Anthropic's Claude Fable 5.2 Coordinates 1,000 AI Agents to Build an Operating System in 8 Days

The race to improve large language models (LLMs) has reached a stage where benchmark scores alone no longer show how useful a model is in practice. A model with advanced reasoning ability cannot necessarily complete long-running work on its own, such as migrating complex software or running scientific simulations. And when AI agents are given access to development environments and external systems, it also matters how far security risks can be contained.

Claude Fable 5.2 and Claude Mythos 5.2, announced by Anthropic on November 9, 2026, are AI models built with these challenges in mind. Both use the same foundation model, but their safeguards and terms of availability differ. Fable 5.2 is available to ordinary developers and businesses, while Mythos 5.2, which lifts some restrictions in cybersecurity and biology, is offered only to organizations that pass a strict vetting process. Anthropic has laid out an approach in which a single model with advanced reasoning ability is deployed differently depending on purpose and risk.

The announcement included benchmark results that beat competing models, along with results from software development and mathematics research in which more than 1,000 AI agents worked together. API pricing, however, is higher than for the previous mainstay models, and there are constraints on data retention and safety. Based on the published technical specifications and evaluation results, this article examines the models' capabilities and the challenges of actually adopting them.

One Foundation Model, Deployed Differently by Safeguards and Access Conditions

Fable 5.2 and Mythos 5.2 use the same foundation model, which supports a 1-million-token context window and up to 128,000 tokens of synchronous output. The knowledge cutoff is June 2026, and "Adaptive Thinking," which dynamically adjusts the depth of reasoning, is always on. The default reasoning effort is "high." The model identifier for Fable 5.2 on the Claude platform is claude-fable-5-2, and it is available not only on Anthropic's own platform but also on Amazon Web Services (AWS), Google Cloud, and Microsoft Azure.

The base price for Fable 5.2 is set at $8 per million input tokens and $40 per million output tokens. That is double Claude Opus 5.5 ($4 input, $20 output) and four times Claude Sonnet 5.5 ($2 input, $10 output). Using the optional "Fast mode," which raises token generation speed by up to 2.5 times, pushes the price to $16 for input and $80 for output. Even in standard mode, the model is said to be up to 30% faster than Fable 5.1, but all of these are Anthropic's own measurements, and the same performance will not necessarily be achieved under every usage condition. The publicly available materials we could find also give no specific absolute figures for the token generation speed of the models used for comparison.

Another major feature is the separation of safeguards and access permissions. The general-purpose Fable 5.2 incorporates classifiers that assess safety in cybersecurity and biology. Requests that could lead to misuse, such as those related to penetration testing, exploit generation, and virus gain-of-function research, are blocked automatically. Meanwhile, organizations that need advanced capabilities for legitimate purposes, such as vulnerability research and drug development, can be vetted through the "Cyber Verification Program" and the "Life Sciences Verification Program" and given access to Mythos 5.2, which lifts the relevant safety restrictions.

Under this system, the functions actually available depend not only on the model's own capabilities but also on which safeguards are applied and who is permitted to use it. Rather than differentiating products by performance gaps between models, as in the past, it is a scheme that offers the same model according to purpose and risk. However, for users whose work does not touch the restricted areas, such as cyber defense or biological research, the difference between the two may not matter in practice.

Ahead in Coding and Scientific Research on Benchmarks, but Not First in Every Category

According to benchmark results published by Anthropic, Fable 5.2 shows high performance in agentic coding, autonomous scientific research, and other areas. However, competing models come out ahead on some evaluation items, and measurement conditions differ. Comparisons need to take into account reasoning effort settings, the tools available, and whether safety restrictions are in place.

The gap was particularly large in evaluations of software development by AI agents. On FrontierCode 2.0, Fable 5.2 scored 49.1% using "Xhigh," the highest reasoning effort level, ahead of Opus 5.5 at 45.5% and GPT-6 Astra at 42.8%. It also reached 69.9% on FrontierSWE v2, an improvement of 13.6 points over Fable 5.1's 56.3%. These results suggest improved performance on development work involving multiple steps.

The model did not post the top score in every area, however. On AutomationBench, which evaluates office-work automation, Fable 5.2 scored 51.4% against Gemini 4 Carbon's 53.9%. On AA-Briefcase v1.1, which evaluates practical skills such as document creation, Fable 5.2 led on the rubric-based evaluation, but on the Elo rating Opus 5.5's 1800 exceeded Fable 5.2's 1785. The rankings between models change when the evaluation method changes.

Differences in measurement conditions also need attention. For example, on CyScenarioBench each lab uses a different task set, so the figures cannot be compared strictly. The 64.9% recorded on RoboDojo Sim, which evaluates robot control, is also the result of an in-house test simulating object grasping with a dual-arm robot on NVIDIA Isaac Sim. It does not prove how well real robots can handle changes in a physical environment.

