Claude Fable 5.1: What Anthropic’s Most Powerful Model Can Do and How to Build a 3D Aircraft Engine With It

What Has Changed in Claude Fable 5.1  

Mythos is Anthropic’s top-tier model class, ranking above Opus in capabilities. The difference between Fable and Mythos is not intelligence, but safety restrictions. Fable 5.1 is the public version with additional classifiers, while Mythos 5.1 is the same model with looser restrictions for vetted specialists.

The road to this release was not an easy one. The original Fable 5 was unveiled in June, but just a few days later Anthropic suspended access to it and to Mythos 5 at the request of U.S. authorities, which restricted model use by foreign nationals. In July, the restrictions were lifted, and the model returned to general availability. Fable 5.1 is launching without such caveats — as a standard update available to all paid users from day one.

Claude Fable 5: Testing the First Publicly Available Mythos-Class Model

10.06.2026

Read now

The update itself is, first and foremost, a quality boost on complex, long-running tasks. Anthropic highlights six areas where the gap versus Fable 5 is most noticeable and grows as the effort level increases:

• agentic programming (multi-file tasks, large refactors, and debugging over the course of hours);
• knowledge work (documents, spreadsheets, and presentations from scratch)
• research and web search
• vision work (dense charts and tables in PDFs)
• reasoning across the entire 1 million-token context window
• computer use.

Anthropic separately notes that Fable 5.1 avoids “cutting corners” that leads to low-quality results, and can dig down to the root cause of a problem rather than masking the symptom. Developers should also account for behavior changes: compared with Fable 5, the model more often makes a single tool call per step instead of batching, less often writes intermediate explanations between calls, and at low effort levels more often answers from memory rather than turning to search. All of this is controlled via settings and prompt instructions, but familiar agent workflows may need a bit of retuning.

A telling example from the announcement is a case from the investment firm Millennium. The model found the cause of an extremely rare failure (about one in a million runs) that engineers could not explain for four to five years, and that all other models missed, including Fable 5. Fable 5.1 disassembled a third-party library, matched it against a memory dump, and traced the failure to a bug in that library.

The second major change is economics. Anthropic kept the price per token, but radically reduced the cost of cache reads — an operation that accounts for most of the spend in agent workflows. More on this below.

The third is the safety mechanisms, which in the original Fable 5 drew the most criticism. They have been made more precise: they less often block harmless requests, but still route truly risky ones to Opus models.

Capabilities and Benchmarks  

According to Anthropic, Fable 5.1 sets a new standard in programming, knowledge work, and long-running tasks.

Scientific research. On Terminal-Bench-Science 0.1 — a benchmark for agentic scientific tasks — Fable 5.1 scores 52.6% versus 24.7% for Fable 5, meaning more than double.

Terminal programming. On Terminal-Bench 4.0, the result rose to 55.8% versus 42.0% for Fable 5. Mythos 5.1 on the same test posts 60.9% — the gap is explained by tasks that Fable 5.1’s triggered cyber classifiers block.

Agentic coding. On CursorBench 3.2.0, the model reaches 73.4% versus 70.5% for Fable 5 — the best result among those tested on this benchmark.

Multidisciplinary reasoning. On Humanity’s Last Exam — 65.0% with tools and 60.9% without them, above the results of Fable 5 and Opus 5.

Computer use. On OSWorld 2.0 — 77.9% under the lenient scoring mode and 41.7% under the strict one, also higher than Fable 5.

In Anthropic’s comparison table, Fable 5.1 beats its own Fable 5 and Opus 5, as well as OpenAI’s GPT-5.6 Sol, on almost every metric — from business workflows (AutomationBench: 31.4%) to knowledge tasks (GDPval-AA v2: 1,853 points).

An important caveat: Fable 5.1 was tested with production safety defenses enabled, so on tasks where the classifiers triggered, the model received zero points — which understates its real performance on some benchmarks like OSWorld 2.0.

Separately, Anthropic collected feedback from early-access partners, and the practical numbers are telling. Investment firm Every described the model with the formula “Fable-level intelligence, Opus-level price, Sonnet-level speed”: in its tests, Fable 5.1 was about twice as fast as Opus 5 and used half as many tokens. Browserbase reported that on its most difficult browser benchmark, the model completed 82% of tasks — versus 74% for Opus 5 and 57% for Fable 5 — while spending fewer tokens than both. Legal startup Crosby recorded an increase from 47.9 to 57.0 on its contract-redlining benchmark, and financial service Rogo saw the same accuracy as Fable 5, but with 20% fewer tokens.

