Claude Fable 5.1 Review: Pricing, Benchmarks, and When to Use It

Claude Fable 5.1 is Anthropic's generally available upgrade to Fable 5. The official pitch is simple: a little more capability, much cheaper cache reads, fewer false refusals on legitimate security work, and writing that no longer sounds like a model. The practical question for builders is different: do you actually need it, or is it an expensive default?

In the companion video below, Will — founder of Dyad and a former Google engineer — pressure-tests the launch with an independent benchmark: three real business apps, not abstract coding puzzles. The result is more useful than the headlines. Fable 5.1 is a genuinely strong model. For most everyday vibe-coding work, it is probably not the model you should reach for first.

What Anthropic actually shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 as the same underlying model with different safeguards. Fable 5.1 is generally available. Mythos 5.1 is restricted to trusted-access programs aimed at cybersecurity and life sciences.

The primary source is Anthropic's launch post, Introducing Claude Fable 5.1 and Claude Mythos 5.1.

Confirmed from Anthropic, not from social recaps:

• API price is still $10 per million input tokens and $50 per million output tokens — the same headline rates as Fable 5.
• Cache reads dropped from $1.00 to $0.25 per million tokens. That is the only line on the public rate card that moved.
• Anthropic estimates typical workloads cost about 25% less than Fable 5, and highly agentic workloads can reach about 45% savings because they reread cached context so often.
• On Anthropic's Terminal-Bench-Science 0.1 score, Fable 5.1 reaches 52.6% versus Fable 5 at 24.7%.
• On Terminal-Bench 4.0, Fable 5.1 scores 55.8% (Mythos 5.1 60.9%) versus Fable 5 at 42.0%.
• Anthropic says Fable 5.1 is about 60% less likely to refuse legitimate cybersecurity tasks that Fable 5 blocked. It can help discover software vulnerabilities; it is not supposed to help develop exploits.
• Default effort is High in Claude Code and Medium in Claude Cowork and Claude.ai. Lower effort is faster and cheaper; raise it when the task is genuinely hard.

Product-page pricing and plan access are also listed on Claude Fable: Pro, Max, Team, and Enterprise, plus the Claude API and major cloud marketplaces.

The cost story is cache, not a 25% list-price cut

This is the easiest fact to get wrong. Fable 5.1 is not 25% cheaper on every token. Input and output are unchanged. Your bill falls only to the extent you hit prompt cache. If your workflow barely caches, you will see almost no savings. If you run long-horizon agents that keep a large project in cache, the 45% figure is plausible.

That is also why cost per task matters more than cost per token. A model that spends fewer tokens — or finishes in fewer loops — can beat a cheaper-looking model on the invoice even when its sticker price is higher.

Independent check: Dyad's three-app benchmark

Will's benchmark asks models to build a CRM, a help desk, and a SaaS admin portal with authentication, user roles, and permissions. That is closer to what AI app builders actually ship than a coding-puzzle leaderboard.

In that test, Fable 5.1 was the expensive option: about $47 for the same three-app workload that cost about $24 on Opus 5 and about $20 on GPT-5.6 Sol. Quality across the top models was close enough that no model won every app. Cursor Bench, which Grok 4.6 currently leads in the numbers Will walked through, is a vendor-adjacent coding benchmark and should be treated with extra scrutiny for that reason.

Takeaway if you build with Cursor, Dyad, Codex, or Claude Code: keep model choice in the tool. Locking every task to Fable 5.1 is how a strong model becomes a bad default.

Where Fable 5.1 still earns the spend

Use Fable 5.1 when the task is long, messy, or expensive to get wrong:

• Multi-hour or multi-day coding agents that plan, verify, and recover without you babysitting every step.
• Security review of an app you actually plan to ship. Fable 5's over-refusal was a real complaint; 5.1 is the version Anthropic says will do defensive work without constantly blocking it.
• Hard debugging. Anthropic published a Millennium researcher anecdote: a one-in-a-million crash that other models, including Fable 5, missed and Fable 5.1 found by tracing a vendor library.
• Writing and research-heavy work. Will's side-by-side on the same prompt showed Fable 5 still producing "Claudisms" — stacked caveats, odd phrasing, em dashes everywhere. Fable 5.1 read like a person wrote it.
• UI design. In the head-to-head web-design ELO numbers Will cited, Fable 5.1 sat at 1765, with open-weight Qwen 3.8 at 1688. Feed it two or three designs you like rather than asking it to clone one site. Watch for AI tells such as two-tone headlines.

When you should not default to Fable 5.1

Skip it for restyling buttons, boilerplate CRUD, first-draft landing pages, and anything you can verify in a minute. A cheaper model at a lower effort setting will usually get you there. If you are already on GPT-5.6 Sol, Opus 5, or a strong open-weight model and the output is good, do not "upgrade" just because Anthropic shipped a point release.

If you need the Fable 5 baseline — access, pricing, and when the original Fable-class model is the right tool — start with our Claude Fable 5 guide. For editor setup, see how to use Claude Fable 5 in Cursor.

If you are choosing between Anthropic and OpenAI for production coding, we also compared Claude Fable 5 vs GPT-5.6 Sol.

Mythos 5.1 vs Fable 5.1

You probably want Fable 5.1. Mythos 5.1 is the same model with tighter safeguards and a restricted access program for cybersecurity and biology work, including a US-government-developed biology access track. Unless you are in that trusted-access bucket, Fable 5.1 is the product you can actually use.

Zero data retention: what is true today

Do not treat "zero data retention" as a toggle every builder gets on day one. Anthropic is rolling out Enterprise Frontier Safeguards (EFS): customer-controlled cloud storage and zero data retention for enterprise customers later in fall 2026. Until EFS is broadly available, Anthropic says eligible users can run Fable 5.1 with zero data retention. If you sell into healthcare, finance, or legal, get the current eligibility in writing from Anthropic rather than assuming the launch blog post is your DPA.

How to try Fable 5.1 this week

• In Claude Code or Claude.ai, confirm you are on Fable 5.1 and start at Medium effort unless the task is long-running.
• In Cursor or any router that bills API usage, run one representative agentic task on Fable 5.1 and the same task on your current default. Compare invoice and retries, not vibes.
• If you use a local or model-picker builder, keep Fable 5.1 as an option rather than the only option. Our Dyad review covers a free local setup that lets you swap models per task.
• For security scans, retry the prompts Fable 5 used to refuse. That is the regression that actually changed for Dyad users.

Bottom line

Fable 5.1 is a better Fable, not a new category. The research-bench jump is real. The cache-read cut is real. The writing quality jump is real. The independent coding-cost check is also real: you can pay more than twice GPT-5.6 Sol to build the same three apps. Pick it for hard, cached, long-horizon work. Do not make it the default for every prompt.