Two versions of the same model with different access limits
Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, two versions of a single underlying model designed for complex technical tasks. The primary difference described by the company is not a distinct architecture, but the applied safeguards and who can access each version.
Fable 5.2 is generally available. Anthropic states it includes protections against misuse in cybersecurity and biology, and that these measures allow it to be offered to a broad audience. Mythos 5.2 is directed at organizations working on beneficial cybersecurity or biomedical research applications; they must request access through the company's verification programs.
General availability does not mean free access or identical conditions across all platforms. The announcement sets API pricing but does not detail the conditions of each Claude plan or separately confirm an API enablement date in the excerpt. Organizations considering Mythos must distinguish between submitting a request and receiving approval: requesting it does not equate to gaining access.
The division also matters for interpreting their uses. Fable includes safeguards that block most cybersecurity-related tasks, according to Anthropic; Mythos offers limited access to capabilities intended for verified professionals and organizations. This does not mean Mythos permits any operation or that every result is safe or correct. Access restrictions are an operational policy, not a guarantee that all risk disappears.
What changes between Fable 5.2 and Mythos 5.2
| Aspect | Claude Fable 5.2 | Claude Mythos 5.2 |
|---|---|---|
| Underlying model | Same model as Mythos 5.2 | Same model as Fable 5.2 |
| Described safeguards | Reinforced protections in cybersecurity and biology; Anthropic says they block most cybersecurity tasks | The company offers it to verified organizations for cybersecurity and biomedical research applications |
| Access | General availability; conditions and price depend on the product or channel used | Limited access via verification programs; the request does not guarantee approval |
| Announced price per million tokens | $8 per input token and $40 per output token | $8 per input token and $40 per output token |
| Optional fast mode | $16 per million input tokens and $80 per million output tokens | $16 per million input tokens and $80 per million output tokens |
Performance and pricing: figures communicated by Anthropic
Anthropic claims both models generate results up to 30% faster than Fable 5.1 and Mythos 5.1, respectively, and are priced lower than those previous versions. The announcement sets a standard rate of $8 per million input tokens and $40 per million output tokens. The optional fast mode costs $16 per million input tokens and $80 per million output tokens, and the company claims it can speed up generation by up to 2.5 times.
Maximum speed and savings do not describe the outcome of every request. Performance depends on the task, the configured effort level, and the platform; the announcement does not turn those figures into a guarantee of lower latency or lower cost in each use. Token prices alone are not enough to calculate the total cost of a workflow involving tools, agents, or human reviews.
In the evaluations Anthropic presents, Fable 5.2 achieves results superior to Opus 5.5 in almost all included capabilities, with variations between tests. For example, the company reports 49.1% on FrontierCode 2.0 with Xhigh effort and 69.9% on SWE-bench Verified v2. For Mythos 5.2, it reports 89.6% on its internal CVE Rediscovery evaluation and 76.3% on CyScenarioBench.
These values are not a universal comparison between models. CVE Rediscovery is an internal evaluation that, according to Anthropic, measures how many published vulnerabilities a model can find from a project version prior to its fix. The described run was performed in October 2026 with 1,829 CVEs. The company also warns that CyScenarioBench results are based on distinct tasks for each lab; therefore, a difference in scores does not necessarily equate to a comparison on exactly the same cases.
Examples of evaluations published by Anthropic
The scores belong to specific tests and reflect the methodology and configuration described by the company. They do not by themselves prove general superiority or predict performance in any real-world work.
| Test | Model and reported score | What it measures or what limit Anthropic declares |
|---|---|---|
| FrontierCode 2.0 | Fable 5.2: 49.1% at Xhigh effort | Evaluates whether a model's code changes would be accepted by maintainers of open-source repositories. |
| SWE-bench Verified v2 | Fable 5.2: 69.9% | Agentic coding evaluation; the figure corresponds to the test cited by Anthropic. |
| CVE Rediscovery | Mythos 5.2: 89.6% | Internal evaluation of published vulnerabilities; the run mentioned by the company used 1,829 CVEs in October 2026. |
| CyScenarioBench | Mythos 5.2: 76.3% | Anthropic notes that each lab uses its own tasks; scores are not necessarily based on an identical set of cases. |
What Anthropic claims about safety testing
Anthropic states it conducted pre-release safety and alignment testing on the model underlying both versions. The company developed evaluations to observe concerning behaviors, such as attempts to escape a sandboxed environment, and claims that Mythos and Fable matched or exceeded recent models on most of its alignment measures.
