Claude Mythos 5 & Fable 5: The Agentic Leap
Anthropic released Claude Mythos 5 and Fable 5, their most capable models ever, featuring breakthrough long-horizon autonomy, superior code generation, and new safeguards. The Claude Platform now offers managed agents with scheduling, memory, and dreaming capabilities. Claude Code expands with dynamic workflows, remote control, and security scanning—enabling developers and enterprises to ship AI-native systems at unprecedented scale.
The New Frontier: Mythos 5 & Fable 5
Two New Fifth-Generation Models Released
Anthropic released Claude Mythos 5 and Claude Fable 5 just hours before the keynote. Fable 5 is the most capable model generally available; Mythos 5 is the same underlying model with cybersecurity and biology safeguards removed for authorized researchers.
Exponential Capability Growth on Accelerating Timeline
Model capabilities have progressed from simple commit messages two years ago, to full feature building one year ago, to overnight autonomous tasks six months ago, to finding a 27-year-old OpenBSD vulnerability two months ago. Each leap is bigger but arrives in shorter time intervals.
Fable 5 Dominates Coding Benchmarks
Fable 5 achieves state-of-the-art performance on nearly all tested AI capability benchmarks, with exceptional strength in software engineering, knowledge work, scientific research, and vision. It scores highest on SWE-bench Pro and outperforms competitors on specialized evals like Cognition's Crunchier Bench and Jensen Spark's UI/game coding tasks.
Single-Shot Correctness & Long-Horizon Autonomy
Fable 5 solves complex well-specified problems on the first attempt without iteration. It maintains coherence over days-long tasks spanning millions of tokens, remembers specifications, dispatches and manages sub-agents reliably, and operates with cost consciousness unmatched by prior models.
Superior Code Reading and Debugging
Fable 5 excels not just at writing code but at reading and understanding it. It is better at troubleshooting outages, digging through repository history to identify what broke and when, and proactively surfacing improvement suggestions.
Enterprise Workflow Mastery
Fable 5 manages end-to-end organizational work including financial analysis, documents, slides, and spreadsheets. It follows instructions precisely, stays on scope, and produces professional-grade outputs. It excels with messy, multi-threaded requests where it outshines other Claude versions.
Industry-Leading Vision Capabilities
Fable 5 reads dense technical images, web applications, plots, diagrams, and charts more accurately than any prior Claude version, making it superior for visual analysis and documentation.
Safeguard System for Responsible Deployment
To safely release Fable 5's powerful cybersecurity, biology, and chemistry capabilities, Anthropic built a safeguard system that routes sensitive requests to Opus 4.8 and charges Opus pricing. Researchers doing legitimate work may experience occasional blocks, but this trade-off enables broad access today rather than months of delay.
The Gap Between AI Capability and Business Reality
Exponential Model Growth vs. Linear Business Adoption
While AI model capabilities improve exponentially, most business capabilities remain linear. This growing gap between what AI can do and what it actually does for organizations represents the collective opportunity for developers building on the Claude Platform.
Platform Company Philosophy
Anthropic is a platform company because developers worldwide produce far more value on top of Claude than Anthropic could build alone. Most people will never call the Claude API directly; they experience AI through applications developers build on the platform.
Year-over-Year API Volume Growth
API volume on the Claude Platform is up nearly 17 times year-over-year, demonstrating rapid developer adoption and increasing reliance on Claude for production systems.
Customer Success Stories
Rakutin: From Code Acceleration to Managed Agents
Rakutin evolved from using Claude for code acceleration to building managed agents powering custom internal agents across engineering, product, sales, and finance. Product managers now coordinate agent teams the same way they manage human teams, shipping major releases every two weeks instead of once per quarter.
Canva: Interactive Design Components via Claude
Canva integrated Claude into its design platform, allowing hundreds of millions of non-coding users to request interactive components like maps, calculators, or widgets. Claude builds these as working mini-apps ready to drop into pages, democratizing interactive design.