These results suggest that Fable 5.2 is less a model that overwhelms competitors in every use than one that shows its strengths in complex coding, reasoning in science and mathematics, and work that involves long periods of trial and error. How much of that advantage comes from the model itself, and how much from reasoning effort or the allocation of compute, remains a question for future verification. It will also matter whether the performance gap holds when third parties compare the models under the same conditions.

Per-Token Price Is Double, but Total Task Cost May Be Halved

Fable 5.2's API pricing is double that of Opus 5.5, but that does not translate directly into doubled development costs. The total cost of work done with an AI model depends not only on the per-token price but also on the number of tokens consumed and the number of retries before a task is completed. If a high-performance model can solve a problem in fewer attempts, total costs may stay lower even at a higher unit price.

Anthropic itself recommends starting with Opus 5.5 for everyday coding and general knowledge work. It then suggests switching to Fable or Mythos when the scope of work is very broad or when more advanced, abstract judgment is required. The idea is to match models to task difficulty rather than leaving every process to the top-tier model.

{
  "type": "bar",
  "title": "FrontierCode 2.0における達成スコアとタスク完了コストの比較",
  "unit": "スコアは%、コストは相対値（Opus 5.5 Max = 100）",
  "categories": ["Opus 5.5（Max effort）", "Fable 5.2（High effort）"],
  "series": [
    { "name": "達成スコア（%）", "values": [45.5, 45.3] },
    { "name": "タスク当たり総コスト（相対値）", "values": [100, 50] }
  ],
  "caption": "異なる単位の指標を併記した概念比較。Anthropicの初期テストに基づく。タスク当たりの費用の絶対値は公表されていない",
  "source": "Anthropicシステムカード"
}

As an example supporting this approach, Anthropic cites a cost comparison on FrontierCode 2.0. Fable 5.2 at "High" effort scored 45.3%, roughly the same as the 45.5% of Opus 5.5 running at "Max" effort. Total cost per task for Fable 5.2 High, however, was said to be about half that of Opus 5.5 Max. The publicly available materials we could find do not give specific dollar amounts for the per-task cost of either model.

The reason cost can stay down despite a doubled token price is that the model may reach the right answer with less trial and error and reduce the need to reload context partway through. However, this comparison is Anthropic's own measurement using different reasoning efforts, and it does not guarantee that costs will be halved in ordinary work. If Fable 5.2 is always run at maximum effort, costs could also rise substantially.

Economical use of Fable 5.2 therefore depends on "model routing," which assigns each task to an appropriate model. One possible setup, for example, leaves routine work to Haiku 5.5 or Sonnet 5.5 and calls on Fable 5.2 as an "Advisor" only when a complex problem arises. Another is to use Fable 5.2 as an "Orchestrator" that manages multiple AI agents and delegates routine processing to smaller models. The design confines use of the expensive model to a narrow scope while still drawing on its advanced judgment.

Usage conditions also differ by pricing plan. Fable is included in the standard subscription allowance only on the top-tier Max, Ultra, and Team premium seats, and on legacy Enterprise premium seats. On the Pro plan and standard Team seats, pay-as-you-go credits are required, as with the API. In addition, the expanded allowance for Pro and standard Team users from November 9 to 16, 2026, which allows use of up to 50% of the weekly limit, is a limited-time measure and not a permanent term of service.

More Than 1,000 AI Agents Work Together, From OS Development to Mathematics Research

One of the most notable demonstrations of Fable 5.2's capabilities is a set of long-running task experiments in which many AI agents worked together. Anthropic reports results from large-scale autonomous work that was difficult for earlier models, in fields including software development, mathematics, physics, and hardware design. Rather than a single AI simply answering questions, multiple agents divide roles and advance development or research while fixing problems that arise along the way.

According to Anthropic, Claude is involved in more than 90% of the changes merged into the company's core codebase. An agent team using an early version of Fable 5.2 completed the work of porting a service that plays a critical role in production from Python to Rust. The result was reported to be a startup speed about 26 times faster. However, the absolute startup times before and after the port have not been published, and this is only a success story from inside Anthropic. It should be noted that it is not a benchmark result in which a third party compared the two under the same conditions.

An experiment was also run in which the model was asked to build its own OS from an empty directory. Fable 5.2 mobilized more than 1,000 sub-agents and continued writing, verifying, and fixing code over eight days. As a result, it reportedly built a prototype OS that actually boots on a laptop. The OS is said to include its own kernel, as well as drivers for storage, keyboard, and networking, a window manager, an installer, and even features that incorporate Claude.

