Anthropic launches Fable 5.2 and Mythos 5.2 at $8 per million input tokens

⏱️ Reading time: 9 min

Anthropic raised the bar on what a model can do without constant supervision: during prelaunch testing, Claude Fable 5.2 formalized the proof of Fermat’s Last Theorem in the Lean language in 30 hours, a problem that took the mathematical community more than three centuries to close on paper. On November 9, 2026, the company introduced Fable 5.2 alongside its restricted-access twin, Mythos 5.2, with improvements in coding, reasoning, and vision.

📑 En este artículo

1. TL;DR
2. What Claude Fable 5.2 is
3. What happened
4. Context and history
5. Technical details
6. Impact and analysis
7. What’s next
8. Frequently Asked Questions
   1. What’s the difference between Claude Fable 5.2 and Mythos 5.2?
   2. How much does it cost to use Fable 5.2 or Mythos 5.2?
   3. Did Mythos 5.2 improve in cybersecurity compared to the previous version?
   4. What effort level should you use with Fable 5.2?
   5. Does the Fable-Mythos duo replace Claude Opus 5.5 for daily use?
9. References

The announcement also brings a price cut: both models now cost $8 per million input tokens and $40 per million output tokens, plus an optional fast mode that speeds up generation by up to 2.5 times for $16 and $80 respectively.

TL;DR

• Anthropic launched Claude Fable 5.2 and Mythos 5.2 on November 9, 2026 at $8 per million input tokens and $40 output.
• Fable 5.2 formalized the proof of Fermat’s Last Theorem in Lean in 30 hours during prelaunch testing.
• The optional fast mode costs $16 and $80 per million tokens and speeds up generation by up to 2.5 times.
• Mythos 5.2 rose to 89.6% on CVE recovery, up from 78.3% for version 5.1.
• On FrontierCode 2.0 with Xhigh effort, Fable 5.2 scored 49.1%, above Opus 5.5’s 45.5%.

What Claude Fable 5.2 is

Claude Fable 5.2 is Anthropic’s most capable model for open technical work: agentic coding, scientific research, and computer use without intermediate steps. It shares its architecture with Mythos 5.2, but Fable integrates cybersecurity and biology safeguards that let it be offered to any user, while Mythos requires prior authorization.

The difference between the two is one of governance, not technical capability. Anthropic designed Fable with restrictions that don’t require any paperwork for general use, and reserved Mythos for organizations enrolled in its Life Sciences and Cyber Verification programs, intended for biomedical research or authorized security testing.

What happened

The launch comes just a few months after Fable 5.1 and Mythos 5.1, and Anthropic presents it as an update within the same family rather than a new generation, with performance jumps the company describes as a step above Opus 5.5 across nearly every capability tested.

During prelaunch testing, Anthropic’s engineers used the models to complete software migrations they previously considered infeasible, design three functional electronic devices from scratch, solve two open problems in black hole physics, and advance robotics evaluations. None of those results came with public methodology beyond the announcement itself, so they remain internal claims until third parties replicate them.

The company also cut prices for the third time this year for this line: $8 per million input tokens and $40 per million output tokens in standard mode, with an optional fast mode at $16 and $80 that multiplies generation speed by up to 2.5 times.

Context and history

Fable and Mythos aren’t separate models. Anthropic describes them as the same underlying model with a different access policy: Fable carries cybersecurity and biology safeguards that block dangerous uses without requiring prior authorization, which allows it to be offered to any paying user. Mythos, on the other hand, is only available to organizations enrolled in the programs the company recently expanded.

That separation addresses a problem known since 2025: models capable of finding zero-day vulnerabilities or designing molecules can also be used for attacks or to create pathogens. Instead of limiting the model’s capability, Anthropic chose to filter access based on who is using it and for what, the same strategy it already applied with versions 5 and 5.1 of both models.

Technical details

The performance jump doesn’t come from a bigger model, but from how it spends compute when responding. Fable 5.2 supports multiple effort levels (Low, Medium, High, and Xhigh) and, according to Anthropic, at the Medium level (the default used by Claude Code and the Claude apps) it solves hard tasks with fewer tokens than version 5.1 needed for the same work. That matters because the cost of an agentic task depends both on the price per token and on how many tokens the model generates while thinking before it responds.

On the FrontierCode 2.0 benchmark, Fable 5.2 with High effort performed on par with Opus 5.5 at Max effort, but at roughly half the cost per task, according to Anthropic’s early testing. With Xhigh effort, Fable 5.2 scored 49.1%, the highest score reported by any published model Anthropic tested, compared to 32.4% for version 5.1 under the same effort level.