A separate announcement block is dedicated to science. Mythos 5.1, having gained access to open protein-design tools, engineered high-affinity binding proteins — across three targets, ten times better than the best entries from Adaptyv Bio competitions, with an Anthropic-record success rate of around 50% (typically 10–15% in this field). Fable 5.1, meanwhile, trained a neural network and built a new elevation map of one-third of Venus’s surface using radar images from NASA’s Magellan mission from 30 years ago: resolution improved from 10–20 km to 2–3 km, and elevation accuracy — by about a quarter.

Prices and Effort Levels  

The token pricing for Fable 5.1 itself hasn’t changed: $10 per million input tokens and $50 per million output tokens. The main change is elsewhere — cache reads (reusing already processed context) are 75% cheaper, down to $0.25 per million tokens.

For agentic workflows, this is fundamental: the model rereads a large shared context over and over, and cache reads make up the bulk of the bill. According to Anthropic, the total cost of a typical workload drops by about 25% versus Fable 5, and for context- and tool-heavy tasks — by up to 45%. Batch processing (Batch API) still provides a 50% discount.

The second practical detail is effort levels, of which the model has five: Low, Medium, High, Extra high (xhigh), and Max. They control how much the model “thinks” about an answer: the higher the level, the more thorough the response, but the longer it takes and the more it costs. Anthropic emphasizes that on Low and Medium, Fable 5.1 delivers results on par with Fable 5 or better, but significantly cheaper — meaning the upgrade pays off even if you’re not chasing maximum quality.

The default effort level depends on the product: in Claude Code — High, in Claude Cowork and on claude.ai — Medium. In the claude.ai interface, the level is switched manually in the menu next to the send button. An important detail: in Fable 5.1, thinking is always enabled at any effort level, and it cannot be turned off.

Technical specs: context window — 1 million tokens, maximum output — 128,000 tokens.

Where it is Available and on Which Plans  

For developers, Fable 5.1 is available via the Claude API under the identifier claude-fable-5-1, and through cloud partners: Amazon Bedrock (anthropic.claude-fable-5-1), Google Cloud Vertex AI, and Microsoft Foundry.

In Claude apps, the model is available on all paid plans — Pro, Max, Team, and Enterprise — on the web version, in the mobile and desktop apps, as well as in Cowork, Claude Code (version 2.1.250 and newer), Claude Design, Claude for Microsoft 365, and Claude Tag. There are no Fable models on the free plan.

The billing scheme for subscribers depends on the plan.

1. On Max and premium seats in Team and Enterprise, Fable 5.1 is included in the plan — you can spend up to 50% of your weekly usage limit on it at no extra cost.
2. On Pro and standard seats in Team and Enterprise, the model runs on usage credits — charges for usage beyond the plan limits.

Unlike Fable 5, which had a free-access window in subscriptions over the summer, there is no promo or one-time credit for version 5.1 — it runs on the standard scheme from day one. 

How to Test the Model: Building a 3D Aircraft Engine Model  

We will put Fable 5.1’s claimed strength — the ability to build complete interactive applications — to the test with a clear, hands-on scenario. We will assemble an interactive 3D cutaway model of a turbofan aircraft engine, with rotation, thrust control, and labeled components. We won’t write a single line of code by hand — the model will generate all the logic, geometry, and UI on its own.

You will need any paid Claude plan.

Step 1: Choose the model and effort level 

Create a new chat. In the menu next to the send button, select Claude Fable 5.1. In the same place, open Effort and set it to High or Max — for complex “single-pass” generation, this will deliver the best results.

Step 2: Describe the task

The technical brief fits into just a few lines — the model will fill in the rest on its own.

Prompt

Build a single-file interactive 3D model of a turbofan jet engine as an HTML artifact. Show the engine in cutaway view so the internal stages are visible: fan, low- and high-pressure compressors, combustion chamber, turbines and nozzle. Let the user rotate the engine with the mouse, and add a throttle slider that speeds up the rotating parts and changes the exhaust glow. Label each major section with a callout. Include a short info panel explaining how a turbofan produces thrust. Use only HTML, CSS and JavaScript in one file, no external assets. 

Fable 5.1 doesn’t write code in the chat — it assembles a working artifact and opens it in the preview window on the right. For complex generations, the model may “think” noticeably longer than usual: at high effort levels, that’s normal.

Step 3: Check the interactivity

Rotate the engine with your mouse to inspect the cross-section, and move the thrust slider — the rotating stages should speed up, and the glow at the nozzle cutaway should intensify.

Expand the info panel — the model adds a short explanation of how it works alongside the “hardware,” even though the prompt only had general wording about “how a turbofan produces thrust.”

Step 4: Polish the result 

The strong point of Fable 5.1 is refinement without losing context. Ask it to add details in a single sentence.

