Claude Fable 5.1 Is Here: The Model Anthropic Pulled After Three Days Is Back, Cheaper and Twice as Capable

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, calling them the world's most advanced models for coding and knowledge work. Fable 5.1 is generally available to anyone with a Claude account. The API model ID is claude-fable-5-1, it carries a 1 million token context window and up to 128,000 tokens of output, and it prices at $10 per million input tokens and $50 per million output tokens.

Three things make this more than a version bump. It more than doubled its predecessor's score on agentic scientific research, going from 24.7 percent to 52.6 percent on Terminal-Bench-Science 0.1. It costs roughly 25 percent less to run for typical workloads and up to around 45 percent less for heavily agentic work, driven by a 75 percent cut to cache read pricing. And it ships alongside a restricted-access twin, Mythos 5.1, which is the structural answer to the problem that took the original Fable 5 offline three days after launch.

The rest of this covers the real numbers, the pricing change that matters more than the benchmarks, and the honest question most coverage skips: whether a small business has any reason to touch a model at this price. All figures here come from Anthropic's own announcement and the published model documentation.

To understand why this launch matters, you have to remember what happened in June. Anthropic released Claude Fable 5, its most capable model, and then disabled it for everyone three days later after a jailbreak collided with a US export-control directive. We wrote about it at the time in why Anthropic took down Claude Fable 5, and the lesson then was that frontier capability and public availability are two different things.

Fable 5.1 is the first update to that model line since Fable 5 launched on June 9, and the way Anthropic has structured it reads as a direct response. Rather than one model that is either safe enough for everyone or available to nobody, there are now two: Fable 5.1, generally available with full safeguards, and Mythos 5.1, the same underlying model with fewer constraints, restricted to vetted cybersecurity and life sciences organizations through trusted access programs, currently only a set of US organizations.

That split is the interesting engineering decision. The capability that made the original release untenable did not go away. It got moved behind an access program, so the model everyone can use is not crippled by safeguards designed for the small number of users who need to work near the edge.

The headline number is agentic scientific research. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6 percent against Fable 5's 24.7 percent, more than doubling it. Benchmarks are imperfect proxies, but a jump of that size in one release cycle is not noise, and it is the kind of gain that changes what a model can be trusted to do without supervision.

The other movements are meaningful but less dramatic. Agentic coding on Terminal-Bench 4.0 goes from 42.0 percent to 55.8 percent. Knowledge work measured by GDPval-AA v2 rises from 1723 to 1853. Multidisciplinary reasoning on Humanity's Last Exam, with no tools, improves from 57.8 percent to 60.9 percent. Anthropic states it outperforms Fable 5, Opus 5, and OpenAI's GPT-5.6 Sol across multiple benchmarks.

One result outside the usual leaderboards is worth noting because it shows what this class of model is being aimed at. On protein binder design, Fable 5.1 reportedly hits close to a 50 percent success rate where 10 to 15 percent is typical. That is a research capability, not a business tool, and it explains why the Mythos access program exists at all.

The per-token headline price did not move. Fable 5.1 still costs $10 per million input tokens and $50 per million output tokens, the same as Fable 5 and double Claude Opus 5's $5 and $25. On the sticker alone, nothing changed.

What changed is cache reads, which now cost 75 percent less, at $0.25 per million tokens. If you have never worked with prompt caching, this is the mechanism where a long, stable chunk of context, such as a large codebase or a document set, gets stored server-side so repeat requests do not pay full price to resend it. Anthropic puts the practical effect at around 25 percent lower cost for typical workloads and up to around 45 percent for agentic work.

That distribution tells you exactly who the release is for. Agentic work means long-running sessions that reread the same context over and over, which is precisely the shape that benefits most from cheap cache reads. A single question with a short prompt sees no benefit at all. This is a price cut aimed at people running the model for hours, not minutes.

For anyone building on it, the practical details are these. The model ID is claude-fable-5-1 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and anthropic.claude-fable-5-1 on Amazon Bedrock. The context window is 1 million tokens, roughly 555,000 words, with a maximum output of 128,000 tokens. Reliable knowledge runs through June 2026, and Anthropic commits to not retiring it sooner than September 1, 2027.

Two behavioral details differ from earlier models and catch people out. Thinking is adaptive and always on, so there is no option to disable it, and the older fixed thinking-budget parameter is rejected outright. Depth is controlled through an effort setting instead, which defaults to high on the API. Anthropic notes it defaults to high in Claude Code and to medium in Claude Cowork and on claude.ai, which is a sensible split given that consumer chat rarely needs maximum reasoning.

