Fable 5 and the AI-Native Company

The Exponential Keeps Moving

Two years ago, Claude’s frontier was drafting a commit message. A year ago, Opus 4 could build an entire feature on its own. Six months later, agents could run overnight on long-horizon tasks. Two months ago, Mythos read the entire OpenBSD source tree and found a 27-year-old vulnerability that had evaded human reviewers and static analysis tools for decades.

In June 2026, Anthropic released Claude Fable 5 and Claude Mythos 5, the fifth generation of Claude models. But the pattern here is not just capability growth — it’s that the intervals between jumps are getting shorter while the jumps themselves are getting bigger. Most business capabilities are still on a linear. The gap between what AI can do and what it’s actually doing in organisations is the collective opportunity.

What Makes Fable 5 Different

Two properties drive the gap between Fable 5 and previous models on agentic tasks:

Single-shot correctness. Give Fable 5 a complex, well-specified problem and it nails it on the first pass. Early testers reported single prompts producing work that would have taken teams days or weeks. This is not about generating more options — it’s about first-attempt precision on hard tasks.

Long-horizon autonomy. Fable 5 can run for days on a single goal and stay coherent throughout, even on tasks that span millions of tokens. It remembers your specifications, dispatches sub-agents, keeps them on track, and manages cost across the entire run more reliably than any previous model. Crucially, Fable 5 is as good at reading code as writing it — better at triaging outages, digging through repo history to find what broke and when, and proactively surfacing improvement suggestions.

The Glasswing safeguard system

High capability cuts both ways. Mythos — Fable 5’s sibling with cyber/bio safeguards lifted — was previewed first to a small set of partners under Project Glasswing because its cybersecurity capabilities were strong enough to be potentially misused.

Rather than restrict access, Anthropic built a safeguard routing layer into Fable 5: when a request touches cybersecurity, biology, or chemistry topics, it routes to Opus 4.8, clearly labeled, at Opus prices. This lets the most capable general model ship to everyone today while research access to Mythos 5 expands gradually through the Glasswing programme. Researchers doing legitimate work in these fields will sometimes hit the routing unnecessarily — that’s a known trade-off being actively improved.

Dynamic Workflows: Parallelism at Scale

The most immediately practical new capability for engineering teams is dynamic workflows in Claude Code.

The problem it solves: tasks like large-scale migrations, full-codebase audits, and localisation to many languages are composed of many independent subtasks. Running them sequentially with one Claude Code session is slow and requires manual re-prompting for each. Running separate sessions requires manual orchestration.

Dynamic workflows let you describe the work once and Claude creates a deterministic, parallel execution plan — then runs it across tens or hundreds of agents simultaneously.

The localisation demo from the Tokyo keynote is a concrete illustration: a single-agent sequential run of 13 language translations would take nearly an hour. With a dynamic workflow, all 12 additional translations run simultaneously, followed by 12 verification agents — and the whole workflow can be saved as reusable JavaScript code for future localisation runs. What previously required 13 separate manual tasks completes in one prompt.

This same pattern applies to migrations, security audits, performance sweeps, or any large job requiring deterministic parallel structure.

The Three-Layer Story

Fable 5 capabilities, the Claude platform, and Claude Code are three layers of the same story — each one depends on the others:

Model layer (Fable 5): More intelligent base behaviour means agents need less elaborate harnesses to produce good outcomes. Claude can dig farther with simpler primitives as intelligence increases — a filesystem and a sandbox are often enough.

Platform layer (Managed Agents): Scheduled deployments, vaults for secrets, memory, dreaming, 1M context window, and sandboxed execution environments. Rakuten’s product managers now coordinate teams of agents exactly as they’d manage teams of humans — and ship major releases every two weeks instead of once per quarter.

Product layer (Claude Code): Dynamic workflows, code review agents on PRs, routines for automating recurring tasks, security scanning overnight, agent view for monitoring parallel sessions. Spotify merges 1000+ PRs per month with migration agents and cut migration time by 90%. Mercari measured 90% year-over-year increase in engineering output.

What an AI-Native Company Actually Looks Like

The keynote opened with a concrete picture: a product detected it was broken at night, read its own error reports, found the bug, wrote the fix, rolled it out to users — and by the time the team woke up, the change log was already written.

