aipoc-lab
/katBusiness/temAgents · AI in Business · LLM/podAnthropic · NVIDIA/typNews

AI Today: Claude Fable 5.2 and Mythos 5.2

Authorship and approval

[AI-generated content][Approved by a human]

Published and updated

Published: Updated:

State of knowledge as of 2026-11-10. Source material collected on 2026-11-10. Anthropic has released Claude Fable 5.2. In remote work task benchmarks, it completes 41.3% of tasks, more than double its predecessor, and paid plans are receiving additional rate limit resets this week. Additionally: Anthropic's call for researchers, AI for Science expanding to 20 institutions, Kuken reducing drawing search times to 4.5 minutes, and NVIDIA's partnership with Korea.

Today's Highlights (TL;DR)

  • Claude Fable 5.2: 41.3% task completion on remote labor benchmarks (compared to 17.9% for its predecessor), up to 30% faster, with 1 to 3 Fable limit resets for paid plans this week.
  • Anthropic Call for Researchers: Selected teams will receive free access to Claude for 8 to 12-week projects.
  • AI for Science: Anthropic's program includes partnerships with 20 foundations and research institutions.
  • Kuken Industries: A tool that searches for parts drawings in 4.5 minutes instead of nearly an hour.
  • NVIDIA and Korea: Expanding collaboration on the national AI ecosystem, ranging from infrastructure to workforce development.

Claude Fable 5.2: Twice the Completed Work, Plus Pro Plan Rate Limit Resets

Anthropic has released Claude Fable 5.2 alongside its twin, Mythos 5.2. They are the same model: Fable features safety mitigations for cybersecurity and biology, making it available to all users. Mythos is restricted to verified organizations.

  • Remote Labor. In the Remote Labor Index benchmark consisting of remote work assignments, Fable 5.2 completes 41.3% of tasks. Fable 5.1 completed 17.9%, Opus 5.5 21.3%, and GPT-6 Astra 20.8%.
  • Coding. In FrontierCode 2.0, which tests whether open source maintainers would accept model-generated code changes, Fable 5.2 scores 49.1%. According to Anthropic, this is the highest score among released models. Opus 5.5 scores 45.5%, and GPT-6 Astra 42.8%.
  • Autonomy. An Anthropic engineer gave the model an empty folder and a single instruction: design an operating system. Over 1,000 sub-agents worked for eight days. The resulting system boots on a laptop and features windows, a login screen, an installer, and built-in Claude.
  • Mathematics. Approximately 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in the Lean language within 30 hours, enabling computer-checked verification step by step. The first such formalization took Anthropic 11 days.
  • Speed and Price. Responses are up to 30% faster than Fable 5.1. API pricing is 8 dollars per million input tokens and 40 dollars per million output tokens, lower than version 5.1. Fast mode provides up to 2.5x faster responses for 16 and 80 dollars.
  • Pro Plan and Resets, This Week Only. Paid plan subscribers receive additional Fable usage: 1 to 3 limit resets depending on subscription tenure. On the Pro plan and standard Team plan seats, Fable can consume up to 50% of the weekly limit.
  • Mythos 5.2 is designed for cybersecurity and life sciences through verification programs. In the CVE Rediscovery benchmark, it found 10.2% more known vulnerabilities than Opus 5.5. For AI builders: Anthropic advises starting with Opus 5.5 as an everyday model and reserving Fable for the heaviest open technical tasks. On the Pro plan, Fable can consume half of the weekly limit, so it is better to save resets for large tasks rather than burning them on casual queries. All benchmark results are reported by Anthropic and are not independent measurements.

Anthropic Invites Researchers for 8–12 Week Projects with Claude

Anthropic has announced an open call for life sciences researchers: selected teams will receive free access to Claude models for collaborative projects. Researchers from universities, non-profits, and government laboratories are eligible to apply.

  • Projects may focus on human biology, molecular biology, medicinal chemistry, drug discovery, and clinical sciences, among other fields.
  • Collaboration with Anthropic researchers will last 8 to 12 weeks, with teams producing analyses, materials, or tools.
  • Collaboration results are intended for publication, with accompanying data and code shared as open source.
  • Researchers retain the rights to their data and independent follow-up research. Anthropic does not require data handover.
  • Anthropic has set an application window of approximately 14 days and plans to notify selected teams about 14 days after the deadline. For AI builders: Research group leaders from universities, non-profits, and government labs can evaluate whether their project can produce preliminary results within this timeframe. The announcement does not specify model usage limits or license terms for shared outputs.

Anthropic Expands AI for Science to 20 Institutions

Anthropic has expanded its AI for Science program by partnering with 20 foundations and research institutions. Partners will provide researchers with complimentary Claude credits for scientific projects.

