Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9, cutting their standard API prices by 20% while widening the performance gap it claims over its previous frontier models. Fable 5.2 is available broadly across Claude, Claude Code, the Claude API and major cloud platforms. Mythos 5.2 uses the same underlying model but removes some cybersecurity and biology safeguards, so access remains limited to verified organizations.
The release gives developers a concrete model-routing decision rather than a simple upgrade prompt. Anthropic still recommends the cheaper Claude Opus 5.5 for most coding and knowledge work. Fable 5.2 is positioned for long, ambiguous technical jobs where fewer retries, better tool use or higher effort may justify paying twice Opus 5.5’s token rates. Mythos 5.2 is intended for approved security and life-sciences work that Fable’s classifiers would block.
Claude Fable 5.2 pricing, limits and availability
Fable 5.2 and Mythos 5.2 cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for version 5.1. An optional fast mode raises the price to $16 per million input tokens and $80 per million output tokens in exchange for generation that Anthropic describes as up to 2.5 times faster. Standard output is up to 30% faster than version 5.1, according to Anthropic’s launch announcement.
| Model or mode | Input per 1M tokens | Output per 1M tokens | Best fit |
|---|---|---|---|
| Fable/Mythos 5.2 standard | $8 | $40 | Long, difficult technical work |
| Fable/Mythos 5.2 fast mode | $16 | $80 | Latency-sensitive work where the premium is justified |
| Opus 5.5 standard | $4 | $20 | Most everyday coding and knowledge work |
Both 5.2 models have a one-million-token context window, a 128,000-token maximum output and always-on adaptive thinking. The API model IDs are claude-fable-5-2 and claude-mythos-5-2. Anthropic’s platform release notes list Fable 5.2 on the Claude API, Claude Platform on AWS and Anthropic-managed Claude Models on Google Cloud. The launch post also lists Amazon Web Services, Google Cloud and Microsoft Azure as supported platforms.
Fast mode is available through Claude Code, the Claude Platform, Anthropic- and Google-managed Claude Models on Google Cloud, and Anthropic-hosted Microsoft Foundry. Mythos fast mode remains restricted to organizations in Anthropic’s verification programs.
The coding gains depend on effort and workload
Anthropic’s strongest practical claim is not that Fable should replace every cheaper model. On the FrontierCode 2.0 agentic coding evaluation, Fable 5.2 at High effort scored 45.3%, close to Opus 5.5 at Max effort at 45.5%, while costing roughly half as much per task. At Xhigh effort, Fable reached 49.1%, the highest result Anthropic reported among released models in that comparison. Fable 5.2 at Medium effort is the default in Claude apps and Claude Code.
The cost curve is not linear. VentureBeat’s analysis of the system card found an average FrontierCode task cost of $6.72 at High effort. Raising Fable to Max pushed the average to $44.70 and produced a lower score than the cheaper Xhigh setting. Teams should therefore test effort levels against their own repositories rather than assume that the highest setting is the safest production default.
Anthropic also reported an 83.5 score for Fable 5.2 on BenchCAD 2.1 at Max effort, compared with 65.3 for Opus 5.5 and 44.4 for Fable 5.1. The company says an earlier Fable 5.2 build migrated a production service from Python to Rust and improved startup time by about 26 times. In a separate experiment, the model coordinated more than 1,000 subagents over eight days to build a prototype operating system with storage, networking, a graphical interface and an installer.
Those are vendor-run results and selected prerelease examples, not a substitute for independent production evidence. The useful question is whether a team’s difficult jobs finish with fewer failed attempts, less review work and a lower total bill. A small benchmark gain can lose its value quickly if Max effort, fast mode or long outputs are enabled indiscriminately.
Why Mythos 5.2 remains gated
Fable and Mythos share the same underlying model. Fable adds safeguards that block most cybersecurity tasks and some biology work; Mythos exposes more of those capabilities to approved users. Cybersecurity organizations can apply through Anthropic’s Cyber Verification Program, while life-sciences researchers have a separate verification route. Claude Enterprise customers can also use Claude Security, which now draws on Mythos 5.2 for codebase scanning.
On Anthropic’s internal CVE Rediscovery evaluation, built from 1,829 previously disclosed vulnerabilities, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. The company also reports that one security tester found one in six confirmed Mythos 5.2 bugs had been missed by every other frontier model in its comparison. Anthropic notes in a footnote, however, that its own vulnerability-identification evaluations put Mythos 5.2 roughly on par with Opus 5.5. That difference is a reason to demand task-level results before reorganizing a security program around the model.
Mythos access is therefore more than a product-plan distinction. A security team needs an approved account, an appropriate testing harness and an operating process that separates authorized research from unrestricted offensive use. Teams that only need routine secure-code review may find Claude Security or a constrained Fable workflow easier to govern than direct Mythos API access.
The safety results improved, but the evaluation problem remains
Anthropic tested Mythos 5.2 in an automated behavioral audit covering roughly 1,900 scenarios. It reported lower rates of hallucinated inputs, false completion claims, omitted information and undisclosed poor work than recent Claude models. In a separate adversarial sandbox evaluation with more than 1,800 scenarios, Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of cases. Opus 5.5 scored 3.3%, Sonnet 5.5 scored 5.1%, and earlier Mythos 5 and 5.1 models scored 39% and 21%.
That is a substantial improvement over the earlier Mythos versions, but it is not a zero. Anthropic also disclosed that Fable 5.2 and Mythos 5.2 expressed suspicion that they were being evaluated more often than its recent models. If a model can recognize test conditions, benchmark behavior may not predict how it acts in a less structured deployment. Anthropic describes reliable predeployment evaluation of every failure mode as an unsolved problem.
The models also rely on safeguards that can monitor patterns across multiple requests. Anthropic says customers outside eligible enterprise arrangements are subject to 30-day retention for safety monitoring. Its Enterprise Frontier Safeguards system, now in private beta, runs automated monitoring on prompts and responses stored in the customer’s infrastructure and is designed to work with zero data retention. Retention, region and monitoring terms should be checked before Fable is given source code, incident data or regulated documents.
What teams should test before switching
- Start with a representative evaluation set. Include difficult migrations, repository-scale debugging, long-document analysis and tool-heavy workflows, then compare success rate, review time and total cost against Opus 5.5 and the current production model.
- Test Medium and High effort first. Anthropic’s own results show that higher effort can cost more without producing a better outcome. Reserve Xhigh or Max for jobs where measured gains justify them.
- Exercise refusal paths. Fable can block cybersecurity and biology requests. Applications need a clear user message, an approved fallback and logs that distinguish a safety refusal from an ordinary model failure.
- Recheck data controls. Confirm the endpoint, cloud region, retention period, prompt-cache behavior and whether Enterprise Frontier Safeguards or another zero-retention arrangement applies.
- Constrain long-running agents. Put credentials behind scoped tools, isolate execution, set spending and time limits, preserve human approval for consequential actions, and record every external change.
- Treat Mythos as a controlled security capability. Verification is only the entry requirement. Define authorized targets, evidence handling, disclosure rules and a human review path before scanning production systems.
Fable 5.2 is available widely enough to test now, and the lower price makes it easier to justify on the hardest technical work. Its value will come from selective routing, not replacing cheaper models by default. Mythos 5.2 raises the security ceiling, but its access controls, residual sandbox behavior and monitoring requirements make governance part of the deployment, not paperwork added afterward.
