Claude Fable 5.1 for Business: What Changed, Who Gets It

Anthropic released Claude Fable 5.1 on September 1, 2026. It is their most capable generally available model, it runs $10 per million input tokens and $50 per million output, and it is built for long-horizon agentic work: multi-hour coding runs, multistep research, dense document handling. Claude Mythos 5.1 shipped alongside it and is the same model with different safety routing, available only to vetted organizations through Project Glasswing.

The release page leads with benchmark tables. For a business the decision is not on that table. It is in the pricing line, the retention terms, and the availability tier, and we think you should read those three before you read a single score.

One note on where we sit. We are multi-provider on purpose. Smile PreVue runs its image generation on Google Vertex AI Gemini under a BAA, our orchestration work runs on Claude, and we have shipped against OpenAI too. This is a reading of one release, not a loyalty post.

What changed from Fable 5 and from Opus 5?

The headline numbers from Anthropic's announcement, with the comparisons that matter:

• Terminal-Bench 4.0: 55.8 percent, against Fable 5 at 42.0 percent and GPT-5.6 Sol at 37.3 percent.
• AutomationBench: 31.4 percent, against Opus 5 at 26.9 percent and Fable 5 at 17.1 percent.
• OSWorld 2.0 (strict): 41.7 percent, against Opus 5 at 39.6 percent and Fable 5 at 36.1 percent.
• GDPval-AA v2: 1853, against Opus 5 at 1824 and Fable 5 at 1723.

Now read those against Opus 5 specifically, because Opus 5 is half the price. On Terminal-Bench, Opus 5 scores 52.3 percent to Fable 5.1's 55.8. On OSWorld strict it is 39.6 to 41.7. On GDPval it is 1824 to 1853. The jump from Fable 5 to Fable 5.1 is large. The jump from Opus 5 to Fable 5.1 is a few points, for double the per-token price.

Anthropic says this out loud in their own docs, which is the most useful sentence in the whole release: start with Claude Opus 5 for most workloads, and use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5 at higher effort still fall short.

Read that again, because it is a vendor telling you not to default to its flagship. It is also exactly how we think model selection should work. The eval decides the tier, not the launch post. If you do not have an eval, you do not have a way to know whether the more expensive model is buying you anything, and you will end up paying double for a rounding error.

What does Claude Fable 5.1 cost?

The cache read column is the strange one, and it is doing a lot of work. Fable 5.1 reads cached tokens at 0.025x its base input rate, where the rest of the Claude lineup sits at 0.1x. That is why Fable 5.1 cache reads are cheaper in absolute dollars than Opus 5 cache reads despite the model costing twice as much per fresh token. Anthropic estimates typical workloads land about 25 percent cheaper than on Fable 5, and highly agentic work up to roughly 45 percent cheaper, almost entirely on that mechanic. We worked through what that does to long agent loops last week, so we will leave it at one line here.

The raw price is still double Opus 5. A workload that reads a large stable prefix over and over benefits enormously. A workload sending fresh tokens every call does not.

Who can get Mythos 5.1, and does it matter for you?

Fable 5.1 is generally available: the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Mythos 5.1 is not. It is limited to participants in Project Glasswing, reached through the Cyber Verification Program for defensive security work and the Life Sciences Verification Program, and it is currently restricted to US organizations. Anthropic's pricing page marks it "limited availability."

For almost every small and mid-sized business, the honest answer is that Mythos does not matter to you and will not for some time. We are flagging it for one reason: if a vendor or an agency tells you their product runs on Mythos 5.1, ask which verification program they are in. It is a specific, checkable claim, and it is not something you get by signing up.

This is a general habit worth building. "The model exists" and "we can actually get it" are different sentences, and the gap between them is where a lot of AI roadmaps quietly break.

The data retention line most buyers skip

Here is the one that decides whether you can use this model at all. From Anthropic's own documentation: Claude Fable 5.1 and Claude Mythos 5.1 carry 30-day data retention and are not available under zero data retention unless expressly authorized by Anthropic.

