Anthropic has unveiled Claude Fable 5.2 and Claude Mythos 5.2, the company’s two most capable large language models to date, delivering outputs 30% faster than their predecessors and cutting API pricing by 20%. The launch, reported by The Verge, positions both models as purpose-built for demanding, long-horizon technical work — from formalising mathematical proofs to modelling viral protein structures at a level competitive biology tools cannot match.
Key takeaways
- Claude Fable 5.2 is available now across Anthropic’s platforms; Claude Mythos 5.2 is restricted to vetted organisations in the Life Sciences and Cyber verification programmes.
- Both models generate outputs 30% faster than the 5.1 generation, and Fable 5.2 scores more than 10% higher than its predecessor across 14 AI benchmarks.
- API pricing drops to $8 per million input tokens and $40 per million output tokens — down from $10 and $50 with Claude Fable 5.1.
- A high-performance mode delivers responses 2.5× faster than the standard tier, according to SiliconANGLE.
- During pre-release testing, the models helped formalise Fermat’s Last Theorem in 30 hours and answered two open questions in black hole physics.
- Fable 5.2 is 55% less likely to erroneously block a safe prompt than Fable 5.1, per SiliconANGLE.
Two Models, One Architecture, Two Access Tiers
Claude Fable 5.2 and Claude Mythos 5.2 share the same underlying capabilities but differ significantly in what they will and will not answer. As SiliconANGLE reports, the two models are identical at the core; the distinction lies entirely in their content policies and availability. Fable 5.2 blocks most research biology questions and many cybersecurity-related prompts; when a user sends a request that falls into either category, the model redirects it to Claude Opus 4.8 or Claude Opus 5 for handling rather than issuing a flat refusal. Mythos 5.2 carries more relaxed guardrails and is therefore available only to trusted researchers through two limited-access schemes: Anthropic’s Life Sciences verification programme and its Cyber verification track, as Yahoo Finance confirms.
The 5.2 generation follows the Claude Fable 5.1 and Claude Mythos 5.1 models released in September, according to The Verge. Anthropic describes the new pair as designed for “daunting” technical challenges — particularly complex, open-ended tasks that benefit from large-scale agent orchestration over extended timeframes. For a full view of the current Anthropic model range and pricing, see our AI models database.
Benchmark Performance: Science, Coding and Reasoning Gains
SiliconANGLE reports that Anthropic tested Fable 5.2 against its predecessor across 14 AI benchmarks, with the new model exceeding Fable 5.1’s score by more than 10% on most of them. Scientific reasoning recorded the sharpest improvement: Fable 5.2 scored 74.6% on Terminal-Bench-Science 0.1, a 22% gain over Fable 5.1’s result on the same test.
Coding performance also moved materially. According to SiliconANGLE, Fable 5.2 outperforms its predecessor by more than 13% across two specialised coding benchmarks — FrontierCode 2.0 and FrontierSWE 2.0. Anthropic frames both new models as well-suited to complex AI coding agents that need to manage long-running, multi-step engineering projects rather than short one-shot queries. The improvements in vision and reasoning, cited by The Verge, round out what Anthropic is calling the most capable generation it has shipped to date.
Scientific Feats: Fermat’s Theorem, Black Holes and Viral Capsids
Both Anthropic and independent tech media have highlighted a series of pre-release demonstrations that underscore the models’ scientific ambitions.
The Verge reports that during pre-release testing, Fable 5.2 and Mythos 5.2 collectively helped formalise a proof of Fermat’s Last Theorem in 30 hours. SiliconANGLE adds important context: earlier this year, Anthropic used an unreleased LLM to generate a computer-verifiable version of the same proof — a task that took 11 days. Fable 5.2 reduced that timeline to 30 hours by spinning up 1,000 parallel agents to distribute the calculations. The models also reportedly answered two open questions about a theoretical physics concept related to black holes, as both SiliconANGLE and The Verge report.
Mythos 5.2 contributed a distinct result in structural biology. According to SiliconANGLE, the model accurately modelled the protein structure of 28 out of 34 adeno-associated virus (AAV) capsids it studied. AAV capsids — the protective shells that some gene therapies use to deliver drugs into cells — are assembled from dozens of individual proteins. SiliconANGLE notes this is significant because other cutting-edge biology models, including Google DeepMind’s AlphaFold2, are optimised to model individual protein chains, not assembled multi-protein complexes. The creators of AlphaFold2 were awarded a share of the 2024 Nobel Prize in Chemistry.
The Verge adds that pre-release testing also saw the models design three working electronic devices from scratch, make substantial progress on robotics evaluations, and help engineers complete what Anthropic described as “otherwise-infeasible software migrations.”
Agentic Coding: Building an Operating System in Eight Days
Beyond benchmark numbers, Anthropic has shared one headline engineering demonstration. As SiliconANGLE reports, one of the company’s engineers asked Fable 5.2 to build a full operating system from scratch. The model orchestrated 1,000 agents, completing the task in approximately eight days. The resulting OS can run on a laptop, provides network access, and includes a Claude sidebar — illustrating the model’s capacity to manage large, parallelised software projects across an extended development cycle.
