Claude Fable 5.1 Officially Released: Anthropic Usher in a New Era of “Long-Term Agents” for AI

Three months ago,AnthropicWith the release of Fable 5, we sought to answer one question: When a model needs to work continuously for hours or even days, can it consistently remember its goals, call upon tools, check results, and complete the task without supervision?

Now, Anthropic is back with Fable 5.1.

Fable 5.1 and Mythos 5.1 Officially Released

On September 1, local time, Anthropic officially released Claude Fable 5.1 and Claude Mythos 5.1.

According to the official definition, both use the same underlying model; the difference lies not in “intelligence” but in the level of security protection: Fable 5.1 is fully available to the general public and businesses, while Mythos 5.1 is provided only to select cybersecurity and life sciences organizations through a Trusted Access Program.

Judging by the version number, this is just a “.1” update. However, judging by the actual changes, Anthropic is clearly doing more than just patching Fable 5. The new model outperforms its predecessor by a factor of two in scientific research agent testing and shows comprehensive improvements in programming, knowledge work, computer operations, and long-duration tasks; at the same time, the actual invocation cost of Agent workloads has decreased by up to approximately 451 TP4T.

More importantly, this time around, Anthropic didn’t just focus on model benchmarks. Topics such as pricing, caching, data retention, false positives, access to vulnerability research, and enterprise deployment architectures accounted for nearly half of the announcement.

The relationship between Fable 5.1 and Mythos 5.1 is the first key to understanding this release; they are two different deployments of the same model.

Fable 5.1 introduces cybersecurity and biosecurity protection mechanisms designed for both public and enterprise environments. When a request triggers a risk rule, some tasks are blocked, while others are routed to the relatively less capable Opus model. Mythos 5.1 retains more of its original capabilities but is accessible only to vetted organizations.

In other words, Anthropic is attempting to manage “model capabilities” and “model openness” separately: rather than directly weakening the entire model to mitigate risk, it provides different levels of capability based on user identity, task type, and usage context.

What are the key highlights?

So what are the highlights of these two “legendary” models?

In fact, the key changes in Fable 5.1 can be summarized as follows: higher capacity limits and cheaper cache reads; however, with inference intensity maxed out, completing a single task may still be more expensive than in the previous generation.

On Terminal-Bench 4.0, Fable 5.1 scored 55.81 TP4T, while Mythos 5.1, with some restrictions disabled, reached 60.91 TP4T. The performance gap between the two is largely not due to differences in their underlying models, but rather the result of Fable’s safety system intervening in certain cybersecurity tasks.

Among the data sets released by Anthropic, the most significant changes were observed in agent-based scientific research.

On Terminal-Bench-Science 0.1, Fable 5.1 achieved a score of 52.61 TP4T, while Fable 5 scored only 24.71 TP4T. Opus 5 scored 29.01 TP4T, and GPT-5.6 Sol scored 22.41 TP4T. Compared to its predecessor, Fable 5.1’s score is not just a slight improvement—it is more than double the original score.

Terminal-Bench-Science 0.1: The standard error for each model is ±3.5–4.5 points. The public leaderboard (3 trials per task, using the Claude Code framework) shows that Claude Opus 5 achieved an accuracy of 30.01 TP4T, while Claude Fable 5 achieved an accuracy of 21.41 TP4T; Our experimental setup replicated these results, yielding accuracy scores of 29.01 TP4T and 24.71 TP4T, respectively—both within the margin of error.

These figures indicate that the improvements in Fable 5.1 are not limited to code generation. It outperforms its predecessor in computer operations, cross-domain reasoning, research tasks, and knowledge work, and also outperforms the lower-cost Opus 5 in most official benchmarks.

Its performance in knowledge tasks also reached the top, though the gap between it and Opus 5 was not significant. Fable 5.1 achieved Elo ratings of 1,853 and 1,694 in GDPval-AA v2 and AA-Briefcase, respectively, representing increases of 130 and 122 points over its predecessor.

However, its lead over Opus 5 in the former category remains within the confidence interval, while in the latter category, it is essentially on par with Opus 5’s score of 1,685. Specifically, its advantages are primarily concentrated in analytical quality and compliance with scoring criteria; its presentation quality does not show a corresponding lead.

The price adjustments primarily address the cost associated with the Agent repeatedly reading the context. In Fable 5.1, the price for cache reads has been reduced from $1 to $0.25 per million tokens, a decrease of 75%; The prices for standard input and output remain at $10 and $50 per million tokens, respectively, while the cache write price remains at $12.50. The context window remains at 1 million tokens, and both text and image input are supported.

However, the price reduction for caching does not mean that all tasks are more cost-effective. At the highest inference intensity, Fable 5.1 costs an average of $3.76 to complete a single intelligence index test task—approximately 201 TP4T more than Fable 5—because the number of output tokens is about 1.7 times that of the previous generation. Even though the cache price reduction saved approximately $1.40 per task, it did not fully offset the cost increase resulting from the higher token output.