Effort is also the lever that makes the model affordable. Anthropic reports Fable 5.1 achieves similar or better results than Fable 5 at low or medium effort, with much higher performance at higher tiers. In plain terms: you can often get last generation's quality for less money by running the newer model at a lower setting, which is the opposite of how people usually assume upgrades work.

Honestly, probably not, and that is worth saying plainly in a week when every AI newsletter will tell you otherwise. At $10 and $50 per million tokens, Fable 5.1 is the most expensive model in the lineup, and Anthropic's own guidance is to start with Claude Opus 5 for most workloads and reach for Fable only for demanding reasoning and long-horizon agentic work, or when your testing on Opus 5 at higher effort still falls short.

For the work most small businesses actually do with AI, drafting copy, summarizing calls, answering email, cleaning spreadsheets, the cheaper models are not a compromise. Claude Sonnet 5 costs $2 and $10 per million tokens, one fifth the price, and a business owner would struggle to tell the difference on a first draft of a service page. Paying frontier prices for routine work is the most common AI budgeting mistake we see.

Where it earns its cost is long-running autonomous work: an agent refactoring a large codebase over hours, research that spans hundreds of documents, or anything where a mistake compounds across many steps and correctness matters more than the token bill. If your use of AI is a person typing into a chat box, this release changes nothing for you. If it is a system running unattended, the cache pricing alone may change your math.

Most business owners never choose a model directly. You use ChatGPT, Claude, or a tool with AI built in, and whatever runs underneath is somebody else's decision. Even so, two things from this release are worth carrying with you.

The first is that model quality keeps improving faster than most workflows adapt. A task your team gave up on eighteen months ago because the output was not good enough may now be entirely feasible, and nobody will send you a notification about it. Periodically retrying the thing that failed before is cheap and occasionally pays off enormously.

The second is the lesson from June that has not changed. Fable 5 went from released to disabled in three days, and businesses that had rebuilt a process around it were stuck. Fable 5.1 is generally available today, and it too has a retirement commitment attached, in this case no sooner than September 2027. If a critical part of your operation depends on one specific model from one specific vendor, that is a single point of failure you do not control. We covered how to think about that in our guide to the 2026 AI tool stack for small businesses.

The current Claude lineup is easier to reason about as a ladder than as a list. Claude Haiku 4.5 sits at $1 and $5 per million tokens with a 200,000 token context, built for speed. Claude Sonnet 5 at $2 and $10 is the balance point most businesses should default to. Claude Opus 5 at $5 and $25 is the recommended starting place for complex agentic coding and enterprise work. Fable 5.1 at $10 and $50 sits above all of them for demanding reasoning and long-horizon work.

All four now carry adaptive thinking, and the top three all offer 1 million token context windows, which is a meaningful change from a year ago when large context was itself the premium feature. The differentiator has moved from how much a model can read to how well it reasons across what it read and how long it can work without going off the rails.

If you want the practical head to head rather than the spec sheet, our comparison of Claude Opus, GPT-5, and Gemini 3 for small business covers how these choices actually play out on everyday work, and the AI tools we actually use every week is the shorter answer for owners who want the recommendation without the benchmark tables.

Buried under the benchmark numbers is a change that matters more for day to day use than any leaderboard result. Anthropic reports 60 percent fewer false positives in its cybersecurity safeguards, meaning the model refuses far less benign work while still declining to help build actual exploits. False refusals are the most underrated tax on working with AI: every time a model declines a legitimate request, someone rewrites the prompt, tries a different tool, or gives up, and none of that shows up in a benchmark. A 60 percent reduction is the kind of improvement people feel immediately and never quite articulate, and it is directly downstream of the safeguard work the June incident forced.

So, the short version. Claude Fable 5.1 launched September 1, 2026, generally available, at claude-fable-5-1. It more than doubles Fable 5 on agentic science, improves meaningfully on coding and reasoning, keeps the same headline price while cutting cache reads by 75 percent, and pairs with a restricted Mythos 5.1 that keeps the sharpest capabilities behind a vetting process rather than removing them for everyone. For most businesses the correct response is interest rather than action: keep using the cheaper models for everyday work, and revisit the tasks you abandoned when AI was not good enough, because the ceiling moved again. For teams running long autonomous jobs, the cache pricing is the line item to recalculate this week.

This also matters if you care about being visible in AI answers rather than just using AI, since these models increasingly decide how customers find businesses at all, a shift we covered in our guide to generative engine optimization. If you want help working out where AI genuinely fits in how your business gets found and how it converts, rather than adding another subscription nobody opens, that is a conversation we have often. Get in touch, or read how we approach web design and local SEO for Orange County businesses.