Fable 5.2 vs Opus 5.5: double the price, so when is it worth it?

Fable 5.2 is the pricier model on every line of Anthropic’s rate card except one. If you build with Claude through the API, here is how the two compare at standard rates:

• Input: Fable 5.2 costs $8 per million tokens. Opus 5.5 costs $4.
• Output: Fable 5.2 costs $40 per million tokens. Opus 5.5 costs $20.
• Writing to the prompt cache: $10 per million tokens for Fable 5.2, $5 for Opus 5.5.
• Reading from the prompt cache: 20 cents per million tokens on both. This is the one line where they match.

A token is a small chunk of text, roughly a short word or part of a word. Input is what you send the model; output is what it writes back. Prompt caching lets the API reuse a big block of text you send again and again, such as a long document, at a fraction of the normal input price.

Both models offer an optional fast mode. Anthropic says it generates tokens up to 2.5 times quicker and costs double the standard rate. For Fable 5.2 that means $16 for input and $80 for output per million tokens.

Fable did get cheaper this week. Fable 5.1, released in September, cost $10 and $50, so the new version cuts both prices by 20%. It is still the most expensive current model Anthropic sells, and Fable accounted for 11% of spending on Anthropic’s models in August, according to Ramp’s AI Index.

Fable 5.2 wins on long, open-ended technical work. Anthropic’s benchmark results show both 5.2 models ahead of Opus 5.5 in multidisciplinary reasoning, agentic coding, computer use, knowledge work, cybersecurity and agentic scientific research.

The clearest head-to-head example is a math project. A team of 1,000 Fable 5.2 agents turned the proof of Fermat’s Last Theorem into a form a computer can check line by line, and finished in 30 hours. Using the same setup, Opus 5.5 got through 26% of that work in 18 hours. Anthropic’s first attempt at the same job, with an unreleased research model, took 11 days.

Anthropic’s other claims for Fable 5.2 are against its own predecessor rather than Opus:

• Coding: 49.1% on the FrontierCode 2.0 benchmark, up from 32.4% for Fable 5.1.
• Science tasks: 74.6%, up from 52.6%.
• Robot control: 64.9% on the RoboDojo test, more than double the 27.6% that Fable 5.1 managed.
• Cyber operations: 76.3% of challenges completed on Irregular’s CyScenarioBench, versus 61.7% for Fable 5.1 and 61% for OpenAI’s GPT-6 Cyber.

Anthropic also says agents running an early build of Fable 5.2 moved a key production service from Python to Rust, and the rewritten service now starts about 26 times faster. One engineer had the model coordinate more than 1,000 agents for eight days to build an operating system that boots on a laptop. The company adds that Fable 5.2 solved two open problems in black hole physics and designed three working electronic gadgets from a single paragraph of instructions.

Speed improved too. Anthropic says the 5.2 models produce answers up to 30% faster than the 5.1 versions, and that Fable 5.2 uses fewer tokens than Fable 5.1 on hard tasks at medium effort.

Anthropic calls Opus 5.5 its strongest daily driver, and recommends it as the default for most coding and knowledge tasks. Its advice is to start there and move up to Fable or Mythos for the most ambitious technical work, or down to Sonnet 5.5 or Haiku 5.5 when a job is simpler.

That guidance matters because the price gap compounds. Every request sent to Fable 5.2 instead of Opus 5.5 costs twice as much for the same number of tokens. For a quick bug fix, a summary or a draft email, Anthropic’s own framing suggests Fable is more model than the job needs.

Here is a simple way to read Anthropic’s lineup:

• Haiku 5.5: the fastest and cheapest model.
• Sonnet 5.5: a mid-priced model for coding and agents.
• Opus 5.5: the everyday pick for agentic coding and business work.
• Fable 5.2: the top public model, built for long-running agents and hard, open-ended problems.
• Mythos 5.2: the same underlying model as Fable 5.2 with looser cyber and biology safeguards, for approved organizations only.

You don’t need the API to try Fable 5.2. It is available to all Claude users and developers, and on Amazon Web Services, Google Cloud and Microsoft Azure.

On subscriptions, Anthropic’s plan page lists Fable on the Pro plan as paid with usage credits. The $100 Max plan includes Fable for up to half of your weekly limit. For launch week, Anthropic is also handing out extras: eligible paid subscribers get one to three Fable resets depending on how long they have subscribed, and Pro plans and standard Team seats can use Fable for up to 50% of their weekly usage limits.

Anthropic notes that how fast you hit a limit depends on which model you use, how long your chats are and how big your files are. Because Fable is the most expensive model per token, it will eat into that allowance faster than Opus.

Fable 5.2 says no less often than before. In early testing, Anthropic counted about 55% fewer blocked harmless requests than with Fable 5.1. When Fable does block a cyber question, Opus 4.8 answers it instead; blocked biology questions go to Opus 5.

Mythos 5.2 is the unrestricted version, and you probably can’t use it. Access requires approval through Anthropic’s Cyber Verification or Life Sciences programs, or Claude Security on Enterprise plans. Anthropic calls it its most capable model for cybersecurity engineering, and it scored 89.6% on the company’s CVE Rediscovery test of finding known flaws in source code, up from 78.3% for Mythos 5.1.

On safety, Anthropic says the new models matched or beat Opus 5.5 on most alignment, honesty and misuse measures. Its system card says the model was about an order of magnitude less prone than Opus 5.5 to post harmful content to public package registries unprompted, though it still did so when a simulated user pushed it. The UK AI Security Institute, METR and Frontier Design tested the model before release.

For most people, the answer is the one Anthropic gives: start with Opus 5.5. It costs half as much per token, and the company names it the default for everyday coding and knowledge work.

Fable 5.2 earns its price on the kind of work where Anthropic’s results come from: multi-day agent projects, big code migrations, research problems and tasks with fuzzy instructions. Testers told Anthropic that Fable 5.2 handles ambiguous requests well and communicates more clearly than earlier Fable models. If your work looks like that, the launch-week resets are a cheap way to test it on your own tasks before paying full rates.