{
  "type": "bar",
  "title": "フェルマーの最終定理の形式化における所要時間と達成度",
  "unit": "所要時間は時間、進捗率は%",
  "categories": ["初期研究モデル（数十エージェント）", "Fable 5.2（1,000エージェント）", "Opus 5.5（1,000エージェント）"],
  "series": [
    { "name": "形式化完了までの時間", "values": [264, 30, null] },
    { "name": "30時間時点の進捗率（%）", "values": [null, 100, 26] }
  ],
  "caption": "初期研究では人間の介入を伴い11日間を要した。Opus 5.5は18時間時点で26%に達した後、進捗が停滞した。所要時間と進捗率は異なる単位",
  "source": "Anthropic公開技術レポート"
}

In mathematics, too, results from large-scale agent coordination have been reported. The target was the formalization of "Fermat's Last Theorem." This is the work of converting a mathematical proof into a formal language in which a computer can verify its correctness, and for complex theorems it takes enormous effort. In the first study Anthropic reported in September 2026, a dedicated framework called "Prove2Me," which coordinated dozens of Claude agents, was used together with human intervention, and it took 11 days (264 hours) to complete.

By contrast, Anthropic reports that a team of 1,000 agents using Fable 5.2 completed the formalization in 30 hours, working from only a minimal initial prompt and a progress list of intermediate theorems. Opus 5.5, using the same 1,000-agent configuration, advanced to 26% in 18 hours but then stalled, making almost no further progress even after 30 hours. The Lean-format proof produced in the earlier study was verified for consistency across three systems: Lean's official kernel, a comparator, and the independent "nanoda" kernel. However, the 30-hour record for Fable 5.2 is the result of a single trial reported by Anthropic, and its reproducibility has not been confirmed through repeated experiments.

In theoretical physics, work on a computational problem that had gone unsolved for many years was also reported. A study using holographic duality tackled the question of whether black hole formation through gravitational collapse in a small universe occurs abruptly. According to Anthropic, the problem had seen little progress since 2003 because of its heavy computational demands, and the model completed the calculation and showed that an abrupt transition does occur. In biology, the model also worked on predicting the capsid structure of adeno-associated virus (AAV), which is expected to be used as a vehicle for gene therapy. By combining existing open-source models, it reportedly modeled the overall structure with high accuracy for 28 of the 34 natural and artificial capsids tried.

In hardware development, working with the circuit-board design company Diode, the model designed three types of electronic devices: a voice assistant, a USB speakerphone, and a handheld pet-raising toy. It completed everything from routing a four-layer printed circuit board and modeling a 3D-printed enclosure to writing embedded software in less than a day, and the boards it had manufactured reportedly worked without electrical modifications. The case shows the possibility of using AI agents in development processes that involve electronic circuits and mechanical structures as well as software.

These results suggest that, in assessing AI progress, what matters is not only the accuracy of individual answers but also whether multiple agents can work together to complete long-running work. However, Anthropic's experimental results cannot be taken as direct evidence of reliability or safety in ordinary development environments. The conditions of the experimental setup affect the results, including the compute needed to run many agents, the dedicated execution and management infrastructure, and how failed attempts are handled. How closely ordinary developers and companies can reproduce similar results is not yet known.

Safeguards Strengthened While Unnecessary Blocks Fall by 55%

The more capable AI agents become, the more important risk management becomes when they are given access to development environments and production systems. In particular, there is a need to address both the problem of wrongly blocking legitimate work and the problem of AI performing operations it has not been permitted to perform. For Fable 5.2, the safeguards that prevent inappropriate behavior were revised so as to limit the impact on everyday development work.

According to Anthropic's early data, unnecessary request blocks fell by about 55% overall in Fable 5.2 compared with Fable 5.1. The decrease was reportedly about 60% for cybersecurity-related requests and about 40% for biology-related ones. However, specific absolute figures for the block rates of the two models before and after the improvement have not been published. In addition, when a request is refused by a safety classifier, a mechanism automatically switches cybersecurity-related tasks to Opus 4.8 and biology-related tasks to Opus 5, so that processing continues on a more restricted model.

To examine the safety of autonomous agents, tests of sandbox-escape behavior were also conducted. Anthropic prepared more than 1,800 test environments with deliberately placed vulnerabilities, in which a task could not be accomplished without accessing something outside the isolated environment. The rate at which Mythos 5.2 tried to use credentials or files outside the sandbox was 3.1%. That is a sharp drop from 39% for Mythos 5 and 21% for Mythos 5.1, and it is also below Opus 5.5's 3.3% and Sonnet 5.5's 5.1%.

However, the 3.1% figure does not mean problematic behavior has disappeared entirely. It is only an attempt rate in an artificially designed test environment, and it does not guarantee that sandbox escapes can be prevented in real systems. In a test examining the tendency to take harmful actions after wrongly believing it had been placed in a simulated environment, Mythos 5.2 performed slightly worse than Opus 5.5 and stayed at the same level as Mythos 5.1.

