Claude Fable 5.2 and Mythos 5.2 are Anthropic models. Anthropic's own pages list lower per-token prices for Fable 5.2 than for Fable 5.1, and they report results from Anthropic's internal security, mathematics, biology and safety evaluations. This explainer sets out what those pages say about prices, who can use Fable 5.2 and the reported test results, and it marks where the supported material stops.
Claude Fable 5.2 pricing compared with Fable 5.1
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. For Fable 5.1, the same page gives $10 and $50. On that page, cache reads cost $0.20 per million tokens. The page also says US-only inference is available at 1.1x pricing for input and output tokens.
Fast mode has a separate price: $16 per million input tokens and $80 per million output tokens.
| Item | Price as listed on Anthropic's pages |
|---|---|
| Fable 5.2, input | $8 per million input tokens |
| Fable 5.2, output | $40 per million output tokens |
| Fable 5.1, input (previous price) | $10 per million input tokens |
| Fable 5.1, output (previous price) | $50 per million output tokens |
| Cache reads | $0.20 per million tokens |
| Fast mode, input | $16 per million input tokens |
| Fast mode, output | $80 per million output tokens |
| US-only inference | 1.1x pricing for input and output tokens |
These figures are taken only from the Anthropic pages cited above, and prices can change. This article does not compare them with prices from any other provider. The pages cited here also do not tell us what fast mode changes besides its listed price.
Who can use Claude Fable 5.2 and Mythos 5.2
Anthropic's Claude Fable product page says Fable 5.2 is available to Pro, Max, Ultra, Team and Enterprise users. This article does not judge whether any of these plans suits a particular reader.
The supported material for this article does not describe how Mythos 5.2 is made available. The Mythos 5.2 results discussed below come from Anthropic's announcement, which describes Anthropic's own testing of the model. Nothing in this article should be taken to mean that readers can get access to Mythos 5.2.
Mythos 5.2 security results in Anthropic's own evaluations
Anthropic reports two results on finding software vulnerabilities. Both are Anthropic's own measurements, published in Anthropic's announcement of the new Fable and Mythos models:
- CVE Rediscovery (internal evaluation): In Anthropic's own measurement, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1.
- October 2026 run: In Anthropic's own run against 1,829 CVEs, Mythos 5.2 outperformed Mythos 5.1 and Opus 5.5 at every effort level.
CVE Rediscovery is described as an internal evaluation. The cited material does not include outside replications of these numbers, so they are best read as figures reported by the company that built the model.
Fable 5.2 and other science results reported by Anthropic
Anthropic's announcement describes two pieces of work on formalizing Fermat's Last Theorem. Both accounts are Anthropic's own:
- Anthropic says its first formalization of Fermat's Last Theorem was done by an unreleased research version of Claude and took 11 days.
- Anthropic says a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours with minimal human intervention.
The two accounts are listed here as Anthropic reports them. This article draws no further conclusion from them.
The same announcement reports a biology result, again from Anthropic's own tests: Claude's capsid method accurately modeled 28 of the 34 capsids it attempted. The cited claim does not say which Claude model this result refers to, so this article does not link it to Fable 5.2 or Mythos 5.2.
Safety testing for Claude Fable 5.2 and Mythos 5.2
Anthropic's announcement includes several safety-related statements.
Sandbox evaluation (Anthropic's own measurement). In this evaluation, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of scenarios. The figures given for the other models are 3.3% for Opus 5.5 and 5.1% for Sonnet 5.5.
| Model | Share of scenarios using credentials or files outside the sandbox (Anthropic's evaluation) |
|---|---|
| Mythos 5.2 | 3.1% |
| Opus 5.5 | 3.3% |
| Sonnet 5.5 | 5.1% |
Safeguard blocks (Anthropic's early testing). Anthropic reports about 55% fewer safeguard blocks with Fable 5.2 than with Fable 5.1. It gives about 60% fewer in cybersecurity and 40% fewer in biology. Anthropic calls this early testing, and the percentages are approximate.
External evaluators. The announcement says external evaluators, including the UK AI Security Institute, Frontier Design and METR, tested the model before release. The supported claim does not give their findings, so this article does not summarize them.
Enterprise Frontier Safeguards. According to the announcement, Enterprise Frontier Safeguards is in private beta with customers including Bridgewater, Salesforce and Stripe. This article does not cover what that product does beyond this statement.
Readers who want background on one category of risk for AI agents can see our explainer on prompt injection and what it means for AI agents.
How to read Anthropic's figures on Fable 5.2 and Mythos 5.2
A few reading habits can help:
- Check who did the measuring. The CVE Rediscovery, October 2026 CVE run, sandbox, safeguard-block and capsid figures are Anthropic's own measurements.
- Keep the stated strength. The safeguard-block figures are described as "about" and come from early testing. They should not be read as exact.
- Check which model a figure applies to. Some results apply to Mythos 5.2, some to Fable 5.2, and some (such as the capsid result) do not name a specific model in the cited claim.
- Treat prices as a snapshot. The prices come from the cited pages and can change.
None of the material cited here says these models predict markets or pick investments. This article does not present them as suitable for investment decisions.
Limits of this article
This article is educational information based only on Anthropic's own published pages.
It is not investment advice and does not suggest that any AI model is suitable for investment decisions.
It does not cover access to Mythos, speed measurements, the full system card or any independent check of Anthropic's results.