On this page 6 sections
Claude Fable 5.2 pricing is $8 per million input tokens and $40 per million output tokens, twice Claude Opus 5.5 and 20% below Fable 5.1. Anthropic released it on November 9, 2026. Keep most product features on Opus 5.5. Pay for Fable on the long agent tasks Opus fails, or call it as an advisor.
Most of the launch coverage is about benchmark scores. For a founder, the useful question is narrower: which of your product’s AI features, if any, should move to a model that costs twice as much per token? The numbers below come from Anthropic’s own pages as of November 10, 2026. Benchmark and speed claims are Anthropic’s. Treat the cost table below as a method to run on your own traffic.
Claude Fable 5.2 pricing next to Opus 5.5
Anthropic’s pricing page lists these rates per million tokens:
| Model | Input | Output | Cache read | 5-min cache write | Batch (in / out) |
|---|---|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 | $10 | $4 / $20 |
| Claude Fable 5.1 | $10 | $50 | $0.25 | $12.50 | $5 / $25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 | $5 | $2 / $10 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | $2.50 | $1 / $5 |
Two details matter more than the headline. First, cache reads cost the same $0.20 on Fable 5.2 and Opus 5.5, because Fable bills them at 2.5% of input and Opus at 5%. An agent that rereads a large cached context every turn pays the same for that context on either model. The 2x gap lands on new input and, above all, on output, which includes thinking tokens.
Second, fast mode doubles everything again. Fable 5.2 in fast mode costs $16 input and $80 output, and Anthropic says it generates tokens up to 2.5 times faster. It isn’t available on the Batch API. Use it only where a user is waiting on the answer.
Anthropic’s own advice is to not start here. The models overview says to begin with Opus 5.5 for most workloads and use Fable 5.2 “for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.” The launch post backs that with numbers: on FrontierCode 2.0, Anthropic reports 49.1% for Fable 5.2 at xhigh effort against 45.5% for Opus 5.5. It also says Fable 5.2 at high effort scores about level with Opus 5.5’s best on that test “at roughly half the cost per task.” That’s a vendor claim on a vendor-picked benchmark. It’s also the claim worth testing, because cost per finished task is the number that matters.
What one agent task costs on each model
Per-token prices don’t map onto a product, so take one agent task. Say your product runs a research or analysis agent that takes about 40 turns per job. Across the job it reads 1.1 million tokens of cached context, writes 100,000 new tokens to the cache, and produces 60,000 output tokens, thinking included. Assume, for the comparison, that every model spends the same tokens.
| Setup | Cost per task | Per 1,000 tasks |
|---|---|---|
| Claude Sonnet 5.5 | $0.96 | $960 |
| Claude Opus 5.5 | $1.92 | $1,920 |
| Claude Fable 5.2 | $3.62 | $3,620 |
| Claude Fable 5.2, fast mode | $7.24 | $7,240 |
| Sonnet 5.5 executor + 3 Fable 5.2 advisor calls | about $1.87 | about $1,870 |
Output is $2.40 of Fable’s $3.62, so the lever is how much the model writes and thinks, not how much context it reads. These are list prices on an assumed token mix. Your own logs will move them, in either direction.
The table hides the case where Fable wins. If some of these jobs fail on Opus 5.5 and need a retry or a person to clean up, the cost of a finished job on Opus is higher than $1.92. If Fable finishes more of them in fewer turns, as Anthropic claims, the gap narrows or flips. You can only learn that from a side-by-side eval on a few hundred of your own jobs. The same per-task method is in our Haiku 5.5 cost breakdown, at the other end of the price range.
Effort is the second lever. Fable 5.2 defaults to high effort, where Opus 5.5 defaults to medium. Anthropic’s effort docs say most of Fable’s gain is already present at high, to step up to xhigh only where you’ve measured a gain, and to reserve max for the rare task “where time and tokens don’t matter”. A request that omits effort runs one level higher on Fable than on Opus, and the bill shows it before anything else does.
