Anthropic Launches Claude Fable 5.1 and Mythos 5.1: Major Coding Gains and Up to 45% Lower Agentic Costs
Anthropic Officially Releases Claude Fable 5.1 and Mythos 5.1
Anthropic has officially released Claude Fable 5.1 and Claude Mythos 5.1, delivering a major architectural update three months after the debut of Fable 5. The launch targets demanding software engineering, long-horizon problem-solving, and scientific research workflows.
Both models share the same underlying model and core intelligence, but differ in their safeguards and access policies. Fable 5.1 is generally available to developers and enterprise customers, while Mythos 5.1 is offered exclusively through trusted access programs designed for cybersecurity defense and life sciences research.
According to Anthropic, Fable 5.1 avoids shortcuts that degrade work quality, focusing instead on diagnosing and fixing the root causes of complex software failures.
Agentic Costs Fall by Up to 45% as Cache Reads Get 75% Cheaper
Addressing developer demand for cost efficiency, Anthropic introduced a 75% price reduction on prompt cache reads, lowering the rate from $1.00 to $0.25 per million tokens.
Because autonomous agents and iterative coding workflows repeatedly re-read stored context, this pricing change yields substantial real-world cost reductions:
- Typical Workloads: Operating costs drop by an estimated 25% for standard token-billed requests.
- Agentic Workloads: For context-heavy, multi-turn agentic loops, cumulative savings reach up to approximately 45%.
- Base Input Tokens: $10.00 per million tokens.
- Output Tokens: $50.00 per million tokens, with an output limit of up to 128k tokens.
- Cache Writes: $12.50 per million tokens (5-minute TTL) and $20.00 per million tokens (1-hour TTL).
- Batch Processing: $5.00 per million input tokens and $25.00 per million output tokens.
- Context Window: Standard 1 million token context window.
How Fable 5.1 Performs Across Major AI Benchmarks
Anthropic’s official evaluation suites show Fable 5.1 delivering significant gains across key software development, scientific inquiry, and general knowledge benchmarks:
- Terminal-Bench-Science 0.1: Fable 5.1 achieved 52.6%, substantially outperforming Fable 5 (24.7%), Opus 5 (29.0%), and GPT-5.6 Sol (22.4%).
- Terminal-Bench 4.0 (Autonomous Coding): Fable 5.1 scored 55.8%, while Mythos 5.1 reached 60.9%. Anthropic notes that safeguard interventions on certain dual-use cybersecurity tasks affected Fable 5.1’s score. Both models outperformed Opus 5 (52.3%), Fable 5 (42.0%), and GPT-5.6 Sol (37.3%).
- Humanity’s Last Exam (Multidisciplinary Reasoning): Achieved 60.9% without external tools and 65.0% when tool use was enabled.
- CursorBench 3.2.0: Reached 73.4% at maximum effort, surpassing Fable 5 (70.5%) and Opus 5 (70.0%).
- AutomationBench (Business Workflows): Scored 31.4%, compared to 17.1% for Fable 5.
- OSWorld 2.0 (Computer Use): Recorded 77.9% on partial completion and 41.7% under strict evaluation criteria.
- GDPval-AA v2 (Knowledge Work): Achieved 1,853 points, outperforming Opus 5 (1,824), Fable 5 (1,723), and GPT-5.6 Sol (1,711).
From Multi-Year Bugs to Autonomous Prototypes
Early testing across quantitative finance, developer platforms, and enterprise software highlights Fable 5.1’s capabilities in mission-critical environments:
Fable 5.1 Solves a Software Bug That Stumped Engineers for Years
Quantitative investment firm Millennium tested Fable 5.1 against an intermittent software crash occurring roughly once in every million runs-an issue that engineers and prior models had been unable to resolve for four to five years. Fable 5.1 disassembled an external vendor library, compared the assembly against core dumps, and traced the issue back to the underlying bug within hours.
Cognition Moves Devin Workloads to Fable 5.1
Walden Yan, Co-founder and Chief Product Officer at Cognition, confirmed that Cognition was moving Opus 5 traffic in Devin to Fable 5.1 on launch day. Yan noted that reduced cache read pricing makes a frontier Fable-class model economically viable for continuous code review workflows.
Fable 5.1 Builds a Complex Prototype in Three Days
Ron Sanzone, Staff Software Engineer at MongoDB, reported that Fable 5.1 built a complex prototype in three days. The model researched architecture documentation and codebases, then ran for hours unattended with robust verification loops to implement the entire prototype.
Additional enterprise feedback includes:
- Red Hat: Successfully identified the root causes of broken builds across all tested effort levels.
- Canva: Built a fully functioning interactive rhythm game in Canva Code, synchronizing gameplay elements directly to generated musical beats.