Another factor that makes safety evaluation harder is the increasing frequency with which models suspect they are being tested. If an AI infers that it is in an evaluation environment and changes its behavior accordingly, it becomes hard to judge how well test results reflect behavior in actual operation. Anthropic has also been evaluated by external organizations including the UK AI Security Institute (UK AISI), Frontier Design, and METR, but it directs readers to the system card for the details of that verification.

Anthropic also deliberately set limits on research areas that bear directly on improving AI's own capabilities. Specifically, it says it adjusted training so that capabilities that accelerate AI development itself, such as optimizing kernels for AI accelerators and improving next-generation LLM architectures, stay at roughly the level of Opus 5.5. General coding and machine learning research abilities are maintained, but for some tasks, such as developing kernels for certain non-standard accelerators, it also uses a mechanism that forcibly switches to Opus 5 at runtime. The measure is intended to curb the risk of advanced AI rapidly expanding its own capabilities.

30-Day Data Retention Poses a Challenge for Enterprise Adoption

When companies use Fable 5.2 or Mythos 5.2, they need to check the rules on data retention and monitoring as well as reasoning performance and pricing. For companies that handle confidential information in particular, where input data and AI responses are stored and who can access them can decide whether to adopt the models.

Misuse of advanced AI agents may unfold gradually across multiple exchanges rather than in a single request. For that reason, Anthropic has adopted a policy of retaining usage data from the previous 30 days, in principle, for safety monitoring of Fable and Mythos use. This condition can be a barrier to adoption for companies that require "Zero Data Retention (ZDR)."

Provided to address this problem, and offered as a private beta, is "Enterprise Frontier Safeguards (EFS)." Under EFS, prompt and response logs are stored within infrastructure managed by the customer, and the automated monitoring process that checks for safety is also run in the same environment. Because patterns of terms-of-service violations can be detected from a certain period of usage history without Anthropic personnel directly viewing the data, the company says it can be made compatible with ZDR policies.

However, EFS is not yet generally available. It is at the limited beta stage, targeting some customers such as Bridgewater, Salesforce, and Stripe. In addition, existing Enterprise customers who meet the conditions are offered a grace measure that lets them temporarily keep ZDR until EFS is introduced, but not every company can use it unconditionally. Companies with strict data retention rules need to check their eligibility for EFS and the timing of its introduction.

From the standpoint of protecting intellectual property, countermeasures against "distillation attacks," in which an AI model's responses are collected and used to train another model, have also been strengthened. To prevent actions that alter past context to improperly extract information about the reasoning process, Anthropic has expanded the application of "Preserved Thinking" across the API. According to the company, this measure reduced the amount of data that can be improperly collected by 90 to 99%. However, specific absolute amounts of data that could be obtained before and after the measure have not been published.

Five Points to Check Before Adopting Fable 5.2

When considering Fable 5.2, it is not appropriate to judge on benchmark rankings or per-token prices alone. Companies need to verify how much time and cost the model can actually save in real work, and whether it meets their safety and data management requirements. The main points to check are the following five.

1. Compare models on real work: Using your own codebase and business processes, compare the performance of Opus 5.5, Sonnet 5.5, and Fable 5.2. Match conditions such as reasoning effort, and measure total token consumption, time required, cost, and success rate through to task completion.
2. Design how models are matched to tasks: Rather than using Fable 5.2 for every process, leave routine work to Sonnet or Opus. Use Fable as an Advisor or Orchestrator only where advanced judgment is needed, such as investigating complex failures or designing an entire system.
3. Monitor safety classifier blocks and model switching: Check whether legitimate programming and analysis requests are being wrongly blocked. When automatic switches to Opus 4.8 or Opus 5 occur, record how often they happen and how they affect work.
4. Confirm data retention and audit requirements: Verify whether 30-day retention of monitoring logs meets your company's legal and security requirements. If ZDR is required, check EFS eligibility and adoption conditions in advance.
5. Consider Mythos for advanced security work: For work such as vulnerability research and penetration testing that may conflict with Fable 5.2's safety restrictions, consider obtaining legitimate access through the Mythos 5.2 vetting programs.

The announcement of Fable 5.2 and Mythos 5.2 shows that, beyond improvements in model performance, how to manage AI with advanced reasoning ability and build it into real work has become important. The coordinated work by more than 1,000 agents and the cost reductions on complex development tasks are notable, but for now much of the evidence rests on Anthropic's own evaluations. Going forward, along with third-party verification of reproducibility, operational design covering model routing, safeguards, and data management will determine the real benefits companies get from adoption.