025, Opus cost $15 per million input tokens and $75 per million output tokens, what the Opus before it had cost.2 Six months later Opus 4.5 cut that by two thirds, and in September 2026 Opus 5.5 arrived at $4 and $20.93

The price list grew in the same eighteen months. Fable now sits above Opus, Haiku 5.5 costs a tenth of the Haiku before it on prompts up to 100,000 tokens, and the deepest recent cuts landed on the cache.

$75Opus output per million tokens, May 2025
$20Opus 5.5 output today
$0.20Opus 5.5 cache read today, against $1.50 on Opus 4
$0.50Haiku 5.5 output today, the cheapest in the table, prompts to 100K tokens

Opus 4 from its launch announcement and Anthropic's price list21; today's prices from the record, confirmed .

What every Claude model costs today

Anthropic sells 16 Claude models through its API. Four are its newest, Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5;5 the rest are older versions it still serves, and Mythos is sold only to organizations Anthropic has verified.1 The Batch API, for work that can wait, costs half on every one of them.

ModelInputOutputCache
read
5-min
write
1-hour
write
Batch
input
Batch
output
Fable 5.1newest$10$50$0.25$12.50$20$5$25
Mythos 5.1verified organizations only$10$50$0.25$12.50$20$5$25
Fable 5still sold$10$50$1$12.50$20$5$25
Mythos 5verified organizations only$10$50$1$12.50$20$5$25
Opus 5.5newest$4$20$0.20$5$8$2$10
Opus 5still sold$5$25$0.50$6.25$10$2.50$12.50
Opus 4.8still sold$5$25$0.50$6.25$10$2.50$12.50
Opus 4.7still sold$5$25$0.50$6.25$10$2.50$12.50
Opus 4.6still sold$5$25$0.50$6.25$10$2.50$12.50
Opus 4.5still sold$5$25$0.50$6.25$10$2.50$12.50
Sonnet 5.5newest$2$10$0.10$2.50$4$1$5
Sonnet 5still sold$2$10$0.20$2.50$4$1$5
Sonnet 4.6still sold$3$15$0.30$3.75$6$1.50$7.50
Sonnet 4.5retires 30 Nov 2026$3$15$0.30$3.75$6$1.50$7.50
Haiku 5.5newest, prompts to 100K tokens$0.10$0.50$0.01$0.125$0.20$0.05$0.25
Haiku 5.5newest, prompts over 100K tokens$0.50$2.50$0.05$0.625$1$0.25$1.25
Haiku 4.5still sold$1$5$0.10$1.25$2$0.50$2.50

USD per million tokens, Anthropic's list prices as the collector read them off its price list, confirmed . A model links to its record. Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are retired; the price list still prints them.6

Every price change since Claude 4

Most of the changes were cuts. Two were not: Haiku 4.5 cost more than the Haiku it replaced, and Opus 4.7 kept its price while counting more tokens for the same text. In June 2026 a new top tier, Fable, arrived above Opus.

DateChangePrice, per million
22 May 2025Claude 4. Opus 4 at $15 in and $75 out, Sonnet 4 at $3 and $15, the prices of the Opus and Sonnet models before them. 2$75Opus 4 output
15 Oct 2025Haiku 4.5 costs more than the Haiku before it. $1 in and $5 out, against the $0.80 and $4 the price list prints for Haiku 3.5. 71$5Haiku output
24 Nov 2025Opus 4.5 cuts Opus by two thirds. $5 in and $25 out. Opus 4.1, in August, had kept Opus 4's $15 and $75. 98$25Opus output
16 Apr 2026Opus 4.7 keeps the price and changes the count. Still $5 and $25, on a new tokenizer that by Anthropic's account counts about 30% more tokens for the same text. Claude 4.7 and later models use it. 101$25Opus output, about 30% more tokens
9 Jun 2026Fable 5 opens a tier above Opus. $10 in and $50 out, alongside Mythos 5 for Project Glasswing participants. 1112$50Fable output
30 Jun 2026Sonnet 5 at $2 and $10. An introductory price, due to rise to $3 and $15 on 1 September. On 10 August Anthropic made it the standard price instead. 1311$10Sonnet output
1 Sep 2026Fable 5.1 cuts cache reads by 75%. From $1 to $0.25 a million, with input and output unchanged at $10 and $50. 1412$0.25Fable cache read
22 Sep 2026Opus 5.5 at $4 and $20. 20% under Opus 5 on input and output, and 60% under it on cache reads, at $0.20. 3$20Opus output
7 Oct 2026Haiku 5.5 at 10 cents and 50 cents. 90% under Haiku 4.5 for prompts up to 100,000 tokens and 50% under it above. The same day, Sonnet 5.5's cache reads fell from $0.20 to $0.10. 411$0.50Haiku output, prompts to 100K
30 Nov 2026Coming: Sonnet 4.5 retires. Anthropic recommends Sonnet 5.5 in its place. Fast mode on Opus 4.8 ends the same day. 61130 Novretirement

