Claude Opus 5.5 costs 20 percent less than Opus 5, and Anthropic says it matches the pricier Fable 5.1 on most work

Anthropic's new Opus is $4 in and $20 out per million tokens, with cache reads down 60 percent. Its own price list keeps Fable 5.1 at $10 and $50, and Anthropic says Opus 5.5 does most work as well. Two outside groups have published their own tests; the 40 percent saving is Anthropic's own figure.

23 September 2026 · 4 min read · L4 Models · The Pulse, 22 September 2026

In this story: Anthropic

What happened

On 22 September Anthropic released Claude Opus 5.5 and priced it below the model it follows. Text sent to the model costs $4 per million tokens and text it writes back costs $20, against $5 and $25 for Opus 5, a 20 percent cut. A token is a small piece of text, a word or part of one. Cache reads, the rate charged when the model reuses text it has already processed, such as a long document or a codebase an agent keeps returning to, fall from 50 cents to 20 cents per million tokens, 60 percent less. Anthropic says cache reads make up most of the cost of agent and coding work, and that because the model also uses fewer tokens per task, it costs 40 percent less than Opus 5 to run on typical workloads. Anthropic's price list for developers shows every Opus from Opus 4.5 to Opus 5 at $5 and $25, so Opus 5.5 is the first of them to cost less.

Why it matters

The sharper comparison is inside Anthropic's own range. The most expensive models on its price list, Claude Fable 5.1 and Claude Mythos 5.1, stay at $10 in and $50 out, and Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work. Where that holds, a job can move to a model that costs 40 percent of Fable 5.1's price per token. Anthropic also says that on FrontierCode, a coding test, Opus 5.5 at its default setting beat the top score of OpenAI's GPT-6 Astra at about a fifth of the cost per task; that is Anthropic's own run of the test.

A worked example shows why the cache rate matters for agents. An agent that rereads 10 million tokens of cached text pays $2 for it on Opus 5.5, against $5 on Opus 5, before any saving on the base price. OpenAI moved on price the same day, at the cheaper end of its range: its new GPT-6 Sol costs $2 in and $10 out, and GPT-6 Luna 10 cents and 50 cents, about half what OpenAI says the GPT-5.6 versions cost at their promotional prices. Its top model, GPT-6 Astra, is still listed at $10 and $50.

What would change it

Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks. Their prices will show whether the cut runs down the whole range or stops at Opus. The other number to watch is Fable 5.1's own price, now that Anthropic sells a model it describes as doing most of the same work for 40 percent of the price per token.

What we do not know

The 40 percent saving and the claim that Opus 5.5 matches Fable 5.1 on most work come from Anthropic's own tests. Of the outside evaluators Anthropic says tested the model before release, METR, a nonprofit that tests AI systems, has published a summary. It had ten business days of access and concluded that Opus 5.5 is likely a modest improvement on Fable 5.1 at AI research work, not a huge leap. METR says Anthropic could review and edit that summary before it was published, that its work was not meant to verify compliance with any threshold in Anthropic's own policies, and that it did not assess the model's alignment.

Artificial Analysis, which runs its own tests of AI models, ranks Opus 5.5 first of the 212 models in its price class on its intelligence index, tested at the model's highest effort setting. It also calls the price somewhat expensive against that class's medians of $2 in and $10 out, and at that setting the model used 260 million tokens on the index against a median of 88 million. What a job costs is the price per token times the number of tokens the model spends, and that number depends on the setting and the work.

What this changes for you

If you pay for an AI coding tool, a support bot or any app that runs on Claude, the company behind it now pays less for Opus, and whether it passes that on is its choice. If you use Claude directly on a Pro, Max, Team or seat-based Enterprise plan, Anthropic says it is raising the five-hour usage limits and giving subscribers a rate-limit reset they can save and use when they choose. For a subscriber, that is the change you will notice first.

Sources

Everything above is written from these. Each line says what that document proves.

1. anthropic.com: Anthropic's own announcement, 22 September 2026: $4 and $20 against $5 and $25 for Opus 5, cache reads at 20 cents against 50 cents, the 40 percent saving on typical workloads from its own tests, Fable 5.1-level performance on most work, its own FrontierCode run, the higher five-hour limits, and Sonnet 5.5 and Haiku 5.5 to follow.
2. platform.claude.com: Anthropic's price list for developers, read 23 September 2026: Fable 5.1 and Mythos 5.1 at $10 and $50, and every Opus from Opus 4.5 to Opus 5 at $5 and $25.
3. metr.org: METR's summary of its pre-release evaluation, 22 September 2026: ten business days of access, a likely modest improvement on Fable 5.1 at AI research work, and its notes that Anthropic could review and edit the text and that the work did not verify policy compliance or assess alignment.
4. artificialanalysis.ai: Artificial Analysis's page for the model, read 23 September 2026: first of 212 models in its price class on its intelligence index at the highest effort setting, $4 and $20 confirmed, prices called somewhat expensive against class medians of $2 and $10, and 260 million tokens used against a median of 88 million.
5. openai.com: OpenAI's own announcement, 22 September 2026: GPT-6 Sol at $2 and $10 and GPT-6 Luna at 10 cents and 50 cents, against GPT-5.6 promotional prices of $4 and $20, and 20 cents and $1.20.
6. openai.com: OpenAI's GPT-6 Astra page, updated 22 September 2026: $10 per million input tokens and $50 per million output tokens.

Keep reading