A Fable advisor costs less than switching
There’s a cheaper way to buy Fable’s judgment than moving a whole feature onto it. The advisor tool, in beta, lets a cheaper executor model consult a stronger advisor mid-task. The advisor reads the conversation, returns a plan or a course correction, and the executor carries on writing the output. Fable 5.2 can advise every executor model, including Sonnet 5.5 and Haiku 5.5.
The economics come from who writes the tokens. Anthropic says advisor output is typically 1,400 to 1,800 tokens including thinking, while the executor writes everything else at its own rate. The advisor also reads the conversation at Fable’s input rate on every call, so long conversations make each call more expensive. Advisor caching exists, and Anthropic says it breaks even at roughly three advisor calls. The docs also say that for teams on Sonnet, an Opus advisor “keeps total cost similar or lower” while a Fable 5.2 advisor “maximizes the quality lift”.
Three limits to know before you build on it:
- It’s beta and not on every platform. It runs on the Claude API, Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed). It isn’t available on Amazon Bedrock.
- Fable’s advice comes back encrypted. The executor can use it, but your code can’t read or log the advice text. If you need to show or audit the advice, the docs point to Opus 4.8 as an advisor that returns plain text.
- It fits long, multistep work. Anthropic lists single-turn Q&A as a weak fit. A chatbot answering one question gains little from it.
For most products this is the first Fable experiment to run: keep the agent on Sonnet 5.5 or Opus 5.5, add a Fable 5.2 advisor with max_uses capped, and compare finished-task rates. Our agent vs workflow guide covers whether the feature needs an agent loop at all, which is the cheaper question to answer first.
Seven things that break moving from Opus 5.5
Fable 5.2 uses the same Messages API as Opus 5.5, and both already reject forced tool choice and disabled thinking. The migration guide still lists differences that turn a model ID change into a failed release:
- 30-day data retention is required. Fable 5.2 isn’t available under zero data retention unless Anthropic expressly authorizes it. If your workspace runs Opus 5.5 under ZDR, every Fable request returns a 400 error. For health or finance products, settle this with Anthropic first.
- The prefix check applies to every account. If your code edits the system prompt, tools or earlier messages between requests, Fable returns a 400 error with the message “The block is bound to a different conversation”. Opus 5.5 only enforces this for accounts created on or after August 31, 2026, so older accounts can pass on Opus and fail on Fable.
- Default effort is higher. Set
effortexplicitly on every request, then rerun your effort sweep. - Rate limits are separate and lower. Fable 5.2 has its own rate limit, with lower token limits at each tier than Opus 5.5. Check your tier before you move production traffic.
- Opus 5.5 can’t read Fable 5.2’s thinking. If a router or fallback moves a conversation from Fable to Opus, the turns after the switch run without Fable’s reasoning, and the first of them costs more. Moving from Opus onto Fable keeps the reasoning.
- Refusals need a code path. Fable 5.2 runs safety classifiers that return
stop_reason: "refusal". Server-side fallback is in beta and retries only on Opus 4.8 or Opus 5. Handle every refusal category, including ones Opus never returned for you. - Customer-managed keys lose structured outputs. In an organization with customer-managed encryption keys, Fable 5.2 doesn’t support strict tool use or JSON outputs, while Opus 5.5 does. Validate the output in your own code there.
None of this forces a move. Anthropic’s deprecations page lists Opus 5.5 as active with retirement not sooner than September 22, 2027, and Fable 5.1 as active until at least September 1, 2027. If you’re already on Fable 5.1, the what’s new page says the move is mostly drop-in, with four breaking changes, and it saves 20%. That switch is worth doing.
Fable 5.2 decision guide for product teams
- You’re on Fable 5.1. Move to Fable 5.2 for the 20% saving. Run the prefix check test first and update any advisor or computer use settings.
- Your feature passes your evals on Opus 5.5. Stay. Fable would double the per-token bill for nothing you can measure.
