Claude Fable 5 is Anthropic’s new Mythos-class model for ambitious, long-running AI agent work: multi-day coding projects, deep research, complex document analysis, and enterprise workflows that need planning, tool use, self-checking, and human review. The short version: use it when the task is valuable enough to justify premium pricing, not when you only need a quick answer.
Anthropic’s Claude Fable page says the model is built for “the hardest knowledge work and coding problems.” The June 9, 2026 launch post describes Fable 5 as a Mythos-class model made safe for general use. It is available through the Claude API, consumption-based Enterprise plans, available marketplaces, Amazon Web Services, Google Cloud, and Microsoft Foundry. It is priced at $10 per million input tokens and $50 per million output tokens, with Anthropic’s existing 90% input-token discount for prompt caching. US-only inference is available at 1.1x pricing. Those details matter because the best Fable 5 use cases are long-running by design.
What is Claude Fable 5?
Claude Fable 5 is a generally available “Mythos-level” Claude model built for difficult, multi-stage work. Anthropic positions it above routine chat or simple coding assistance: it is meant for work that previous models could not sustain, especially projects that require planning across stages, delegating to sub-agents, testing outputs, and continuing asynchronously.
The most important phrase on Anthropic’s product page is not “state-of-the-art.” It is “days-long, complex, and asynchronous tasks.” That is the search intent behind Claude Fable 5: buyers and developers are trying to understand whether this is just a stronger model, or whether it changes the shape of agent work. The answer is that Fable 5 is aimed at the second category.
For teams already exploring multi-agent workflows, this moves the model layer closer to the workflow architecture many teams have been trying to build: one agent plans, another gathers context, another writes code or analysis, another checks the work, and a human approves the final action.
Claude Fable 5 pricing and access
Claude Fable 5 pricing is straightforward on paper: $10 per million input tokens and $50 per million output tokens. Prompt caching can reduce repeated input cost by 90%, which matters if your agent repeatedly loads the same repository instructions, policy files, design system, or project context. For US-only inference, Anthropic lists pricing at 1.1x for input and output tokens.
Developers can call the model as claude-fable-5 through the Claude API. Anthropic also says it is available natively on the Claude Platform, through available marketplaces, and in AWS, Google Cloud, and Microsoft Foundry. AWS notes that customers must configure data retention before invoking Fable 5 on Bedrock because Anthropic requires 30-day input and output retention for Mythos-class traffic.
The practical pricing question is not “Is Fable 5 expensive?” It is. The practical question is “Does it reduce cost per completed project?” If a model finishes a migration, research memo, legal review, analytics investigation, or agentic coding task in fewer human handoffs, the token price can be rational. If the task is a short summary, a simple classification, or a high-volume support response, a cheaper model will usually be the better default.
Where Claude Fable 5 fits in enterprise AI coding
Claude Fable 5 is especially relevant to enterprise AI coding because large codebases are full of work that cannot be solved by autocomplete. Migrations, refactors, dependency upgrades, test repairs, UI rebuilds, and multi-service changes require context gathering, judgment, execution, verification, and iteration.
Anthropic says Fable 5 is its most capable model for ambitious coding projects, including large migrations, complex implementations, and multi-day autonomous sessions. It can write tests to check its work, implement designs with high fidelity, and use vision to compare outputs against the goal. In its launch post, Anthropic also highlights customer feedback from teams using Fable 5 on long-horizon coding, agentic prototyping, and complex engineering workflows.
The operational takeaway: do not treat Claude Fable 5 like a faster code completion model. Treat it like a senior agent inside a harness. Give it a scoped goal, repository access, tests, a progress log, stop conditions, and a review process. If you skip those controls, a more capable model can still spend too long exploring or produce too much output for reviewers to trust.
That is why Fable 5 pairs naturally with the patterns in agentic workflow examples: define the trigger, gather context, let the agent draft or implement, require verification, and route risky actions to a human before they affect customers or production systems.
Long-running AI agents: the real Fable 5 use case
The biggest shift is long-running AI agents. Anthropic says Claude Fable 5 can work for days at a time in an agent harness such as Claude Code or Claude Managed Agents. The model can plan across stages, delegate to sub-agents, and check its own work. That does not mean you should let it run without structure. It means the structure can now support more ambitious work.