  • The program covers research in biology, medicine, physics, chemistry, climate science, and mathematics.
  • HHMI Janelia is co-developing the MHS standard, which defines how AI agents can safely operate laboratory equipment; the standard is currently in research testing.
  • The Allen Institute is building agent teams to analyze open data on the brain, cells, and immune system. According to the institute, work previously measured in months can now take hours.
  • A portion of the credits will go directly to specific research initiatives, while partner organizations will distribute the rest to researchers. For AI builders: Research teams and foundations should review the credit allocation terms. The announcement does not specify credit amounts, eligibility criteria, or regional availability.

Kuken Cuts Drawing Search Time from an Hour to 4.5 Minutes

Kuken Industries partnered with Japanese startup Recursive to build a tool that finds parts drawings in 4.5 minutes instead of nearly an hour. Engineers can independently search the documentation archive of older cooling towers.

  • The archive spans 70 years of installations, including equipment built to custom client specifications.
  • Engineers enter a part or drawing number or upload an image or PDF. The application reads numbers from parts lists and points to potential matches.
  • Kuken estimates that the tool will reduce annual hours spent searching for drawings by approximately 90%.
  • Search results must be verified against tower data and replacement part histories. In a separate prototype tool, about 90% of drawing discrepancies flagged by AI turned out to be correct. In the Lab: The lab database also features ENEOS Materials brings ChatGPT Enterprise to manufacturing. There, the company deployed ChatGPT Enterprise in manufacturing, whereas Kuken uses AI for a specific task: retrieving historical drawings. For AI builders: Maintenance teams can test drawing retrieval across various equipment versions and verify matches against specifications. The source does not specify deployment costs, supported languages, or data processing policies.

NVIDIA and Korea Expand Sovereign AI Collaboration

NVIDIA and the Korean Ministry of Science and ICT (MSIT) are expanding their partnership on the national AI ecosystem. The initiative covers infrastructure, chips, workforce development, Korean foundation models, and startups.

  • NVLink Fusion: Korean NPU and XPU chipmakers can explore participation in this ecosystem.
  • AI for All: The collaboration will include three consortia and four foundation model creators. NVIDIA Dynamo will scale inference, which is the generation of model responses.
  • NVIDIA plans to offer training, certifications, hackathons, and technical seminars.
  • SK Telecom and Rebellions, supported by MSIT, are developing infrastructure for Korean models and workloads utilizing NVLink Fusion. In the Lab: On August 10, the lab covered discussions between Korea and NVIDIA at the AI Summit; the current announcement details specific areas of cooperation. For AI builders: Korean teams building models can check eligibility terms for AI for All. The announcement does not specify application rules, costs, or access terms for Dynamo and NVLink Fusion.

Frequently Asked Questions (FAQ)

How much Claude Fable 5.2 usage do I get on the Pro plan?

This week, paid subscribers receive additional Fable usage: 1 to 3 limit resets depending on subscription tenure. On the Pro plan and standard Team plan seats, Fable can consume up to 50% of the weekly limit.

Who is eligible to apply for access to Claude Mythos 5.2?

Fable is available to everyone, whereas Mythos is available to organizations conducting biomedical or cybersecurity research. Both models share the same base model; Fable includes mitigations that limit misuse risks in cybersecurity and biology.

What terms are outlined in Anthropic's call for researchers?

Selected teams from universities, non-profits, and government labs will receive free access to Claude models. Collaboration with Anthropic researchers will last 8 to 12 weeks, and results are expected to be published alongside open data and code.

How does Kuken's drawing search tool work?

The archive covers 70 years of installations, including custom client-specified equipment. Engineers can search by part or drawing number or upload an image or PDF. Matches must be verified against tower data and replacement part history.

Sources

Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.

Related posts

/katTech/temAgents · AI in Business · LLM/podGoogle · OpenAI/typNews

AI Today: GPT-6.2 Astra and Gemini

Discover the latest updates on GPT-6.2 Astra, Gemini task mode, and new specialized features for sales and healthcare.

/katEcosystem/temAgents · LLM · Open Source/podMicrosoft · NVIDIA/typNews

AI Today: Nemotron and PDD Kit

Discover how NVIDIA promotes open Nemotron models and Microsoft Research uses PDD Kit for verifying agent-written code.

/katTech/temAgents · Automation · LLM/podBig Tech (other) · NVIDIA/typNews

AI Today: Apertus 1.5 and NVIDIA in Industry

Discover the latest updates on Apertus 1.5, NVIDIA's industrial AI strategy in Germany, and multimodal use cases for Gemini.

/katTech/temAgents · Hardware · LLM/podAnthropic · Big Tech (other)/typNews

AI Today: DeepSeek V4.1 and NVIDIA RTX Spark

Check the latest updates featuring DeepSeek-V4.1-Flash-1017, NVIDIA RTX Spark laptops, and autonomous AI post-training systems.