Reducing the inference intensity improves cost-effectiveness: The xhigh tier has an intelligence score of 65, and the cost per task drops to $2.72, though this is still higher than the $2.34 for the highest tier of Opus 5, which has a score of 63. Fable 5.1 occupies the high-end Pareto frontier in terms of both intelligence level and token output efficiency, meaning it is more efficient in the high-capability range, though this does not equate to the lowest cost per dollar.

The value of this upgrade thus becomes clear: Fable 5.1 raises the upper limit for complex tasks without eliminating the trade-off between capabilities, computational resources, and cost. For developers, the real question remains whether the additional expense will enable them to accomplish tasks that were previously impossible.

No longer content with just “fixing code,” but instead getting to the root cause

In its press materials, Anthropic repeatedly emphasized that Fable 5.1 will reduce “shortcuts”—solutions that appear quick but actually compromise quality—and instead focus on identifying the root causes of problems.

The investment firm Millennium provides a representative example.

A section of internal code in the system had a crash that occurred about once every million runs. Engineers spent four or five years trying to figure out the cause but were unable to explain it, and other models—including Fable 5—were also unable to resolve the issue.

The following chart shows how Fable 5.1 compares across various benchmark tests:

Fable 5.1 was evaluated with its production security safeguards enabled. In tasks where these security safeguards were activated, both Fable 5.1 and Fable 5 scored zero on the OSWorld 2.0 test, and Fable 5 also scored zero on the AutomationBench test. With all other security measures enabled, the cybersecurity tasks were completed by Claude Opus 4.8, and the biological tasks were completed by Claude Opus 5. This likely resulted in reduced performance for Fable 5.1 and Fable 5 in these benchmark tests.

Fable 5.1 analyzed the library provided by an external vendor, compared the disassembly results with the kernel dump file, and ultimately pinpointed the issue as a defect in that external library.

While this case study, of course, comes from one of Anthropic’s early partners and cannot replace independent testing, it illustrates the distinction that Fable 5.1 aims to make: it’s not about generating more code faster, but rather about whether it can, like a senior engineer, systematically rule out hypotheses layer by layer within complex systems to ultimately identify the true source of the problem.

The tests conducted on CodeRabbit, an external code review platform, offer a relatively objective perspective.

Out of 45 evaluation tasks and 105 known issues, Fable 5.1 identified 64, while Fable 5 identified 65; there was virtually no change in recall between the two. However, the number of reviews ultimately generated by Fable 5.1 decreased from 253 to 166, and the number of trivial comments dropped from 265 to 79, while the precision rate rose from 32.8% to 37.3%.

In other words, while it didn't uncover any additional issues, it significantly reduced the number of invalid and nitpicky comments, and its output became more restrained.

The trade-off is speed. On average, Fable 5.1 takes 18 minutes and 38 seconds to complete a single evaluation, while Fable 5 takes 12 minutes and 32 seconds—an increase of nearly 491 TP4T. Furthermore, in this set of tests, increasing the inference intensity did not yield better results: the "High" mode was slower than the "Low" mode and actually produced a lower recall rate.

This means that Fable 5.1 isn’t the default choice for all programming tasks. For routine, small-scale pull requests, it may be too expensive and slow; however, for complex tasks that span multiple files and require an understanding of the entire codebase, the longer analysis process may prove valuable.

The unit price hasn't gone down, so why is Agent 45% cheaper?

The standard API pricing for Fable 5.1 remains unchanged at $10 per million input tokens and $50 per million output tokens.

The price for cache reads actually underwent a significant adjustment: it dropped from $1 per million tokens to $0.25, a decrease of 751 TP4T.

Anthropic estimates that the total cost for typical workloads will be approximately 251 TP4T lower than that of Fable 5; for highly agent-driven tasks, the cost reduction could reach up to approximately 451 TP4T.

A programming agent that runs for an extended period may repeatedly access the same set of code, system prompts, tool definitions, project specifications, and task history. If the tool recalculates the price based on standard input every time it is called, the longer the task runs, the higher the cost associated with the repetitive context.

Context caching allows the system to store previously processed context, incurring only a low read cost in subsequent operations. For agents that need to call tools dozens of times and run continuously for several hours, cache reads may account for the majority of their input volume.

Therefore, the recent price adjustment for Fable 5.1 is not a general price reduction, but rather a targeted revision aimed at Agent workloads.

This also explains why Cognition decided to migrate some of the traffic originally routed to Opus 5 to Fable 5.1 in Devin. Previously, while Fable-class models offered higher performance, it was difficult to demonstrate their cost-effectiveness for high-frequency tasks such as code reviews; following the cache price reduction, enterprises are no longer evaluating the price of a single token, but rather the total cost required to complete a single task.

