Short answer
Claude Fable 5.1, released September 1, 2026, is Anthropic's most capable generally available model, built for hours-long agentic work, document-heavy analysis and research. Pricing stays at $10 per million input tokens and $50 per million output, but cache reads drop 75% to $0.25, which Anthropic says cuts typical costs by about a quarter and agentic workloads by up to 45%. Anthropic's own guidance is to start most workloads on Claude Opus 5 and reach for Fable 5.1 when the cheaper model falls short on your own tests.
Model launches are written for developers and investors, so business owners tend to get either breathless coverage or a benchmark table. Here is the version we would give a client who asked "should we care?"
What it is
Anthropic describes Claude Fable 5.1 as its most capable generally available model, aimed at "jobs that take hours and span many applications". It is the successor to Fable 5, which launched earlier this summer, and it is the same model as Claude Mythos 5.1 with a different safeguard level. Mythos 5.1 goes only to vetted organisations in cybersecurity and life sciences through Anthropic's Project Glasswing programme. Everyone else gets Fable.
Concretely, Anthropic says the gains concentrate in six areas: agentic coding over long sessions, knowledge work that ends in a finished document, spreadsheet or slide deck, multi-step research, reading dense charts and tables inside PDFs, reasoning across a 1-million-token context window, and operating a browser or desktop more reliably. Early-access customers cited by Anthropic include one that let the model run an unattended 38-hour analysis job.
What it costs
| Claude Fable 5 | Claude Fable 5.1 | |
|---|---|---|
| Input | $10 per million tokens | $10 per million tokens |
| Output | $50 per million tokens | $50 per million tokens |
| Cache reads | $1.00 per million tokens | $0.25 per million tokens |
| Batch processing | $5 in, $25 out | $5 in, $25 out |
The only line that moved is cache reads, and that is the one that matters for agents. An agent re-reads its instructions, tools and history on every step, and Anthropic charges a reduced rate for that repeated prefix. Cutting the rate by 75% is why Anthropic estimates typical workloads get about 25% cheaper and highly agentic ones up to 45%, while a single short request costs almost exactly what it did before. The full breakdown is in Anthropic's developer notes.
The safeguards, in one paragraph
Fable 5.1 carries safeguards for cybersecurity and biology. Requests that trip them are rerouted to a less capable model and you are not charged Fable prices for the rerouted request. Anthropic says the refined safeguards trigger roughly 60% less often in Claude Code sessions than Fable 5's did, and that the model can help find software vulnerabilities but will not develop exploits for them. All text it writes carries Anthropic's statistical watermark, and data retention defaults to 30 days. None of this affects a normal business workflow, but it is worth knowing why an occasional request gets a different answer than you expected.
The part vendors rarely say out loud
Anthropic's own documentation for the model opens with this advice: for most workloads, start with Claude Opus 5, and use Fable 5.1 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short". We would put it more bluntly. The workflows that pay back fastest in a small or mid-size company do not need the most expensive model on the market:
- Reading an invoice and entering it into accounting
- Triaging and drafting replies to inbound inquiries
- Building a quote from your pricing rules
- Assembling a weekly report from the CRM and a spreadsheet
These are high-volume, low-judgment tasks. They are best served by a mid-tier model, a clear prompt and an approval step. Where Fable 5.1 genuinely changes the picture is in work that used to be too long or too messy to automate at all: a research memo that needs forty sources reconciled, a due-diligence pass over a data room, a monthly board pack assembled from raw exports, an agent that works through a backlog for hours while a person checks the output in the morning.
How to decide, in an afternoon
- Pick one workflow and collect twenty real examples of its input, including the ugly ones.
- Run them through a mid-tier model with the prompt you would actually use. Score the output honestly.
- Run the same twenty through Fable 5.1. If the score does not move, you have your answer and it is cheaper.
- If it does move, work out what the difference is worth per month and compare it to the cost difference. Long jobs with heavy re-reading are where the new cache pricing tilts the maths.
That is the whole method. It is also why we run an audit before recommending any model: the right answer depends on your inputs, not on the launch post.
Frequently asked questions
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?
They are the same underlying model with different safeguard levels. Fable 5.1 is the public version with Anthropic's production safeguards for cybersecurity and biology. Mythos 5.1 is available only to vetted organisations through Anthropic's trusted-access programme, for work that those safeguards would otherwise block.
Is Fable 5.1 the right model for a small company's first automation?
Usually not. Inbox triage, invoice extraction, quote drafting and report assembly run well on cheaper models, and Anthropic's documentation says to start on Claude Opus 5 for most workloads. Fable 5.1 earns its price on long, multi-step jobs: deep research, large document sets, spreadsheet and slide work, and agents that run for hours.
What does the 75% cheaper cache read actually mean?
Agents re-read the same instructions and context on every step. Anthropic caches that prefix and charges a reduced rate each time it is reused. Dropping that rate from $1 to $0.25 per million tokens is why long agentic sessions get much cheaper while short one-off requests barely change.
Does Fable 5.1 watermark what it writes?
Yes. Text from Fable 5.1 carries Anthropic's statistical watermark on every platform, in line with the EU AI Act's content-marking rule. It does not change meaning or readability and adds no hidden characters. Anthropic is rolling out a detection API in private preview.
