Claude Fable 5.2 for business: cost and when to use it
Fable 5.2 costs twice what Opus 5.5 does per token. Keep Opus 5.5 for everyday work and put Fable on long, multi-step agent jobs, ideally directing cheaper models.
Claude Fable 5.2 is Anthropic's newest model, released on 9 November 2026. Anthropic calls it its most capable model for technical work and its strongest lead agent: the model you put in charge of long, multi-step jobs that run for hours and end in a finished document, spreadsheet or piece of code.
On the API it costs US$8 per million input tokens and US$40 per million output tokens, down from US$10 and US$50 on Fable 5.1. That is twice the price of Opus 5.5 (US$4 and US$20). Batch jobs are half price, and a fast mode costs US$16 and US$80. On a Claude plan, Fable 5.2 is available to Pro, Max, Ultra, Team and Enterprise users.
Should you switch? For most business work, no. Anthropic's own advice is to start everyday coding and knowledge work on Opus 5.5 and move up to Fable only for ambitious, open-ended work. In a business that means long agent runs: a tender response, a quote built from a dozen documents, a month-end reconciliation, or a clean-up across thousands of CRM records.
What is Claude Fable 5.2?
Fable 5.2 sits above Opus in Anthropic's current lineup. Its model ID on the Claude API is claude-fable-5-2. It reads up to 1 million tokens at once, writes up to 128,000 in a single reply, and its thinking is always on, with an effort setting that controls how hard it works. On the API the default is high.
Anthropic released a second model on the same day. Mythos 5.2 is the same underlying model with fewer safety blocks on cybersecurity and biology work, and it is only open to organisations verified through Anthropic's programs. For a normal business, Fable is the one you can use.
Anthropic's published gaps over Opus 5.5 are widest on agent work. On AutomationBench, which tests automated business tasks, Fable 5.2 scores 51.4% against 42.5%. On the Remote Labor Index it scores 41.3% at high effort against 21.3%. On FrontierSWE v2, a software engineering test, it is 69.9% against 62.3%. Anthropic also says Fable 5.2 writes up to 30% faster than Fable 5.1 and, in early testing, trips about 55% fewer safety blocks. These are Anthropic's own numbers, so test on your own work.
Scores published by Anthropic at the 9 November 2026 launch. Anthropic's own tests, not ours.
What does Fable 5.2 cost next to Opus, Sonnet and Haiku?
Prices below are Anthropic's API list prices per million tokens, in US dollars.
| Model | Input, US$ | Output, US$ | Batch, US$ in / out | Best for |
|---|---|---|---|---|
| Fable 5.2 | $8 | $40 | $4 / $20 | Long agent runs with many steps, hard code, the lead agent that directs others |
| Opus 5.5 | $4 | $20 | $2 / $10 | Most hard business work: tenders, analysis, writing code. Anthropic's recommended starting point |
| Sonnet 5.5 | $2 | $10 | $1 / $5 | Weekly judgement work, research, most business writing and extraction |
| Haiku 5.5 | $0.10 | $0.50 | $0.05 / $0.25 | High-volume sorting, summarising and record updates. Prices rise five times for prompts over 100,000 tokens |
Two details matter more than the headline. First, cached input on Fable 5.2 costs US$0.20 per million tokens, the same as Opus 5.5. An agent re-reads its own instructions and earlier steps on every turn, and most of that re-reading comes from the cache, so the gap on a long run is smaller than the list prices suggest. Second, fast mode doubles the price for output that arrives up to 2.5 times faster. Pay that only when a person is waiting on the answer.
On a Claude plan you pay by the seat. Our Claude pricing guide has every plan in Australian dollars. At launch, Anthropic is including Fable for up to 50% of weekly usage limits on Pro and standard Team seats, and calls that a launch promotion, so do not plan a year around it.
Which business jobs are worth paying Fable prices for?
Fable earns its price when a job has many steps, each step depends on the last, and a mistake in step three spoils everything after it. If a job is one question and one answer, a cheaper model is the right call. Our guide to which Claude model for which HubSpot job covers those everyday jobs.
If agents are new to your team, this video from our YouTube channel is a crash course for business owners.
- Tender responses. Reading the conditions, building the compliance table, drafting each criterion from your past bids and checking every claim against a source. Our tender guide runs this by hand in a Claude Project. Fable suits an agent that runs the whole sequence and hands a person the draft and the gaps.
- Quoting from many documents. Pulling a scope out of drawings, emails and a supplier price list, pricing it against your rates and writing the quote, with every assumption listed for a person to approve.
- Month-end reconciliation. Matching transactions across a bank feed, invoices and the CRM, chasing down the ones that do not match, and writing up what needs a person's decision.
- CRM clean-up over thousands of records. Working out the rules for duplicates, missing fields and bad owners, then applying them. Fable is useful for designing the rules and handling the strange cases. The bulk of the record updates belongs on Haiku.
- Code. Anthropic calls Fable its most capable model for ambitious software engineering. For a custom integration or a large change to existing code, it is worth trying.
Everything else, such as call notes, lead sorting, email drafts and weekly pipeline summaries, stays on Haiku, Sonnet or Opus. We measured what moving to Opus 5.5 did to our own work in Opus 5.5 halved my tokens and my wait, and for most teams Opus 5.5 is still the right default.
Which of your jobs would actually need Fable?
Our AI engineering team helps Australian businesses work out where Claude fits and builds the agents that run on it. Talk to Matt about the jobs you have in mind.
What does one long agent job cost on Fable?
The figures below are an illustration, not a measurement. The prices are Anthropic's published API rates in US dollars. The token counts are our assumption for one run of a tender response agent, and your job will be shorter or longer. Say the run reads 300,000 fresh tokens of tender documents and past bids, re-reads 2 million tokens from cache across its turns, and writes 80,000 tokens of drafts and working. In millions of tokens, that is 0.3 in, 2 from cache and 0.08 out, priced per million below.
- Fable 5.2: 0.3 × $8 + 2 × $0.20 + 0.08 × $40 = about US$6.00 per run.
- Opus 5.5: 0.3 × $4 + 2 × $0.20 + 0.08 × $20 = about US$3.20 per run.
- Sonnet 5.5: 0.3 × $2 + 2 × $0.10 + 0.08 × $10 = about US$1.60 per run.
Cache write charges are left out. If all 300,000 fresh tokens were also written to the five-minute cache, add about US$0.60 on Fable, US$0.30 on Opus and US$0.15 on Sonnet. Twenty tenders a month on Fable comes to about US$120 in model costs. Against the senior hours a tender takes, that is close to nothing. The expensive outcome is a wrong draft that someone has to rebuild.
Assumed token counts, not a measurement: 300,000 fresh input tokens, 2 million cached tokens and 80,000 output tokens per run, at Anthropic's list prices checked 11 November 2026. Cache write premiums left out.
Per-token price can mislead, too. A cheaper model that takes more turns, or needs a retry, can cost more per finished job. In Anthropic's early testing, Fable 5.2 at high effort scored 45.3% on FrontierCode 2.0, close to Opus 5.5's 45.5% at max effort, at roughly half the cost per task. That is a coding test, so run your own job on both models, count the tokens and count the corrections a person had to make.
Can Fable direct cheaper models instead of doing all the work?
Yes, and this is where most of the savings sit. Anthropic describes two patterns.
- Fable as the orchestrator. Fable runs the job and hands simpler tasks to smaller models. On a CRM clean-up, Fable plans the rules and handles the unusual records, and Haiku updates the other thousands at a fraction of the price.
- Fable as the advisor. A cheaper model such as Sonnet 5.5 or Haiku 5.5 does the work and asks Fable for guidance when it reaches a hard step. Anthropic's advisor tool runs this inside one API request, and only the advisor's tokens are billed at Fable rates. The tool is in beta on the Claude API, Claude Platform on AWS and Anthropic-managed Google Cloud. One catch for audit trails: when Fable 5.2 is the advisor, its advice comes back encrypted, so you cannot read what it told the other model.
The payoff from Fable comes from deciding which model does which step of an agent. Run the whole job on Fable and you pay Fable rates for the easy steps. Run it all on Haiku and the hard steps go wrong. Splitting it is design work: what each step reads, what it may write, where a person approves, and how you find out when a step fails. That is what we build in our AI engineering work. For the API basics behind it, read Claude API for business.
What happens to your data on Fable 5.2?
This is the part to check before anyone sends customer records through it. Fable 5.2 requires 30-day data retention for safety monitoring by default.
Zero data retention is only available on Fable where Anthropic expressly authorises it. Anthropic's answer for larger customers is Enterprise Frontier Safeguards, which runs the safety monitoring on prompts and responses stored in your own infrastructure, so it can work with zero retention. It is in private beta. If your contracts or privacy policy promise customers that their data is not kept, keep Fable away from that data until this is sorted, or run those steps on Opus 5.5.
Organisations with zero retention elsewhere can turn on 30-day retention for one workspace only and keep Fable there. On Amazon Bedrock and Google Cloud's Google-managed service, the cloud provider handles the data, so read their terms too.
What should you do this week?
- Keep Opus 5.5 as your default. Nothing about this launch changes the advice for everyday work.
- List your three longest multi-step jobs. The ones that take a senior person most of a day and follow the same steps each time.
- Run one on both models. Same inputs and instructions. Count tokens, minutes and corrections.
- Check retention before you test. Use documents you can safely have kept for 30 days, or a workspace set up for it.
- Price the build properly. The model bill is small. The design, testing and approvals take the time. Our AI projects start from A$10,000 and ongoing AI partnerships from A$8,500 a month, both ex GST, as listed on our AI consulting page.
Frequently asked questions
How much does Claude Fable 5.2 cost? On the API, US$8 per million input tokens and US$40 per million output tokens. Batch jobs are half that, and fast mode is US$16 and US$80.
Is Claude Fable 5.2 better than Opus 5.5? On Anthropic's own agent and coding benchmarks it scores higher, for example 51.4% against 42.5% on AutomationBench. Anthropic still recommends starting everyday work on Opus 5.5, which costs half as much per token.
Which Claude plans include Fable 5.2? Anthropic lists Pro, Max, Ultra, Team and Enterprise. At launch, Fable is included for up to 50% of weekly usage limits on Pro and standard Team seats as a promotion.
Does Claude Fable 5.2 keep my data? It requires 30-day retention for safety monitoring by default. Zero data retention needs Anthropic's express authorisation, and Enterprise Frontier Safeguards, which works with zero retention, is in private beta.
What is the difference between Claude Fable 5.2 and Mythos 5.2? They are the same underlying model. Mythos 5.2 has fewer safety blocks on cybersecurity and biology work and is only available to organisations verified through Anthropic's programs.
Sources
- Claude Fable and Mythos 5.2, Anthropic, checked 11 November 2026, for the release date, benchmarks, speed, blocks, the advisor and orchestrator roles, the launch promotion and data retention.
- Claude Fable, Anthropic, checked 11 November 2026, for the description, plans, use cases and the recommendation to start on Opus 5.5.
- Claude Fable 5.2 overview, Claude Platform documentation, checked 11 November 2026, for the model ID, context, output limit, effort defaults and platforms.
- Pricing, Claude Platform documentation, checked 11 November 2026, for token, cache, batch and fast mode prices for every model in the table.
- Advisor tool, Claude Platform documentation, checked 11 November 2026, for how the advisor pattern works, billing and availability.
- API and data retention, Claude Platform documentation, checked 11 November 2026, for the 30-day retention requirement and workspace settings.
Fable 5.2 is worth paying for on the few long jobs where a better plan saves a person a day. Finding those jobs and splitting them across the right models is what our AI engineering team builds.
Got a long job you want an agent to run?
Talk to Matt about where AI could help your team. He can explain how we work and whether our AI services fit what you need.