Fable 5.2 vs. Mythos 5.2: Anthropic's New Models Differ on Policy, Not Power

Anthropic reports that Fable 5.2 improves on its predecessor across science and coding benchmarks. The stated distinction between Fable and Mythos 5.2 is policy: Fable blocks some biology and cybersecurity requests, while Mythos has more relaxed guardrails. The cited account does not specify the boundaries of either model's policies. ² For companies choosing between them, checking what each model refuses matters alongside evaluating its capabilities.

What changed in Anthropic's Fable 5.2 and Mythos 5.2

Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, describing them as its most capable large language models to date. The company says their core capabilities are identical; Fable applies stricter limits to some biology and cybersecurity use cases, while Mythos has more relaxed guardrails. ²

Anthropic positions both for complicated, open-ended technical work and says Opus 5.5 is a better fit for routine workflow automation. ² That distinction is a reason to assess model behavior against the tasks a team plans to send, rather than choosing on capability claims alone.

Fable 5.2, Mythos 5.2, and Fable 5.1 compared

The cited account gives more detail about Fable 5.2 than Mythos 5.2: it reports a benchmark comparison against Fable 5.1, but no equivalent set of benchmark scores for Mythos. It also does not give a full specification of either model's policy boundaries. ²

Model	Reported capabilities and restrictions	Access and price
Fable 5.2	Shares the new models' core capabilities. Blocks most research-biology questions and many cybersecurity prompts, routing requests in those categories to Opus 4.8 or Opus 5. 2	The cited account does not specify access conditions. $8 per million input tokens and $40 per million output tokens. 2
Mythos 5.2	Same core capabilities as Fable 5.2, with more relaxed guardrails. 2	Limited to trusted researchers through two access programs; eligibility criteria are not stated in the cited account. $8 per million input tokens and $40 per million output tokens. 2
Fable 5.1	Predecessor used as the benchmark baseline. Anthropic also reports a comparison of erroneous refusals of safe prompts, but the cited account provides no fuller policy specification. 2	The cited account provides no price or access details. 2

Anthropic lists both new models at prices it describes as 20% below those of its previous-generation models. ² This is not a complete buying guide: the cited account does not specify Fable's access terms, Mythos's eligibility rules, or the exact boundaries of either policy. ²

What the Fable 5.2 benchmarks actually show

Anthropic tested Fable 5.2 against Fable 5.1 on 14 benchmarks and reported gains above 10% on most. On Terminal-Bench-Science 0.1, Fable 5.2 scored 74.6%, which Anthropic describes as a 22% improvement. Across the coding benchmarks FrontierCode 2.0 and FrontierSWE 2.0, Fable 5.2 scored more than 13% higher than its predecessor. ²

The account also describes two large-scale projects. In one, 1,000 agents parallelized work on a mathematical-proof workflow, reducing its reported runtime from 11 days to 30 hours. In another, 1,000 agents built an operating system in about eight days; the account says it could run on a laptop, access networks, and included a Claude sidebar. ²

These results do not establish what a single model call can do or what the agent runs cost. The cited account does not describe evaluation methods, datasets, error margins, or task-level benchmark breakdowns, and it provides no direct benchmark comparison between Fable 5.2 and Mythos 5.2. ² Whether the reported gains predict performance on a company's own tasks is not established by the cited results.

Analysis: Treat the benchmarks as a reason to run your own evaluation, not a substitute for one. The 1,000-agent examples combine model capability with orchestration; the account does not separate those contributions. ²

What Fable blocks, and what stays unclear about Mythos

Fable 5.2 blocks most research-biology questions and many cybersecurity prompts, routing requests in those categories to Opus 4.8 or Opus 5. Anthropic also says Fable 5.2 is 55% less likely than Fable 5.1 to wrongly refuse a safe prompt. ²

That figure is a reported comparison, not a guarantee about false-refusal rates on a particular company's prompts. The cited account does not define which individual requests are blocked, describe the routing logic in detail, or explain how borderline cases are handled. ²

For Mythos, the account reports a biology test in which the model accurately modeled 28 of 34 AAV capsids. It does not describe the test method, a comparison baseline, or whether the result reflects production conditions. ² Mythos access is limited to trusted researchers through two programs; the account does not state eligibility or application details. ²