The company also holds that Fable 5.2's cybersecurity and biology safeguards are as robust as those of Fable 5.1 and generate fewer unnecessary blocks. It also describes its strongest-ever protections against distillation attacks. These are Anthropic's claims about its own systems and evaluations; they do not equate to a demonstration that the models cannot fail or be abused.
The System Card for Fable 5.2 and Mythos 5.2 is the documentation Anthropic identifies for expanding evaluation details. The summary announcement describes some conclusions and metrics, but those claims must be interpreted within the specific method of each test. A pre-release evaluation does not anticipate all possible failures nor prove by itself that safeguards function identically in all environments.
The access policy reduces who can directly use Mythos, but it does not replace an organization's own controls. A team evaluating these models must define permissions, supervision, action logging, and human review appropriate for its tasks. The ability to generate a response, analyze code, or propose a fix does not demonstrate that the result is safe, complete, or correct.
Technical applications and results presented as experiments
Anthropic describes uses of Fable 5.2 in coding, reasoning, visual analysis, engineering, and robotics. Among its examples, the company claims an experimental configuration coordinated over 1,000 subagents for eight days to create a prototype operating system that boots on a laptop. The result corresponds to an experiment communicated by Anthropic, not an independent test that the model can build a production-ready operating system by itself.
In mathematics, the company claims 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in 30 hours, with minimal human intervention. Anthropic compares that result with a previous formalization done with an unreleased research version, which took 11 days and used a specific support system and human guidance. The comparison, therefore, does not pit two models under identical conditions; the company also reports a single run per model for the cited graph.
In computational biology, Anthropic explains that Mythos 5.2 combined open-source protein prediction tools and known symmetries to model adeno-associated virus capsids. The company says the method succeeded in 28 of the 34 capsids it attempted to model. That result is limited to the structures tested and does not demonstrate that the procedure works with any protein complex or replaces experimental validation.
In hardware engineering, Anthropic recounts that the model designed three electronic devices and their boards, enclosures, and programs with the help of Diode, the company that manufactured the boards. The company holds that the three devices worked without electrical modifications. To interpret this example, it is useful to distinguish AI-assisted design from the material process of manufacturing, assembly, and testing, in which people and a collaborating company also participated.
Availability and criteria for evaluating its use
Anthropic offers Fable 5.2 generally and reserves Mythos 5.2 for organizations that request access through its Life Sciences and Cyber Verification programs. The company also indicates that Claude Enterprise customers can use Claude Security, a code analysis tool that incorporates Mythos; that product does not imply unrestricted access to the model.
Anthropic's help documentation identifies Fable 5.2 as available in Claude apps, Claude Platform, and certain cloud services, while Mythos 5.2 is listed as limited access for approved partners. Availability may vary by product and access conditions. The announcement date, November 9, 2026, is not enough to deduce that all integrations or accounts received access simultaneously.
To choose between versions, an organization can start from the task and the risk, not just a benchmark score. Fable 5.2 fits general technical uses within its safeguards; Mythos requires a compatible professional need and the corresponding approval. In both cases, local tests must measure quality, cost, latency, error rate, and effectiveness of controls with representative data and tasks.
Before incorporating a model into a critical process, teams can test normal and adversarial cases, check results with deterministic tools when possible, and maintain human review proportional to the impact. Access restrictions, fast mode, and a low temperature can influence behavior or operation, but they do not guarantee accuracy, valid format, security, or reproducibility.
Practical process for evaluating a model in an organization
- 01Define the specific task, expected outcome, and conditions under which an error could cause harm.
- 02Confirm which version, platform, safeguards, and access modality are enabled for the team.
- 03Prepare representative and adversarial cases; set quality, cost, and latency criteria in advance.
- 04Compare results with a verifiable reference and log errors, blocks, and need for human intervention.
- 05Limit permissions and automated actions; require deterministic controls or human review for high-impact operations.
- 06Repeat the evaluation when the model, configuration, tools, or data type change.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction