Anthropic Unveils Claude Fable 5.1 and Mythos 5.1 with Breakthrough Scientific Reasoning and Slashed Cache Pricing

Key points

Claude Fable 5.1 achieves a 52.6% success rate on the Terminal-Bench-Science 0.1 benchmark, more than doubling the 24.7% score of its predecessor ¹³.
Cache read pricing has dropped 75% to $0.25 per million tokens, cutting overall costs for agentic workloads by up to 45% ¹¹.
The models are released in two versions: Fable 5.1 for general availability, and Mythos 5.1 with specialized safeguards for cybersecurity and life sciences research ².
Anthropic introduced Enterprise Frontier Safeguards to let organizations store data on their own infrastructure, ensuring complete data privacy ².

A Major Leap in Scientific AI

Anthropic has officially introduced its newest flagship models, Claude Fable 5.1 and Claude Mythos 5.1, marking a major advancement in how artificial intelligence handles complex coding, knowledge work, and scientific research ². Built to tackle multi-step problems that require sustained reasoning, Fable 5.1 more than doubles its predecessor’s score on the rigorous Terminal-Bench-Science 0.1 benchmark, jumping from 24.7% to 52.6% ¹³. This benchmark tests an agent’s ability to operate autonomously inside a command-line environment across diverse disciplines including life sciences, physics, and mathematics ¹³. Early testing also highlights practical breakthroughs, such as the model’s ability to write custom GPU kernels that speed up open-source deep learning workflows by up to 2.5 times ². In biology, the system has successfully designed high-affinity protein binders with hit rates reaching nearly 50% across twelve targets, significantly outperforming traditional design benchmarks ².

Redefining Agentic Economics

Beyond raw capability scores, Anthropic is directly addressing customer feedback regarding the high cost of running autonomous agents ². While base input and output token pricing remains unchanged at $10 and $50 per million respectively, the company has lowered the cost of cached input reads by 75% down to just $0.25 per million tokens ¹¹. Because long-running agents continually revisit system instructions, codebases, and accumulated conversation history, this sharp discount on cached context reduces overall expenses for heavy agentic tasks by up to 45% ². This pricing adjustment arrives as enterprise adoption faces scrutiny over unpredictable token bills, positioning the new models to better compete in production environments where cost-per-task determines economic viability ¹¹.

Security Architecture and Data Privacy

The release comes with a dual-tier structure designed to balance capability with rigorous oversight ². Fable 5.1 is broadly available with standard production safeguards, while Mythos 5.1 is restricted to vetted researchers through specialized access programs dedicated to cybersecurity and the life sciences ². To alleviate corporate privacy concerns, Anthropic is rolling out Enterprise Frontier Safeguards later this fall ². This architecture allows enterprise clients to store monitoring data on cloud infrastructure entirely controlled by their own organization, achieving zero data retention while maintaining robust safety oversight against misuse ². Furthermore, updated safety filters are designed to reduce false positives significantly, blocking roughly 60% fewer benign queries in cybersecurity and significantly easing restrictions on elementary medical and biological questions ² ¹⁰.

Broader Industry Dynamics

The launch arrives amid a competitive and sometimes turbulent period for frontier AI labs ¹⁰. Alongside Anthropic’s rollout, rival labs have reported major milestones and safety evaluations, underscoring the rapid escalation of agentic capabilities across the industry ⁹. At the same time, regulatory pressures are shaping how these models reach the market, with Anthropic’s outputs incorporating invisible watermarks to comply with transparency mandates under European AI regulations ¹². By pairing substantial performance gains with targeted cost reductions and more flexible enterprise deployment controls, Fable 5.1 represents an ambitious effort to transition AI from a conversational assistant into an autonomous partner for deep scientific and technical discovery ² ¹¹.