- Jane Street & Rakuten Medical: Noted sustained reasoning clarity over long financial tasks and the ability to formulate novel clinical hypotheses from complex datasets.
From Protein Design to Mapping Venus
Beyond standard software development, Fable 5.1 and Mythos 5.1 demonstrated meaningful contributions to scientific discovery and computational modeling:
Mythos 5.1 Designs High-Affinity Protein Binders
In laboratory validations conducted by external organizations, Mythos 5.1 generated high-affinity protein binders targeting 12 distinct biological sites, achieving a viable binder hit rate of nearly 50% (compared to industry averages of 10% to 15%). On three specific targets, binder affinities were 10 times higher than the top designs previously submitted to Adaptyv Bio competitions.
Fable 5.1 Maps One-Third of Venus in Greater Detail
Fable 5.1 trained a neural network to reconstruct a 3D digital elevation map covering one-third of the planet Venus. Using historical radar data from NASA’s Magellan mission, the map reveals surface features at a resolution of 2-3 kilometers (compared to 10-20 kilometers in earlier maps), with up to a 25% improvement in elevation accuracy. Anthropic released this map under a Creative Commons license to assist future exploration, including NASA’s VERITAS and the European Space Agency’s EnVision missions.
Mythos 5.1 Speeds Up Computational Biology by Up to 2.5x
Mythos 5.1 authored custom GPU kernels and cached intermediate results for seven open-source genomics and protein deep learning models (including Enformer, Evo 2, and Flashzoi). This accelerated inference speeds by up to 2.5 times while maintaining identical numerical outputs. Anthropic estimates that these optimizations could reduce GPU costs by 30% to 60% for genome-wide analyses.
What Fable 5.1 Changes for Claude API Developers
Integrating Fable 5.1 introduces important architectural requirements and breaking changes within the Claude API:
- Deprecation of Forced Tool Choice: Setting tool_choice to “any” or explicitly forcing a named tool now returns an HTTP 400 error. Because adaptive thinking is permanently enabled in Fable 5.1, forcing tool calls bypasses internal reasoning, reducing argument quality. Developers should use tool_choice: {“type”: “auto”} combined with explicit prompt instructions or structured outputs.
- Unidirectional Thinking Block Compatibility: Thinking blocks generated by Fable 5.1 cannot be read by earlier Claude models. While Fable 5.1 reads legacy reasoning transcripts, shifting a conversation to an earlier model causes reasoning blocks to be dropped.
- Dynamic Mid-Conversation Effort Control: Developers can now adjust the effort parameter dynamically between turns without invalidating prompt caches, raising reasoning depth for complex steps and lowering it for routine queries.
- Turn-Scoped System Messages: The new clear_at: “next_user_message” parameter allows system reminders to apply to a single turn before expiring, keeping prompt caches warm.
- Real-Time Progress Updates: Setting display: “updates” returns concise textual status updates between intermediate tool executions while keeping internal chain-of-thought blocks hidden.
Anthropic Tightens Safeguards Against Model Distillation
New API accounts created from September 1 onward can no longer manually edit Claude’s previous context in a multi-turn conversation while preserving the transcript of Claude’s prior thinking. Anthropic is rolling out this restriction gradually to protect model capabilities, while existing accounts are not currently affected.
Enterprise Frontier Safeguards Bring Stronger Data Controls
Anthropic introduced Enterprise Frontier Safeguards (EFS), providing enterprise customers with privacy protections equivalent to a Zero Data Retention (ZDR) agreement while preventing adversarial misuse.
With EFS, customers store data within cloud infrastructure they control on AWS, Google Cloud, or Microsoft Azure, rather than storing it on infrastructure controlled by Anthropic. EFS will roll out in phases starting later this fall, with eligible customers able to use Fable 5.1 under zero data retention in the interim.
Additionally, Anthropic updated its cybersecurity safeguards to be more precise: Claude Code users can expect roughly 60% fewer security interventions per session compared with the previous safeguards on Fable 5.
Who Can Access Mythos 5.1?
Dual trusted access programs govern sensitive domains:
- Cyber Verification Program (CVP): Provides vetted defenders access to advanced security modeling tools.
- Life Sciences Verification Program (LSVP): Developed in partnership with the U.S. government to enable verified researchers to utilize Mythos 5.1’s advanced biological capabilities.
Anthropic’s Push Toward Long-Horizon AI Agents
The release of Claude Fable 5.1 and Mythos 5.1 reflects Anthropic’s broader push to move AI beyond conversational assistance toward autonomous, long-horizon workflows.
By combining sustained reasoning, lower caching costs, and stronger enterprise safeguards, Fable 5.1 and Mythos 5.1 target developers and organizations building increasingly autonomous AI systems for critical software engineering and scientific discovery.