Where the newest cuts went

An agent rereads its whole conversation on every turn, and when that conversation is cached, each reread is billed as a cache read. Anthropic says cache reads make up the majority of what agentic and coding work costs.3 Since September it has cut them by 75% on Fable, 60% on Opus and half on Sonnet, deeper than any cut to those models' input or output.14311

On older Claude models a cache read costs a tenth of the input price. On Opus 5.5 and Sonnet 5.5 it is a twentieth, and on Fable 5.1 and Mythos 5.1 a fortieth.1 Anthropic puts the saving from Fable's cut at about 25% on typical work and up to about 45% on highly agentic work.14

Writing to the cache costs more than plain input: 1.25 times the input price for a cache that lasts five minutes, twice it for one that lasts an hour. The batch discount stacks on top of caching.1 Fast mode, quicker output on Opus at a premium, costs $8 in and $40 out on Opus 5.5, or $10 and $50 on Opus 5 and Opus 4.8, and is not sold through batch.1 Asking for US-only processing adds 10% to every price on Claude 4.6 and later.1

What one coding session costs

Anthropic's price list prices one job itself: an hour of coding on Opus 5.5 that reads 10,000 fresh input tokens and 40,000 cached ones and writes 15,000 output tokens.1 Here is that job priced on each line's newest model, for a thousand sessions. This is an illustration, our arithmetic on the prices above, not a measured bill.

Model1,000 sessions
Opus 4, May 2025at its launch prices$1,335.00
Opus 4, the same wordsour estimate, on its older tokenizer$1,026.92
Fable 5.1today$860.00
Opus 5.5today$348.00
Sonnet 5.5today$174.00
Haiku 5.5today, prompts to 100K tokens$8.90

Our arithmetic: fresh input, cache reads and output at each model's list price, times a thousand, rounded down to the cent. The prompt is 50,000 tokens, so a model priced by prompt length is billed in the band that holds it. Anthropic says Claude 4.7 and later count about 30% more tokens for the same text, so the second Opus 4 row divides the first by 1.3, an approximation by Anthropic's own account.101 The price list's own example adds $0.08 an hour for Claude Managed Agents, left out here because only that product charges it.

The same thousand sessions cost $348.00 on Opus 5.5, 66% less than Opus 4 would have charged for the same words, by our estimate. On Opus 5.5, output is 86% of this session's bill, which is why the figure at the top follows the output price. A long agent run rereads far more from the cache, and there the cache cuts count for more.

Our interest, again. A member of the team behind Vitna holds an economic interest in Anthropic through a special-purpose vehicle and a fund, and stands to gain if its valuation rises. We are not neutral on Anthropic or its rivals.

Claude, Anthropic's model, helped draft this page, and the person with the interest holds no other position in a company named here. Nothing here is investment advice.

Sources

  1. 1Price listPricingAnthropicread
  2. 2ReleaseIntroducing Claude 4Anthropicread
  3. 3ReleaseIntroducing Claude Opus 5.5Anthropicread
  4. 4ReleaseIntroducing Claude Haiku 5.5Anthropicread
  5. 5DocsModels overviewAnthropicread
  6. 6DocsModel deprecationsAnthropicread
  7. 7ReleaseIntroducing Claude Haiku 4.5Anthropicread
  8. 8ReleaseClaude Opus 4.1Anthropicread
  9. 9ReleaseIntroducing Claude Opus 4.5Anthropicread
  10. 10ReleaseIntroducing Claude Opus 4.7Anthropicread
  11. 11ChangelogClaude Platform release notesAnthropicread
  12. 12ReleaseIntroducing Claude Fable 5.1 and Claude Mythos 5.1Anthropicread
  13. 13ReleaseIntroducing Claude Sonnet 5Anthropicread
  14. 14PageClaude FableAnthropicread

More posts

  1. Technical note3 min read

    Why a Google SERP API now returns five results, not ten

    Google now unlocks each result link a little later than the one above it, the tenth about 24 seconds after the page. Five results arrive before the long waits: 2,273 of 2,500 live searches answered, 12 at a time, at a median of 4.19 seconds.

  2. Ownership guide10 min read

    Who owns the AI companies, and who still holds the votes

    Elon Musk owns under half of SpaceX, home of xAI, and casts 82% of its votes. Microsoft owns about a quarter of OpenAI and names none of its directors. How the money and the votes came apart at 13 AI labs.

  3. Technical note8 min read

    How Beam 1.0 got past its tokens-per-second ceiling

    Beam 1.0 runs at 348.8 tokens a second across its hosts and 537.3 on the fastest. One token a pass stops near 300. What sets that ceiling, and the ways past it.

  4. Market note6 min read

    How DeepSeek went from price war to price rise

    DeepSeek's cheapest model costs $1.20 per million output tokens at peak, up from $0.28 in April. Every seller's price today, and every release's price and score.

  5. Company story6 min read

    How a stock trader came to own DeepSeek

    Liang Wenfeng still controls DeepSeek. His hedge fund paid for it for three years, and the outside investors who bought in this June hold under 9%.