Fast mode is a separate lever: it doesn’t change the reasoning effort, it speeds up token generation by up to 2.5 times in exchange for doubling the price ($16 input and $80 output per million tokens). It’s useful for workflows where latency matters more than cost, such as agents responding in a live conversational interface.

flowchart TD
    A["Routine task"] --> B["Claude Haiku 5.5"]
    C["Standard coding task"] --> D["Claude Opus 5.5"]
    E["Extreme or open-ended task"] --> F["Claude Fable 5.2"]
    F -- "Advisor role" --> D
    F -- "Orchestrator role" --> B
    G["Mythos 5.2"] --> H["Biomedical research or authorized pentesting"]

Anthropic also suggests an orchestration pattern: Fable 5.2 can act as Advisor, consulted by smaller models when they get stuck on a hard problem, or as Orchestrator, delegating simple subtasks to cheaper models. The company’s operational recommendation is to start any task with Opus 5.5 and escalate to Fable or Mythos only when complexity demands it, or step down to Sonnet 5.5 or Haiku 5.5 when it doesn’t.

Impact and analysis

The biggest numbers aren’t in general coding, but in the domains where Mythos has expanded access. On known vulnerability recovery (CVE Rediscovery), Mythos 5.2 reached 89.6%, up from 78.3% for Mythos 5.1: an improvement of just over 11 percentage points in finding already-documented flaws from their description. On CyScenarioBench, a cybersecurity scenario benchmark where each lab uses its own tasks (so it isn’t directly comparable between companies), Mythos 5.2 scored 76.3% against 61.7% for Mythos 5.1.

In protein design, Mythos 5.2 rose to 85.1% on the benchmark for de novo design of target-binding molecules within 24 hours, up from 79.1% for the previous version. These are the two fields (offensive cybersecurity and molecular biology) where Anthropic most insists the safeguards didn’t hold back the improvement: the company states that Fable 5.2’s protections are as robust as 5.1’s but block fewer legitimate requests by mistake, and that it implemented its strongest protections yet against distillation attacks, the technique where a third party tries to copy a closed model’s capabilities by training its own model on that model’s responses.

In robotics, Fable 5.2 went from 27.6% to 64.9% on RoboDojo Sim, more than doubling the previous version’s result, though Anthropic clarifies it ran that benchmark at the highest available effort level and internally, without third-party replication yet. On general knowledge with tools (Humanity’s Last Exam), Fable 5.2 reached 74.2%, above Opus 5.5’s 67.7% and GPT-6 Astra’s 57.2%. This confirms that Claude Fable 5.2 improves more in agentic and tool-use domains than in pure knowledge benchmarks, where the gap with Opus 5.5 is smaller.

💭 Key point: none of the prerelease results (Fermat’s Last Theorem, the three electronic devices, the two black hole physics problems) come with public methodology beyond Anthropic’s own announcement: they’re claims from the company itself, not peer-reviewed papers.

What’s next

Anthropic leaves the door open for third parties to replicate its internal benchmarks, something that already happened with previous generations when outside labs ran FrontierSWE or Humanity’s Last Exam on Fable 5.1 and Opus 5.5. The easiest thing to verify first will be price and speed, since those are public API parameters. The hardest will be checking the scientific results, like the physics problems or the device designs, since that depends on Anthropic publishing the technical details or on outside researchers replicating the experiment themselves.

For teams already using Mythos, the immediate question is whether the improvement in CVE Rediscovery and CyScenarioBench translates into fewer false negatives in real bug bounty programs. That will only be known after months of production use, not from the benchmark released on launch day.

If you already have access to the Anthropic API, you can compare the real cost of your own task by running it with Fable 5.1 and Fable 5.2 at the same effort level and checking the output tokens each response reports.

Frequently Asked Questions

What’s the difference between Claude Fable 5.2 and Mythos 5.2?

They share the same underlying model. Fable includes cybersecurity and biology safeguards that allow it to be offered to any user; Mythos requires enrollment in Anthropic’s Life Sciences or Cyber Verification programs.

How much does it cost to use Fable 5.2 or Mythos 5.2?

$8 per million input tokens and $40 per million output tokens in standard mode. The optional fast mode raises that to $16 and $80 per million respectively, and speeds up generation by up to 2.5 times.

Did Mythos 5.2 improve in cybersecurity compared to the previous version?

Yes. On CVE Rediscovery it went from 78.3% (Mythos 5.1) to 89.6%, and on CyScenarioBench from 61.7% to 76.3%.

What effort level should you use with Fable 5.2?

Anthropic recommends Medium for daily use, the default level in Claude Code and the Claude apps; High when aiming for performance close to Opus 5.5’s maximum at a lower cost; and Xhigh only for the most demanding tasks, where Fable 5.2 scored the highest result recorded on FrontierCode 2.0.

Does the Fable-Mythos duo replace Claude Opus 5.5 for daily use?

No. Anthropic still recommends Opus 5.5 as the default model for most coding and knowledge tasks, reserving Fable or Mythos for the most ambitious or open-ended cases.

References

• Anthropic: official announcement of Claude Fable 5.2 and Claude Mythos 5.2, with benchmark and pricing table.
• Wikipedia: history and mathematical context of Fermat’s Last Theorem.
• Anthropic Docs: official documentation on models, pricing, and effort levels for the Claude API.
• GitHub: repository for Lean 4, the language used to formalize the proof mentioned in the announcement.

📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion

Featured image: Foto de Growtika en Unsplash