This is not “people using AI to do their work.” It is work itself running on the substrate of AI, with people deciding what the outcome should be. The distinction matters:

AI-assisted	AI-native
Engineers use AI tools in their workflow	Work runs on AI agents; humans set goals and review outcomes
AI accelerates tasks	AI owns tasks; humans own the direction
Deployment is human-triggered	Deployment can be agent-triggered via schedules or webhooks
Context lives in the engineer’s head	Context lives in memory stores, runbooks, and skills

The infrastructure to become AI-native exists today: agentic harnesses (tools + environment + permissions), context management (memory, dreaming, skills), and production infrastructure (scheduled deployments, sandboxes, observability).

Design for the Next Model

One of the most actionable principles from the keynote: design your architecture for the next version of Claude, not the current one. The teams that win are the ones whose harnesses and product experiences are ready to absorb the next jump in intelligence.

As models get more intelligent, they can do more with simpler primitives. Sophisticated, over-engineered harnesses add friction when models improve. A simpler architecture — filesystem, sandbox, clear goals — ages better. And because capability jumps keep accelerating, building prototypes for experiences that don’t quite work yet is how you know when the exponential has moved underneath you and something magical has become possible.

🧠 Check your understanding

4 questions · your answers are saved in this browser only

1. 1. What are the two properties that most distinguish Fable 5 in agentic contexts?

   Faster token generation and lower API costs Single-shot correctness on complex tasks, and long-horizon autonomy that stays coherent over days and millions of tokens Larger context window and better tool use Better benchmark scores and multi-language support

   The keynote specifically named single-shot correctness (nailing complex tasks on the first pass) and long-horizon autonomy (coherent operation over days and millions of tokens, with reliable sub-agent orchestration) as the two properties that widen the gap between Fable 5 and other models on hard tasks.

2. 2. What does the Glasswing safeguard system do in Fable 5?

   It blocks all cybersecurity-related requests entirely It routes requests that touch cybersecurity, biology, or chemistry to Opus 4.8, clearly labeled, allowing the most capable general model to ship broadly today It requires human approval for sensitive requests before processing It reduces Fable 5's capability on sensitive topics to match Sonnet

   Rather than restricting access entirely, Glasswing routes sensitive topics to a known-safe model tier. The trade-off: legitimate researchers sometimes hit unnecessary routing, but the most capable model can be in everyone's hands today rather than being restricted for months.

3. 3. What makes a dynamic workflow different from just running multiple Claude Code sessions manually?

   Dynamic workflows use a more powerful model One prompt creates a deterministic parallel execution plan that runs tens or hundreds of agents simultaneously, and the plan can be saved as reusable code Dynamic workflows have access to the internet Dynamic workflows run on cloud infrastructure rather than locally

   The key differences: parallel execution (all agents run simultaneously, not sequentially), deterministic structure (not ad-hoc), and reusability (the workflow saves as JavaScript for future runs). Manual multi-session approaches require manual orchestration and can't run in a structured parallel graph.

4. 4. What does it mean to 'design for the next version of Claude, not the current one'?

   Use only the latest model API version in your code Build architectures and harnesses simple enough that they improve naturally when model capabilities jump — avoid over-engineered scaffolding that adds friction as intelligence increases Plan your product roadmap around anticipated future model releases Write evals that will pass on current models to prepare for upgrades

   As models get smarter, simpler primitives (a file system, a sandbox) become more powerful — while complex harnesses become friction. Designing for the next model means keeping your architecture lean and absorbent: ready to leverage new capability without needing major rewrites when it arrives.

Reset quiz

🛠️ Build it yourself

Follow these exact steps to reproduce it yourself

Try it yourself: your first dynamic workflow

Claude Code’s dynamic workflows are available in the Claude Desktop app and via the Agent SDK.

1. Identify a parallelisable task in your codebase — something with many independent subtasks: translating strings to N locales, running a migration across N modules, auditing N endpoints for a security property.

2. Write a single natural-language prompt describing the full task and asking Claude to “use a dynamic workflow to run the subtasks in parallel.”

3. Watch the workflow plan. In Claude Desktop, the side panel shows all agent instances spinning up. Notice: Claude creates the deterministic execution graph, not you.

4. Inspect the generated workflow. After completion, Claude saves the workflow as reusable JavaScript. Read it — this is the serialised plan. You can modify it and re-run without re-prompting.

5. Add a verification stage. Re-prompt Claude to add a verification pass after the worker agents complete: “After each translation, spawn a review agent that checks for accuracy and tone.” This is the pattern Spotify uses for migration PRs.