Claude Platform: Harness, Context, Infrastructure
Three Ingredients for AI-Native Companies
Raw intelligence becomes business outcomes through three components: the harness (tools, environment, permission to act), context (information and memory), and infrastructure (scale and reliability). Together, these enable truly AI-native companies where work runs on AI substrate and humans decide outcomes.
Managed Agents: Purpose-Built for Fable 5
Claude Managed Agents is a new product offering that packages agentic harnesses, context management, and production infrastructure. When combined with Fable 5, it delivers better agent outcomes with significantly less developer effort than building custom solutions.
Agentic Harness: Brain and Hands Separation
The harness separates the brain (model decision-making) from the hands (sandboxed execution). Models don't just suggest fixes; they make them happen. Managed agents iterate toward specified outcomes, allowing the model to keep working until it achieves the goal.
Context Management: 1M Token Window with Memory
Managed agents operate with a 1 million token context window, enabling agents to consume large amounts of content without degradation. Agents can read and write their own skills to fill knowledge gaps, and use dreaming to inspect past trajectories and self-improve.
Infrastructure: Automatic Scaling and Reliability
Managed agents automatically spin up and down sandboxes and generate multiple agentic fleets as needed. This handles the hardest part of building long-running autonomous agents: extreme scale and reliability without manual infrastructure management.
New Managed Agents Features: Scheduling and Vaults
Developers can now schedule agent deployments to run on any cadence needed. Environment variables can be stored in vaults, allowing agents to make authenticated API requests securely without exposing keys.
Managed Agents Adoption: 10x Faster Development
Companies building agentic systems on managed agents complete projects 10 times faster because they avoid rolling their own harness, managing context, and building infrastructure from scratch.
Notion: Agent Orchestration in Product
Notion uses managed agents to power agent orchestration directly within their product, allowing users to delegate complex long-running work to Claude inside their workspace.
Asana: AI Teammates for Collaboration
Asana built AI teammates using managed agents—collaborative AI agents that work alongside humans inside Asana projects, taking on tasks and completing deliverables.
Claude Code: Developer Agents at Scale
Mission: Bridge Idea to Shipped Product
Claude Code exists to close the gap between an idea and a shipped product. It builds tools that elicit frontier intelligence from Claude models and makes them accessible to every builder, evolving alongside increasing AI capabilities.
Evolution from Manual Review to Auto Mode
A year ago, developers reviewed every edit Claude Code made with detailed instructions. Now, many developers use auto mode to delegate permissions to Claude, only checking in after Claude Code has tested changes and created a PR ready for review.
Multiple Interfaces: CLI, IDE, Desktop, Agents View
Claude Code offers four interfaces: CLI for power users wanting minimal text and control; IDE for following code changes; Claude Desktop for full-screen graphical interface with previews and visual indicators; and Agents View in CLI for terminal-based control plane.
Multi-Clotting: Running Multiple Agents in Parallel
Developers run multiple Claude Code agents in parallel (affectionately called multi-clotting). Claude Desktop and Agents View provide visual indicators showing which agents are working and which need input.
Anthropic Internal Impact: 8x Code Shipping
Anthropic engineers ship 8 times more code than in previous years using Claude Code, even as the engineering team has grown substantially.
Code Review Agent: Catch Critical Bugs
Claude Code deploys a team of agents to catch critical bugs in pull requests. Thousands of companies use this daily, including all internal Anthropic teams.
Remote Control: Code on iOS and Android
Claude Code remote control on iOS and Android lets developers code on the go without being tethered to a laptop, enabling work from anywhere.
Routines: Scheduled or Triggered Code Runs
Routines configure Claude Code to run on a schedule, webhook, or API call. Work that previously required manual human triggering can now run automatically.
Cloud Security: Overnight Vulnerability Scanning
Claude Security scans codebases overnight, flags vulnerabilities with severity assessment, and lets developers kick off Claude Code sessions to fix critical issues.
Dynamic Workflows: Parallel Agents for Large Tasks
Dynamic workflows run Claude Code in parallel across tens or hundreds of agents in deterministic structure for ambitious tasks like major refactors and migrations. Workflows can be saved as JavaScript code and reused.
