Claude Fable 5 for business: pricing, API, use cases and adoption strategy

On 9 June 2026, Anthropic launched Claude Fable 5, the first Mythos-class model available to everyone. For a business owner or an IT lead, the question was never whether the model is impressive. It is what to do with it: which tasks justify its price, what it really costs, how to integrate it, and what its data retention changes for a European company.

This article was first published on launch day and revised on 17 October 2026. The prices, the model line-up and the API behaviour below are the ones Anthropic documents on that date, with links. In four months, the line-up has changed a lot: Fable 5 is now a legacy model, succeeded by Fable 5.1 on 1 September, and Anthropic now recommends starting most workloads on Opus 5.5, released on 22 September. For the launch itself, the Mythos class and the science results, read our complete guide to Claude Fable 5 and Mythos 5.

Claude Fable 5 pricing, and the current line-up

Fable 5's API price is 10 dollars per million input tokens and 50 dollars per million output tokens. Fable 5.1 keeps exactly that base price. Here is Anthropic's price list, read on 17 October 2026, for the models a company is likely to compare:

Model	API ID	Input (per M tokens)	Cache read	Output (per M tokens)	Status
Claude Fable 5.1	claude-fable-5-1	10 dollars	0.25 dollar	50 dollars	Current, released 1 September 2026
Claude Fable 5	claude-fable-5	10 dollars	1 dollar	50 dollars	Legacy, retirement not before 9 June 2027
Claude Opus 5.5	claude-opus-5-5	4 dollars	0.20 dollar	20 dollars	Current, released 22 September 2026
Claude Opus 4.8	claude-opus-4-8	5 dollars	0.50 dollar	25 dollars	Legacy
Claude Sonnet 5.5	claude-sonnet-5-5	2 dollars	0.10 dollar	10 dollars	Current, released 28 September 2026
Claude Haiku 5.5	claude-haiku-5-5	0.10 dollar (prompts up to 100,000 tokens)	0.01 dollar	0.50 dollar	Current, released 7 October 2026

All four current models have a 1-million-token context window and 128,000 tokens of maximum output, according to the models overview. Two details matter for budgets: Batch API requests are 50% off, and models from Opus 4.7 onwards use a tokenizer that produces about 30% more tokens for the same text than Sonnet 4.6 and earlier.

Three worked scenarios

The calculations below are ours, at list price, without batch or cache discounts. The token counts are assumptions to adapt to your own documents.

Scenario 1: analysing a 100-page report

Assume 75,000 input tokens and a structured 2,000-token summary. On Fable 5 or 5.1: 75,000 tokens at 10 dollars per million is 0.75 dollar, 2,000 tokens at 50 dollars per million is 0.10 dollar, so 0.85 dollar per document. The same task costs 0.34 dollar on Opus 5.5 and 0.17 dollar on Sonnet 5.5. Against half a day of an analyst's time, the gap barely matters, provided the task actually needs that level of reasoning.

Scenario 2: an internal assistant at 500 requests a day

Take 500 requests a day, with 3,000 input tokens of context and 800 output tokens per answer: 1.5 million input tokens and 400,000 output tokens a day. Over 22 working days, that is about 770 dollars a month on Fable 5.1, 308 dollars on Opus 5.5, 154 dollars on Sonnet 5.5 and less than 8 dollars on Haiku 5.5. For high-volume, low-complexity work, the most capable model is the wrong choice.

Scenario 3: an agentic code migration

A generous assumption for a heavy project: 20 million input tokens (repeated file reads, context, tool round-trips) and 1 million output tokens. That is 250 dollars on Fable 5.1 before caching, and 100 dollars on Opus 5.5. In agentic work, most input tokens are re-read context, so prompt caching changes the picture: a cache read on Fable 5.1 costs 0.25 dollar per million tokens instead of 10. Anthropic estimates that Fable 5.1 costs about 25% less than Fable 5 on typical workloads and up to about 45% less on highly agentic work (Anthropic, September 2026).

The conclusion fits in one word: routing. Send each request to the cheapest model that handles it correctly, and keep the most expensive one for the tasks where its lead is largest.