- A long agent task fails on Opus 5.5 even at
highorxhigheffort. Add a Fable 5.2 advisor first. Move the executor to Fable only if the advisor doesn’t close the gap. - The task is classification, extraction, routing or short replies. Fable is the wrong tier. Price it on Haiku 5.5 or Sonnet 5.5, as in our guide to choosing a model by task type.
- Your workspace runs under zero data retention. Don’t plan on Fable until Anthropic confirms your eligibility.
- Users wait on the answer and Fable passes your evals. Price fast mode at $16 and $80 against what the speed is worth to your conversion or retention.
How Better Analyst runs multi-step agents for 1.5M+ users
The model is one line in a request. The architecture around it decides whether you can test Fable on a single feature this week. On Better Analyst (formerly FormulaBot), we designed the analysis agents with LangChain agents, LangGraph workflows and GPT-4o: they reason through multi-step analysis, transform the data and recover when a step fails, while Python runs in E2B sandboxes. We also split AI inference, data processing and the user-facing API into independently deployable services, which let the platform scale from thousands to 1.5M+ users.
That layout is what makes a model decision cheap. Agent steps that recover from failure give you a finished-task rate to measure, which is the number that settles Fable against Opus. Inference in its own service means a model ID, an effort setting, an advisor tool and a refusal path change in one place, with one deploy and one rollback. If your product’s model calls are spread across the codebase, an AI audit maps each one to the cheapest model that passes your evals. Our AI services overview covers the agent and LLM work we take on from there.
Frequently asked questions
- How much does Claude Fable 5.2 cost per million tokens?
- Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens on the Claude API. Cache reads are $0.20 per million, 5-minute cache writes are $10 and 1-hour writes are $16. The Batch API halves the base rates to $4 and $20, and fast mode costs $16 input and $80 output. Mythos 5.2 has the same prices but is limited to verified organizations.
- Is Claude Fable 5.2 cheaper than Fable 5.1?
- Yes. Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, so Fable 5.2 is 20% cheaper on both. Cache reads drop from $0.25 to $0.20 per million. Anthropic also says Fable 5.2 generates output up to 30% faster than Fable 5.1, and moving from 5.1 is mostly a model ID change plus four breaking changes listed in the migration guide.
- Should I switch my product from Claude Opus 5.5 to Fable 5.2?
- Not by default. Anthropic's own docs say to start most workloads on Opus 5.5 and move to Fable 5.2 only when your evals on Opus 5.5 at higher effort still fall short. Fable costs twice as much per input and output token, defaults to a higher effort level, has lower rate limits and requires 30-day data retention. Switch only the specific long agent tasks that fail.
- What is the Claude advisor tool and can Fable 5.2 be the advisor?
- The advisor tool is a beta API feature where a cheaper executor model, such as Sonnet 5.5 or Haiku 5.5, consults a stronger advisor model mid-task for a plan or a course correction. Fable 5.2 can advise every executor model. Advisor calls bill at the advisor's rates, but they typically produce 1,400 to 1,800 tokens including thinking, so most output stays at the cheaper executor's rates.
- Does Claude Fable 5.2 work under zero data retention?
- No, not by default. Fable 5.2 requires 30-day data retention and isn't available under a zero data retention arrangement unless Anthropic expressly authorizes it. On the Claude API, a request from an organization or workspace without 30-day retention returns a 400 error. If your product runs Opus 5.5 under zero data retention for compliance reasons, sort this out with Anthropic before any switch.
References
- Anthropic - Introducing Claude Fable 5.2 and Claude Mythos 5.2
- Anthropic Docs - Pricing (model, cache, batch and fast mode pricing)
- Anthropic Docs - Models overview
- Anthropic Docs - What's new in Claude Fable 5.2
- Anthropic Docs - Claude Fable 5.2 and Claude Mythos 5.2 migration guide
- Anthropic Docs - Advisor tool
- Anthropic Docs - Effort
- Anthropic Docs - Model deprecations