For governance and security teams, practical questions include which use cases meet the models' policies, who can access Mythos, what audit evidence is available, and how rerouting affects oversight. The cited account does not answer those questions. ²

Cost and workflow implications

Both models are priced at $8 per million input tokens and $40 per million output tokens. Anthropic says a high-performance mode generates responses 2.5 times faster than the standard tier for double the price. The cited account does not specify availability terms or how that pricing applies across workload types. ²

Per-token rates do not predict what a multi-step workflow will cost. Anthropic's examples involved 1,000 agents, so token volume and orchestration overhead may matter alongside the listed rate. That is an inference from the examples, not a total-cost figure reported by Anthropic. ²

A separate release note says Fable 5.2 and Mythos 5.2 were added to the llm-anthropic plugin. The author reports that one maximum-reasoning Fable 5.2 run used 128,000 tokens without returning an image. The note does not establish whether the issue is repeatable or what caused it. ¹

How to evaluate the two models for enterprise use

Test capability and policy separately. For capability, run representative science and coding tasks on whichever model your organization can access. For policy, build a distinct test set covering requests you expect to be allowed, refused, redirected, and genuinely borderline. Treating that policy test as a gate before rollout is the same logic we describe in checkpoint safety as a release-gate question.

Before committing to either model in a workflow:

• Ask Anthropic for specifics — including biology and cybersecurity thresholds, redirect logic, Mythos eligibility, and methodology behind the headline benchmarks. The cited account does not provide those details. ²
• Measure outcomes, not just scores — task completion, refusal and redirection rates, review overhead, latency, token spend, and reproducibility against your own prompts. The published results do not establish how they transfer to your workload. ²
• Test the deployment path — reproduce image-generating workflows at the reasoning levels your application will use, including the maximum-reasoning case reported in the release note. ¹
• Log the decision conditions — policy assumptions, access dependencies, routing behavior, and expected cost, so the choice can be revisited if any of them change.

Analysis: Fable may be a practical default when its restrictions fit the work and the organization's governance posture. Mythos may be worth considering for teams that can obtain access and have a reason to use a model with looser guardrails. Neither choice is supported by a head-to-head performance comparison: Anthropic says the models share core capabilities, but the cited account reports no benchmark testing them against each other. ²

One practical note for teams that use more than one provider: when models differ in behavior, it helps to enforce your own limits rather than rely only on each vendor's policy. An enterprise AI gateway with hard routing constraints on allowed providers and models, plus a governed model catalog, lets you restrict which models a given workload can reach. That does not replace testing a model's refusal behavior, and it says nothing about which models any catalog contains.

What buyers should conclude now

The reported evidence shows Fable 5.2 outperforming Fable 5.1 on several science and coding measures. As Anthropic describes them, the practical distinction between Fable 5.2 and Mythos 5.2 is their safety posture and Mythos's restricted access, not a demonstrated capability difference. ²

Do not treat the benchmark and workflow claims as proof of performance on your own tasks. The cited account does not provide evaluation methods, error margins, or evidence of transferability to a company's workload. ² Keep the decision reversible: validate the work, test policy boundaries, confirm access, and record what would change your mind.

Frequently asked questions

What's the difference between Fable 5.2 and Mythos 5.2?

Anthropic says their core capabilities are identical. Fable blocks some biology and cybersecurity use cases; Mythos has more relaxed guardrails. ²

Is Mythos 5.2 publicly available?

It is limited to trusted researchers through two access programs. The cited account does not give eligibility requirements. ²

How much do Fable 5.2 and Mythos 5.2 cost?

Both are priced at $8 per million input tokens and $40 per million output tokens. A high-performance mode generates responses 2.5 times faster than the standard tier for double the price. ²

How much better is Fable 5.2 than Fable 5.1?

Anthropic reports gains above 10% on most of 14 benchmarks, a 22% improvement on Terminal-Bench-Science 0.1, and gains above 13% across two coding benchmarks. ²

What should companies test before choosing one?

Test representative workloads, refusal and redirection behavior, token consumption, latency, access requirements, and deployment workflows. The published results do not establish how these will map to your own tasks. ¹²