What Fable 5 changed for a company

Anthropic's announcement sums up the positioning in one idea: the longer and more complex the task, the larger Fable 5's lead over Anthropic's other models. The published results point to four families of use:

• Large code changes. At Stripe, a codebase-wide migration in a 50-million-line Ruby codebase took one day, against over two months for a team by hand. For a smaller company, the equivalent is the framework migration or the technical-debt project nobody has time for.
• Senior-level document and financial analysis. Top score of any model on Hebbia's Finance Benchmark; at IMC, near-perfect trading-analysis evaluations. Due diligence, contract review, long sector reports.
• Long-running agents. Fable 5 stays focused across millions of tokens and improves its work with notes in files; on Slay the Spire, file memory helped it three times more than Opus 4.8.
• Vision. Extracting precise numbers from figures, and rebuilding a web app's source code from screenshots alone.

Since 22 September, there is a new question before choosing Fable: Anthropic describes Opus 5.5 as matching Fable 5.1 on most work while being cheaper to run, and its documentation recommends Fable 5.1 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short". Test Opus 5.5 first.

Which model for which task

Model	Task profile	Examples
Fable 5.1	Long, complex, multi-step work where Opus 5.5 falls short in your tests	Codebase migrations, full due diligence, agents running for hours with file memory
Opus 5.5	Default for demanding work	Agentic coding, analysis, writing that needs reasoning; and requests that Fable's classifiers would decline
Sonnet 5.5	Volume with a good speed and quality balance	Internal document assistant, content, level-2 support, search over a knowledge base
Haiku 5.5	Real time and micro-tasks	Ticket classification, routing, field extraction

Integrating Fable through the API

The call is the same as for any Claude model, on the Messages endpoint. With the current Fable model:

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-fable-5-1",
    max_tokens=2048,
    messages=[
        {"role": "user", "content": "Analyse this contract and list the liability clauses."}
    ],
)

print(message.stop_reason)
print(message.content[0].text)

If you already call claude-fable-5, Anthropic lists three breaking changes when moving to 5.1, including forced tool use, which now returns an error (Fable 5.1 overview).

Refusals and fallback: what your pipelines must handle

Fable models ship with safety classifiers on cybersecurity, biology and chemistry, and distillation. At launch, Anthropic wrote that flagged requests were answered by Opus 4.8 in less than 5% of sessions on average. On the API, the current documentation describes it this way: a declined request returns a normal response, not an error, with stop_reason: "refusal" and the policy area in stop_details.category. You can let the API retry on the model Anthropic recommends by setting fallbacks to "default" with the server-side-fallback-2026-07-01 beta header, or give your own list of up to three models. When a fallback happens, the response names the model that served it.

Three practical rules follow. Log every refusal and every fallback, per use case. Treat a fallback answer as a valid answer. And do not resend a refused request in a loop: if a use case is declined systematically, call the model that serves it directly. With Fable 5.1, Anthropic says its cyber safeguards produce about 60% fewer false positives and its biology safeguards trigger 85% less on harmless basic-biology and medical queries.

Data retention and GDPR

This is the section your data protection officer should read. Anthropic applies a 30-day retention to all Mythos-class traffic, on its own products and on third-party platforms, business customers included. The data is not used to train models or for any non-safety purpose, all human access is logged, and deletion happens after 30 days in almost all cases (announcement of 9 June 2026). With Fable 5.1, Anthropic announced Enterprise Frontier Safeguards, which keep customer data in customer-controlled storage, rolled out in phases from this autumn.

For a European company, the work is standard but should not be skipped:

• Record of processing: name Anthropic as a processor and mention the 30-day safety retention.
• Data sent: if prompts contain personal data of customers or employees, check that your legal basis and privacy notices cover the transfer and the retention period.
• Internal AI policy: say what may or may not go through Mythos-class models.
• Vendor file: keep the announcement and the system card next to your data processing agreements.

When not to use Fable

• Simple, high-volume tasks. Classification, extraction, rephrasing, first-line answers: Haiku 5.5 or Sonnet 5.5 do the job for a fraction of the price.
• Domains that trigger the classifiers. Pharma, chemistry, offensive security: if your requests are declined or fall back often, call the model that serves them directly. Anthropic says dual-use cyber work such as pentesting is routed to Opus models.
• Latency-critical work. Real-time chat or synchronous scoring: Anthropic itself lists Fable 5.1 as the slowest of the current line-up.
• A data policy incompatible with vendor retention. Until Enterprise Frontier Safeguards reach your account, Mythos-class models keep 30 days of traffic.

A note on the June episode

Three days after launch, on 12 June, Fable 5 and Mythos 5 were suspended for all users after US export controls, and came back on 1 July. Any product wired to Fable 5 alone stopped for 19 days. The story is told in our article on the suspension and the redeployment, and the lesson is the same as for routing: keep the model behind your own interface, with a tested fallback to another one.