Claude Fable 5.1

Fable 5.1 can be connected and compared on your own work. There is no free tier for it, and a Pro subscription alone does not cover this model's usage.

The model is worth trying where the accepted result matters more than the price of one short response. We use it for website code and editorial tasks; a good draft still goes through review.

You need to understand someone else's project, fix a complex bug, or build a prototype. In the official Terminal-Bench 4.0, Fable 5.1 scores higher than Fable 5; the conditions and model substitution are covered below.

You work through Claude Code with a long history and files. Reading already cached context in the API is now four times cheaper; standard input and output cost the same.

You need a cheap model for large volumes of short requests. Caching will not help if there is nothing to reread, while Fable's base rates remain high.

You want to deliver text or a program without proofreading. In our runs, Fable 5.1 invented a date and missed requirements; for an independent author, one code response did not run at all.

Ways to start are collected in chapter 3, and our experience and failures are in chapter 6. If you are just getting started with AI-assisted work, begin with an introduction to vibe coding.

Fable performs these tasks through the app's tools. For a map, it needs data and calculations; for a website, project files and a browser. The examples below show the model's output together with these tools.

Venus terrain. In an Anthropic demonstration, Fable 5.1 trained a neural network on Magellan mission radar data and an earlier elevation map, reconstructing the terrain of about one-third of the planet. According to the authors, the new map distinguishes details 2–3 km across instead of 10–20 km, and determines elevations up to 25% more accurately. The data has been published for verification.

Cities in Three.js. PhiloLabs published Union Square in San Francisco and a walk through Kyoto: buildings, streets, and camera movement are assembled in browser code. According to the author, several Fable agents studied the materials, wrote code, and checked the scenes. The source code and a verification report are available alongside the result.

Portfolio site. Thomas Wiegold asked Fable 5.1 to build a site for the fictional Marrow studio and says he got the result from one prompt with no edits. The published site has working transitions between projects, changing compositions, and typography. The clients in the demo are fictional too.

Our newsletter. On 7 September, Fable 5.1 adapted the Canvas UI component for a VHS effect on the newsletter page: the original component needed support for an experimental browser API. Grain and scan lines cover the lighthouse illustration, while the text and form remain sharp. The illustration already existed; Fable's role in building the effect is recorded in our log.

In the same issue, two small images with test results were replaced by Russian HTML tables. Fable helped move numbers and labels into layout data, then a test checked the key values.

Editor of the week. On 4 September, Fable 5.1 sorted 582 publications into seven story threads after we removed individual news ratings and asked it to assemble the week's picture. Changing the task improved the selection of key items, but the editor still rejected two already measured models as unverified. Here, quality also depended on the task framing.

Finding a rare failure. In its release feedback, Millennium says Fable 5.1 analyzed a third-party library, matched it against crash data, and found the cause of an error engineers had been unable to explain for several years. The feedback has no open reproducible test. The measurements below include results with described test conditions.

In Claude, select Fable 5.1 from the model list. Before a long task, check exactly how your plan bills Fable. Having Pro does not mean model usage is included in its standard limit.

To launch Claude Code from the terminal, use claude --model claude-fable-5-1. In an open session, the /model command changes the model. If the client requests an update, update it first; manually specifying an ID does not grant account access.

Reasoning effort. According to Anthropic's release, Claude Code defaults to High, while Claude and Cowork use Medium. This is how much work the model puts into an answer, not a subscription name. In the API, reasoning is adaptive and always enabled; the Max mode in the test table is also not the Max subscription.

Start with a familiar task whose result you can verify. Provide source materials, constraints, and a definition of done. For code, ask to see test results; for research, ask for sources behind claims. When the model switches to Opus, check the notification and the actual responder.

Fable 5.1 improved on Fable 5's results in Anthropic's published run. The model name in a row does not guarantee that it completed every task itself: when safeguards triggered, some tasks were handed to Opus. Compare the same test version under the same conditions.

Our code run. On 6 September, we gave Claude Code + Fable 5.1, effort High, 25 tasks with three attempts each. Initially, 60 of 75, or 80%, counted as successful. After checking references to ready-made code, three successes were removed: 57 of 75, or 76%, remained. We found no substitution of another model in these 75 attempts.

The full RuBench analysis also includes failures. Two tasks were not solved in any of the three repeats, even though the model attempted them. For us, this is a useful distinction between refusal, switching to another executor, and an ordinary wrong answer.