Zero data retention means the provider does not store your prompts or the model's outputs after the request completes. It is a contractual term, and for a lot of businesses it is not optional. If you handle patient information, privileged legal material, or financial records, your own customer contracts probably promise something about where data sits and how long it lives. A model with mandatory 30-day retention may simply be off the table, regardless of how it scores.

Anthropic is building the answer to this. Enterprise Frontier Safeguards is described as giving enterprise customers the privacy of a zero data retention agreement while still letting Anthropic detect misuse, and it is rolling out in phases beginning this fall. Eligible customers can use Fable 5.1 under zero data retention before that lands. What "eligible" means is not published, so if this is you, ask your account team in writing and keep the answer.

For our own regulated work the calculus is simple. Smile PreVue runs on Vertex AI under a BAA because a dental practice is a covered entity and we sign an agreement with the clinic about where protected health information goes. We would not move that workload onto any model whose retention terms we could not put in writing to the clinic, and no benchmark would change that. The compliance surface picks the model, then capability picks among what is left.

What breaks when you switch

Fable 5.1 is not a drop-in for Fable 5. Three things to plan for:

Forced tool use returns an error. tool_choice set to any or tool now returns a 400. If you were forcing a tool call to guarantee structured JSON back, move to strict tool use or structured outputs instead. This is a code change, not a prompt change.

Thinking blocks are bound to the model and to an append-only history. A router that switches models mid-conversation drops them. Editing earlier turns invalidates them, and that is enforced with an error for accounts created on or after August 31, 2026. If your harness rewrites history, that is now a design problem to fix rather than a habit to keep.

It batches fewer parallel tool calls. Where Fable 5 might issue several tool calls in one turn, Fable 5.1 may issue one. In an agent loop that means more round trips and more wall-clock time unless your prompt explicitly asks for batching.

There is also a behavior to handle rather than a break. Safety classifiers can decline a request, returning a normal HTTP 200 with a refusal stop reason and a category. Your application has to check that before it reads the response body. Anthropic supports falling back to Opus 5 or Opus 4.8 on a refusal, with credit covering the prompt cache cost of the switch.

None of this is exotic. It is just the reason you re-run the eval suite before you change a model ID in production, instead of after.

When we would pick it, and when we would not

We would reach for Fable 5.1 on long unattended agent runs, multi-repo refactors, research that has to follow up on what it finds, and dense financial or technical PDFs. The clean trigger is the one Anthropic gives you: Opus 5 at high effort ran your eval and did not clear the bar.

We would not reach for it on short high-volume calls, where Sonnet 5 or Haiku 4.5 does the job for a fraction of the spend. We would not reach for it under a zero-retention contract until Enterprise Frontier Safeguards reaches you or you have written authorization. We would not reach for it in a codebase that depends on forced tool calls until that is refactored. And there is no point reaching for it for offensive security or life-science R and D work, because those queries route to Opus models anyway by design.

The frame we would hand any operator: price, retention, and availability first, then capability. A model you cannot legally run is not a cheap model or an expensive one. It is not an option.

FAQ

Is Claude Fable 5.1 generally available? Yes, on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry since September 1, 2026.

Is Claude Mythos 5.1 available to my company? Only through Project Glasswing, for vetted cyber defenders and life scientists at US organizations.

Does Claude Fable 5.1 support zero data retention? Not by default. Thirty-day retention applies unless Anthropic expressly authorizes otherwise, with Enterprise Frontier Safeguards rolling out in phases beginning this fall.

Should I switch from Opus 5? Anthropic's own guidance is to start with Opus 5 and move up only when your evals at higher effort fall short. Run the eval before you run the migration.

Where we would start

If you are trying to decide whether a frontier model earns its price in your product, the answer is an eval on your actual traffic, not a benchmark table. That is the first thing we build on most engagements, and it is the part teams most often skip. This is how we practice AI integration for business.

If you want that built with you, start a project. If you would rather learn to run this yourself, that is a real option too and it is what /learn is for.