Prompt

Add a second view mode that shows the airflow as animated streamlines splitting into the bypass and core paths, with a toggle to switch between the cutaway and airflow views. 

Result:

The model writes the new mode directly into the same artifact, without breaking the already working 3D scene.

Privacy for Business: Enterprise Frontier Safeguards 

Alongside the model, Anthropic introduced Enterprise Frontier Safeguards (EFS) — a system designed to resolve the long-running tension between privacy and abuse prevention. It gives enterprise customers zero data retention–level confidentiality, while still preserving abuse-detection mechanisms.

Here’s how it works: data is stored in cloud infrastructure fully controlled by the customer, not Anthropic, under the customer’s own encryption keys. Abuse checks are automated, and any incident review is handled by the customer by default. EFS is an optional feature that affects neither model behavior nor pricing.

The system was developed with more than 100 companies across finance, healthcare, manufacturing, telecom, legal, retail, and the public sector, as well as cloud partners AWS, Google Cloud, and Microsoft Azure. EFS will be supported in Claude Code, Claude Enterprise, on the Claude Platform, in Amazon Bedrock, Google Agent Platform, and Microsoft Foundry. The rollout will be phased, starting this fall. Until it is ready, eligible customers will be able to use Fable 5.1 with zero data retention.

Recall that the mandatory 30-day data retention for business traffic was one of the reasons some companies avoided Fable 5. 

More Accurate Safety Mechanisms 

Anthropic retuned Fable 5.1’s classifiers so they interfere less often with harmless work. In cybersecurity, false positives dropped by about 60%: in Claude Code, users should see roughly that many fewer interventions per session on average. What’s more, Fable 5.1 is now allowed to search for vulnerabilities in code — that is, to do defensive work to improve software security — although exploit development, pentesting, and binary vulnerability analysis are still routed by the classifiers to Opus models.

In biology, the updated safety mechanisms for Fable 5.1 and Fable 5 trigger 85% less often on harmless queries about basic biology and medicine. Requests at the level of scientific research and R&D in the life sciences are still routed to Opus.

The routing mechanism remains the same as in the original Fable 5: if a classifier triggers, the request in the same chat is reprocessed on an Opus model — currently Opus 5 for biology, chemistry, and life sciences, and Opus 4.8 for offensive cyber tasks. The user sees a notification about the switch, after which the selector stays on Opus until the end of the conversation. 

Separately, Anthropic strengthened protections against distillation — the large-scale “siphoning” of a model’s capabilities to train competitors. New API accounts (created starting on release day) can no longer manually edit Claude’s prior context while preserving a transcript of its reasoning — closing off a common technique for extracting the model’s “thoughts.” 

Limitations and What to Watch Out For 

The main caveat is the same as with the original Fable 5 — safety classifiers. They have been made more precise, but they have not gone away: a significant share of prompts related to virology, toxicology, drug and molecule design, as well as offensive cybersecurity, still get routed to Opus, so Fable 5.1 is not recommended for professional biology and drug development. The checks analyze not only the user’s message, but also everything the model reads — memory, data from connectors, web search results, and files — so a block can be triggered by content you did not write yourself.

The second point is availability and subscription pricing. There is no free window for version 5.1 like Fable 5 had in the summer: on Pro and standard Team seats, the model immediately consumes usage credits, and there is no one-time compensatory credit for it either. 

The third is Mythos 5.1. Full capabilities in cybersecurity and biology are available only to vetted participants in trusted access programs — the Cyber Verification Program for cyber defenders and the Life Sciences Verification Program for life sciences specialists (the latter Anthropic is developing in partnership with the U.S. government). For now, access is open only to a limited set of U.S. organizations, but the company says it will expand it, including to international partners. Claude Security is already running on Mythos 5.1 — it scans codebases for vulnerabilities and suggests patches for human review.

Finally, keep the watermark in mind. Under the European Union’s Code of Practice on AI content transparency, Anthropic embeds an invisible statistical marker into the text responses of models released after August 2, 2026. It does not affect answer quality and contains no user data, but it is formally present in everything Fable 5.1 generates.

What the Fable 5.1 Launch Changes for Users 

Fable 5.1 is a telling release: instead of chasing new records “at any cost,” Anthropic is, for the first time, so explicitly balancing capability, cost, and usability. The model is stronger on complex, long-running tasks, but the real news is not the benchmarks — it’s that, thanks to cheaper cache and flexible effort levels, the same work now costs a quarter less, and sometimes almost half as much. This is a direct response to businesses migrating to cheaper solutions.

One question remains open — can the revamped economics bring Fable back into the set of models businesses choose by default, rather than only for the most demanding scenarios.