In Intelligence Index v4.3 dated 7 September, Astra and Fable 5.1 each have 53 points at maximum effort; Fable is listed with fallback mode. This is a different index, not the Coding Agent Index. One overall score does not erase differences in price, access, and results on individual tasks.

Fable 5.1's base API rates have not fallen: input costs $10 per million tokens and output costs $50. Rereading already cached context is cheaper, at $0.25 per million. This is where a long session with files and tools can save money.

According to Anthropic's estimate across four weeks of August workloads, the total bill for typical work falls by about 25%, and for complex agentic work by up to 45%. This is an estimate for specific workloads at standard effort. A request with no repeated cache reads does not automatically receive this discount.

Subscription and credits. On Pro, Fable 5.1 uses separate usage credits from the start. On Max, it is included in the overall plan but can use up to half of the weekly pool; after that, credits or another model are needed. This is a limit within the shared pool, not an additional 50% or unlimited use.

Access in Russia. Russia is not on Anthropic's official list of supported countries. Installing the client or buying a subscription does not by itself remove the regional restriction. An intermediary may have its own prices, data-retention terms, and model selection; check the exact ID and service rules before paying. A review of AI service payment terms can help you understand intermediaries.

We trust Fable 5.1 with drafts and work that has verifiable results. In the editorial workflow, a separate review found errors that sounded plausible. Even a large context window did not guarantee compliance with every rule.

Three passes instead of one. In the September 5 email, the first version overloaded the opening with numbers, the next invented a date, and the third passed our review without the required fixes. Meanwhile, the missing image in another run turned out to be an input-materials problem: the model did not receive a suitable image. Repeating the prompt is pointless until the supply is fixed.

Code also needs checking. In BitsMinds' September 5 test, Fable 5.1 created an expressive amusement park, but its Rubik's Cube did not run. The author attributed the error to referencing a variable before initialization. The conditions were strict: one attempt per task, maximum effort, no browser testing. This is one author's result, not our estimate of error frequency.

Who actually responded. In Claude and Claude Code, automatic switching to Opus is enabled by default. It may trigger because of the request or material the model has read. After switching, the model remains Opus until you change it manually. In the API, this switch must be configured separately; without it, a response may end in a refusal with HTTP 200.

According to Anthropic, the new safeguards trigger about 60% fewer interventions per session on cyber tasks than the previous Fable 5 safeguards. Vulnerability discovery is allowed, but some tasks are still redirected. Our 0 substitutions in 75 attempts confirm only our own run; they do not support promises that it will "never refuse again".

Data retention. Fable 5.1 uses the Covered Models policy. By default, it requires at least 30 days of retention for prompts and responses, with exceptions for investigations and legal requirements. For consumer Free, Pro, and Max plans, this policy does not change prior retention; retention duration alone does not determine whether data is used for training.

We build the site with agents, but do not attribute its entire history to version 5.1. Astra created the first card itself, Fable refined it later, and a separate review found the remaining errors. For us, the value of a strong model shows in this kind of work with checks, not in a promise to finish everything without a human.

Fable or Astra. The base API rates are the same: $10 for input and $50 for output, but Fable cache reads cost four times less. The models are tied in Artificial Analysis's overall rating; Astra leads in the web arena. The choice depends on the task and the app you will use.

On September 7, Viktor Taelin stuck with Fable for coding for its quality, sound judgment, and stability. At the same time, he considered Astra stronger for interfaces, computer use, and 3D. This is a practitioner's view based on his tasks. Public PhiloLabs work shows that Fable can also build 3D scenes, but does not establish an overall winner.

Fable or Opus. Since September 22, 2026, Anthropic documentation has recommended starting most tasks with Opus 5.5 and moving to Fable when more complex reasoning or long autonomous work is needed. In the Opus 5.5 release table for Terminal-Bench 4.0, the new Opus scores 66.4% against Fable 5.1's 55.8% at rates 2.5 times lower; Fable's own results there have not changed. For a short edit in a familiar project, try the Opus available to you and assess the finished result.

Fable 5 or 5.1. If you have already chosen Fable, the new version keeps the base rates and lowers the cost of cache reads. It also improved a number of results in Anthropic tests. Fable 5's own July result cannot be directly subtracted from 5.1's September result: the conditions and run sets differ.

The Astra profile includes its real-world work and access conditions. For an initial comparison, give both models one familiar task and track the time, number of fixes, and cost of the accepted result.

Fable 5 launched on 9 June 2026; three days later, Anthropic temporarily suspended access, then restored it on 1 July. Version 5.1 appeared on 1 September with updated capabilities and cheaper cache reads. June events and the previous version's results retain their dates.