Spotify: Migrate Thousands of Repos with 90% Time Reduction
Spotify built a background agent on the Claude Agent SDK that reads migration plans in plain English and kicks off a fleet of agents to open PRs. They merge over 1,000 PRs per month and cut migration time by 90%.
Mercari: 90% Engineering Output Growth Year-over-Year
Mercari, Japan's consumer-to-consumer marketplace, runs its entire engineering team on Claude Code and measured 90% engineering output growth year-over-year.
Average Developer Time with Claude
The average developer now spends 20 hours per week with Claude, reflecting deep integration of AI into daily development workflows.
Design Principles for the Exponential Future
Design for Next Version, Not Current One
Developers who win are those whose architectures, harnesses, and product experiences are ready to absorb the next jump in intelligence. Building for future capability rather than today's constraints enables breakthrough products.
Simpler Primitives as Models Get Smarter
As Claude becomes more intelligent, it needs fewer sophisticated harnesses. Simpler primitives like file systems and sandbox environments become sufficient because the model can dig farther with basic tools.
Build Harder Evals for Emerging Capabilities
Developers should create evaluation prototypes for experiences that may not work yet. When a prototype that wasn't working suddenly works, that signals an opportunity to ship something magical that wasn't possible before.
Make Model Upgrades Easy and Frequent
Teams that win treat model upgrades as business opportunities. This requires automated evals, testing processes, and hands-on experimentation with new Claude versions to discover new capabilities and deliver new customer experiences.
Real-World Demo: Dynamic Workflows for Localization
Sequential vs. Parallel Translation Workflow
Localizing a website to 13 languages sequentially with Claude takes about an hour with manual prompting. Using dynamic workflows, Claude creates a repeatable process and runs all 12 translations in parallel, then verifies each one—completing in one prompt what previously required 24 separate tasks.
Workflow Reusability as Code
Dynamic workflows can be saved as JavaScript code and reused for future translations or similar large-scale tasks, creating repeatable patterns for ambitious work.
The Three Layers of the AI-Native Future
Layer 1: Model Capabilities (Diane's Curve)
Frontier models like Fable 5 provide exponentially increasing intelligence and capability across coding, knowledge work, vision, and specialized domains.
Layer 2: Managed Agents Infrastructure (Angela's Platform)
Managed agents provide harness, context, and infrastructure that developers control, enabling production-grade autonomous systems without building from scratch.
Layer 3: Developer Tools (Cat's Code)
Claude Code and dynamic workflows bring the same leverage to developer work, enabling developers to ship code faster and tackle ambitious refactors and migrations.
Closing the Gap: Speed of Implementation
The remaining gap between AI capability and business reality is how fast developers can put these capabilities to work. All three layers run on Fable 5, the best model shipped for agentic work.
Notable quotes
The landscape is changing faster than ever. Things that weren't possible yesterday have become possible today. — Caitlyn Les
Work itself runs on the substrate of AI and people decide what the outcome should be. — Angela Jang
The teams who win are the ones that get the most out of Claude with model upgrades and treat them as business opportunities. — Diane Penn
Action items
- Try Fable 5 and Mythos 5 in your projects today; evaluate how single-shot correctness and long-horizon autonomy improve your use cases.
- Adopt Claude Managed Agents for long-running autonomous tasks; use scheduling and vaults features to build production agents without custom infrastructure.
- Experiment with Claude Code's dynamic workflows for large-scale tasks like migrations, refactors, or parallel processing jobs.
- Build evals for emerging capabilities that don't work yet; use them to detect when new Claude versions unlock new product experiences.
- Design your system architecture to absorb future Claude capability jumps; use simpler primitives and avoid over-engineered harnesses.
- Set up Claude Security scanning overnight to identify vulnerabilities, then use Claude Code to fix them automatically.
- Explore Claude Code's multiple interfaces (CLI, IDE, Desktop, Agents View) to find the workflow that best fits your development style.
- Configure Routines to automate recurring Claude Code tasks on schedules, webhooks, or API calls instead of manual triggering.