Companies mentioned: Anthropic

Primary sources

Run NVIDIA BioNeMo NIM Microservices for Protein Structure Prediction in Claude Science (developer.nvidia.com) – A technical guide from NVIDIA detailing the integration of the BioNeMo Agent Toolkit with Claude Science for protein structure prediction workflows using specialized NIM microservices. Published on NVIDIA Developer.
Claude Fable 5.1 and Claude Mythos 5.1 (anthropic.com) – Anthropic's official company announcement detailing the release of Claude Fable 5.1 and Mythos 5.1, highlighting improvements in scientific reasoning, molecular design, GPU kernel optimization, pricing adjustments, and security safeguards.

Further sources

Claude Fable 5.1 & Claude Mythos 5.1 System Card (www-cdn.anthropic.com) – The official system card for Claude Fable 5.1 and Claude Mythos 5.1, published by Anthropic in September 2026, providing thorough pre-deployment evaluations on cybersecurity risks, biological safety, agentic behavior, and model alignment.
Anthropic is cutting Claude Code's current weekly limits by 17% (bleepingcomputer.com) – A short news brief from BleepingComputer noting adjustments to Claude Code weekly limits alongside plan upgrades.
Anthropic launches Claude Fable 5.1 and says it’s up to 45 percent cheaper for agentic work (theverge.com) – A Verge report covering the launch of Claude Fable 5.1 and Mythos 5.1, detailing early customer impressions from industry leaders regarding coding performance, agentic efficiency, and data retention policies.
Anthropic’s new Fable release is cheaper, less restrictive (techcrunch.com) – A TechCrunch analysis of the Fable 5.1 release, focusing on reduced token costs, zero data retention initiatives via Enterprise Frontier Safeguards, and system card evaluations concerning model misbehavior.
OpenAI and Anthropic are risky for different reasons than their Chinese AI rivals (businessinsider.com) – A Business Insider feature examining AI safety concerns, researcher investigations into autonomous agent behavior at frontier labs, and the growing complexity of risks associated with advanced AI systems.
Anthropic upgrades Claude with new Fable 5.1 model, details here (9to5mac.com) – A 9to5Mac report outlining the arrival of Claude Fable 5.1, its model hierarchy position, performance benchmarks in complex troubleshooting, and cached token pricing changes.
OpenAI confirms Astra has reached critical cyber threshold, but will be available soon (mashable.com) – A Mashable report detailing OpenAI's announcement regarding its unreleased Astra model reaching critical cybersecurity capability thresholds while preparing for public release.
Anthropic launches Fable 5.1 as AI security worries mount (mashable.com) – A Mashable feature covering the launch of Claude Fable 5.1 and Mythos 5.1 against the backdrop of recent industry scrutiny over autonomous agent behavior and safety evaluations.
Anthropic's Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads (venturebeat.com) – A VentureBeat analysis breaking down the enterprise implications of Fable 5.1, deep-diving into the 75% reduction on cached input costs, benchmark comparisons, and recent cyber incident contexts.
Anthropic releases Claude Fable 5.1 and Mythos 5.1, cutting cache read prices by 75% (thenextweb.com) – A Next Web report covering the Fable 5.1 release, highlighting EU AI Act compliance, invisible watermarking, detection APIs in private preview, and access restrictions for the Mythos tier.
Anthropic doubles a science benchmark score with Fable 5.1 while OpenAI says its Astra model crosses critical cyber threshold (rdworldonline.com) – An R&D World article examining Anthropic's performance gains on scientific benchmarks alongside concurrent announcements from OpenAI regarding cybersecurity thresholds.
Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% Cheaper Cache Reads (marktechpost.com) – A MarkTechPost brief summarizing the core release metrics for Claude Fable 5.1, including its Terminal-Bench-Science score and cache pricing adjustments.
Anthropic Ships Claude Fable 5.1, More Than Doubling Its Predecessor on Key Benchmark (decrypt.co) – A Decrypt report detailing the release of Claude Fable 5.1 and Mythos 5.1, breaking down benchmark results against Opus 5, subscription credit allocations, and historical context from previous launch cycles.
Claude Fable 5.1 made me a nice animated pelican (simonwillison.net) – A blog post by Simon Willison testing Claude Fable 5.1 across various reasoning effort levels by prompting the model to generate an animated SVG of a pelican riding a bicycle.