Anthropic sees Fable 5.2 and Mythos 5.2 as the right choice for highly complicated, open-ended technical tasks, while positioning Claude Opus 5.5 — introduced in July — as the preferred model for more routine, day-to-day workflows. SiliconANGLE describes one architecture where Opus 5.5 handles the majority of tasks and calls Fable 5.2 or Mythos 5.2 only when it reaches a bottleneck, letting teams balance cost and capability automatically. For more on the broader agentic coding landscape, our AI coding agents guide tracks the leading tools in this category.
Fable 5.2 vs Fable 5.1: Pricing and Speed Compared
The pricing shift is meaningful for teams running inference at scale. As both Yahoo Finance and SiliconANGLE report, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens — a 20% reduction compared with the $10 input / $50 output pricing carried by Fable 5.1. That figure is confirmed by our own models database. SiliconANGLE notes both Fable 5.2 and Mythos 5.2 also offer a high-performance mode that generates responses 2.5 times faster than the standard tier, intended for latency-sensitive enterprise workloads.
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Speed vs predecessor |
|---|---|---|---|
| Claude Fable 5.1 | $10.00 | $50.00 | Baseline |
| Claude Fable 5.2 | $8.00 | $40.00 | +30% standard; 2.5× high-perf mode |
Teams optimising for cost can compare the full range of Anthropic models — including the lighter Claude Opus 5.5 at $4.00 input / $20.00 output — alongside the wider competitive field using our AI API cost calculator.
Safety Architecture and Guardrail Improvements
Anthropic has tightened Fable 5.2’s refusal behaviour compared with its predecessor. According to SiliconANGLE, the model is 55% less likely than Fable 5.1 to erroneously block a safe prompt — a meaningful improvement for enterprise users who have encountered over-refusals in earlier generations. When Fable 5.2 does determine that a request touches on research biology or cybersecurity, it routes the query to Claude Opus 4.8 or Opus 5 rather than simply declining, preserving the user experience while maintaining the guardrail boundary.
Mythos 5.2 operates under a separate policy framework with more permissive defaults, tailored to the specific research needs of verified Life Sciences and Cyber programme participants. This two-tier approach — a broadly available model with conservative defaults alongside a restricted variant for authorised researchers — reflects an emerging pattern in how frontier AI labs handle dual-use capability at the top of their model ranges.
Pricing War and Enterprise Strategy
The 20% price reduction on Fable 5.2 arrives amid what Yahoo Finance characterises as an ongoing AI pricing war, as models become more computationally efficient and enterprise customers look for ways to cut AI service costs. Industry analysts cited by Yahoo Finance suggest that over time, large organisations will likely deploy a mix of closed-weight models like Anthropic’s and open-weight alternatives — avoiding lock-in with any single vendor while keeping overall costs under control. For a data-driven view of where the closed-versus-open cost gap currently stands, our open vs closed AI cost study provides a detailed breakdown.
Yahoo Finance also notes that the launch comes ahead of Anthropic’s anticipated initial public offering later this month, with market conditions and investor sentiment presenting some uncertainty around the company’s debut valuation.
Frequently Asked Questions
What is the difference between Claude Fable 5.2 and Claude Mythos 5.2? The two models share the same core capabilities but differ in their content policies. Fable 5.2 blocks most research biology and many cybersecurity prompts, routing them to Claude Opus 4.8 or Opus 5; Mythos 5.2 has more relaxed guardrails and is available only to organisations verified through Anthropic’s Life Sciences and Cyber programmes.
How much does Claude Fable 5.2 cost? Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens — 20% less than Claude Fable 5.1, which was priced at $10 per million input tokens and $50 per million output tokens, as reported by Yahoo Finance and SiliconANGLE.
How much faster is Claude Fable 5.2 than Fable 5.1? Outputs are 30% faster in the standard tier, according to Yahoo Finance. SiliconANGLE reports that a separate high-performance mode is 2.5 times faster than the standard tier for latency-sensitive workloads.
What scientific tasks has Claude Fable 5.2 demonstrated? During pre-release testing, the model helped formalise a proof of Fermat’s Last Theorem in 30 hours — reduced from an 11-day baseline using 1,000 parallel agents — and answered two open problems in black hole physics, as reported by SiliconANGLE and The Verge. The Mythos 5.2 variant accurately modelled 28 out of 34 AAV virus capsid structures, per SiliconANGLE.
When were Claude Fable 5.1 and Mythos 5.1 released? Claude Fable 5.1 and Claude Mythos 5.1 were released in September 2026, according to The Verge.
The Bottom Line
Claude Fable 5.2 represents a meaningful step forward from its predecessor: outputs arrive 30% faster, API pricing falls 20%, coding benchmark scores rise by more than 13%, and the false-refusal rate drops by 55%. The scientific demonstrations — a 30-hour Fermat proof, open black hole physics questions, complex viral capsid modelling — are designed to signal that Anthropic is targeting research institutions and advanced engineering teams as a core customer segment, not just the general enterprise market. Mythos 5.2 extends those capabilities further for vetted organisations prepared to operate within Anthropic’s structured access programmes.
For everyday workloads, Claude Opus 5.5 remains Anthropic’s recommended workhorse, with the two flagship models available on demand when a task exceeds its capabilities. Developers and teams can compare the full Anthropic lineup — and the wider competitive field — in our AI models database.
Sources: finance.yahoo.com, siliconangle.com, theverge.com, mashable.com, techeblog.com. Reported November 10, 2026.