The competition among models has thus shifted from “how much per million tokens” to “how much per qualified task completed.”

The real secret weapon might not be programming, but scientific research.

If one looks only at Claude Code, Fable 5.1 could easily be interpreted as yet another upgrade to the programming model. However, Anthropic devoted a significant portion of this release to showcasing its scientific research capabilities.

In molecular design experiments, Anthropic had Mythos 5.1 utilize open-source protein design and folding tools to generate protein ligands, which were then submitted to two external institutions for experimental validation.

Official data show that, across three targets, the binding affinity of the ligand designed by Mythos 5.1 is 10 times higher than that of the best solution from the Adaptyv Bio protein design competition; the hit rate for effective ligands across 12 targets is close to 50%, while Anthropic cites the typical industry benchmark as ranging from 10% to 15%.

In another project, Fable 5.1 used radar data collected more than 30 years ago by NASA’s Magellan spacecraft to train a neural network and generate a high-resolution elevation map of approximately one-third of Venus’s surface.

Previous maps could only show details on a scale of about 10 to 20 kilometers, but the new map increases the resolution to 2 to 3 kilometers and improves elevation measurement accuracy by up to 25%. Anthropic has released this map under a Creative Commons license, hoping it will serve as a reference for selecting observation targets for NASA’s VERITAS and the European Space Agency’s EnVision missions.

Mythos 5.1 also writes custom GPU kernels for seven open-source deep learning models for proteins and genomes and caches intermediate results, boosting inference speed by up to 2.5 times while maintaining exactly the same output. For certain whole-genome analysis tasks, GPU costs are expected to decrease by 301 TP4T to 601 TP4T.

The key difference between these cases and ordinary question-and-answer tasks is that the model’s outputs have moved beyond the textual context and are now being validated in the real world: proteins must actually bind in the laboratory, GPU kernels must produce exactly consistent results, and maps of Venus must be cross-validated against existing observational data.

Of course, this is still a long way from “AI making scientific discoveries on its own.” The model relies on open-source tools, data, evaluation methods, and experimental verification systems provided by humans, and its final conclusions still require review by scientists.

However, the shift in the model’s role is already evident: it is no longer limited to searching for papers and summarizing knowledge, but has begun to enter a closed-loop research process involving “proposing solutions—invoking tools—running computations—generating candidate results—and undergoing experimental validation.”

Should the data be retained or not? Anthropic has taken a different approach.

Fable 5 previously required, by default, that usage data be retained for 30 days to enable Anthropic to identify model abuse across time and accounts. While this policy aids in security monitoring, it has become a barrier to the adoption of these models in regulated industries such as finance, healthcare, and law.

Fable 5.1 also introduces Enterprise Frontier Safeguards, which aim to strike a balance between security monitoring and zero data retention.

Under this architecture, logs are stored in the customer’s own AWS, Azure, or Google Cloud environment, using keys, access policies, and audit systems managed by the customer. Anthropic provides automated risk detection, but risk flagging is handled directly by the customer; by default, no manual review by Anthropic employees is required.

Technically speaking, the data is still retained for a certain period of time to identify anomalous behavior across sessions and accounts; however, the actual control over the data and the authority to conduct manual reviews remain with the company.

Anthropic stated that the system was designed in collaboration with more than 100 companies, with participants spanning the finance, healthcare, manufacturing, telecommunications, legal, retail, and public sectors, and including several U.S. systemically important banks.

This may be one of the changes in Fable 5.1 that has more practical significance than benchmark scores.

When an agent can only answer questions, a 30-day data retention period may be nothing more than a single line in the privacy policy; but when the agent begins to access code repositories, financial data, customer information, and production system credentials, questions about where data is stored, who can view it, and how it is audited will directly determine whether the model can be deployed.

Fable 5.1 is really all about “high-value missions”

Fable 5.1 does not aim to be the cheapest or fastest general-purpose model.

The pricing of $50 per million tokens issued makes it impractical for a large volume of low-value Q&A. CodeRabbit’s testing also shows that it is willing to spend more time on complex tasks and is not suitable as the default model for all code reviews.

Anthropic is targeting a different kind of market: tasks that may run for hours, involve multiple applications, files, and tools, and would take days or even weeks to complete manually—yet are costly enough to cover the cost of the model.

Root cause analysis of software failures, modifications to complex contracts, financial research, cross-codebase refactoring, scientific computing, and extended browser use all fall into this category.

Therefore, the most significant change in Fable 5.1 is that Anthropic is simultaneously addressing three long-neglected areas of infrastructure: making caching costs suitable for long-running tasks, ensuring that safety systems minimize false positives, and keeping sensitive data within environments controlled by customers.

This is a competition that has shifted from “model performance” to “model deployability.”