Anthropic’s engineering guidance on long-running agents is useful here: agents need a harness that preserves progress across context windows. A good harness gives the model an initialization step, a feature or task list, a progress file, a clean working state, tests, and a way to hand off context to the next session. Without that, even strong models can try to do too much at once, forget why a change was made, or declare victory too early.
For business teams, this is the difference between “ask an AI a question” and “assign an AI a project.” A question ends when the answer is generated. A project needs state. It needs artifacts. It needs review. It needs a record of what changed. If your team has not yet mapped which tasks are safe to delegate, start with an AI agent readiness checklist before routing real work to a long-running model.
Claude Fable 5 safeguards and fallback behavior
Claude Fable 5 includes new safeguards for cybersecurity, biology, chemistry, and model-distillation risk. Anthropic says queries flagged in these domains are routed to Claude Opus 4.8 instead of Fable 5. The product page also says customers are not charged Fable prices for rerouted requests.
Anthropic’s launch post says the safeguards trigger on average in less than 5% of sessions, meaning more than 95% of Fable sessions involve no fallback. That figure is important, but so is the caveat: if your workload is in cybersecurity, biology, chemistry, or anything that looks like distillation, the fallback rate may be much higher than the average. Teams in those domains should test with their own prompts, logs, and review process before assuming Fable 5 behavior will be consistent.
There is also a data-retention requirement. Anthropic says using Fable requires 30-day data retention for safety monitoring. The company states in its launch materials that this data is not used to train new Claude models and is retained to help detect misuse patterns and improve safeguards. For regulated enterprises, this is not a footnote. It belongs in the security review, vendor assessment, and data-flow diagram before production rollout.
When to use Claude Fable 5 — and when not to
Use Claude Fable 5 when the problem has three traits: high value, long horizon, and measurable completion. A codebase migration is a fit because the outcome is testable. A multi-document diligence review can be a fit because the output is a structured memo with cited evidence. A complex analytics investigation can be a fit because the agent can inspect data, form hypotheses, and produce a decision-ready report.
Do not use Fable 5 as the default for every chat, support reply, classification, extraction, or short-form content task. Those workflows usually need cost control, latency, and consistency more than maximum reasoning depth. A useful routing rule is simple: start routine tasks on a cheaper model, escalate to Fable 5 only when the task requires multi-step reasoning, tool use, self-verification, or long context.
For teams comparing AI agents vs automation, Fable 5 belongs on the “agent” side of the map. Rules and scripts are still better for predictable actions. Fable 5 is for ambiguous, high-context work where the model has to decide what to inspect next.
Example use cases for Claude Fable 5
Large code migrations. Give the agent a scoped migration objective, repository context, tests, and a completion checklist. Ask it to change one subsystem at a time, run tests, summarize changes, and stop for review before merge.
Complex product implementation. Fable 5 can help convert a product spec into a working implementation plan, generate code, compare UI output against design goals, and write tests. The key is to give it the design artifacts, acceptance criteria, and a clear review gate.
Deep research and analysis. For enterprise research, Fable 5 can work through large document sets, extract evidence, compare sources, and produce a structured deliverable. This is useful for finance, legal, analytics, and strategy work where charts, tables, and PDFs matter.
Multi-agent orchestration. In a more mature setup, Fable 5 can act as the planner or lead agent while smaller models handle cheaper sub-tasks. A lead agent might break a project into research, implementation, QA, and summary tasks, then reconcile the outputs before human review.
Workflow supervision. Fable 5 is also relevant to teams designing human approval systems. The more autonomous the model becomes, the more important it is to decide which outputs can ship automatically and which require review. Our guide to human in the loop AI agents covers that approval model in more detail.
A practical rollout plan
Start with one high-value workflow, not a company-wide model switch. Pick a project where completion is measurable: a migration with passing tests, a research memo with cited sources, a customer-analysis report with reproducible data, or a document-review workflow with reviewer acceptance.
Next, define the harness. Give the agent a task list, a progress log, repository or document access, a testing method, budget limits, and stop rules. Require it to leave artifacts: changed files, test results, notes, unresolved questions, and a short handoff summary. This turns an impressive model into a repeatable workflow.
Then add approval gates. A model that can work for days should not automatically merge code, email customers, change pricing, delete data, or publish public content without review. Route public, financial, irreversible, or customer-facing actions to a human. That keeps the benefit of autonomy without pretending the model is an accountable employee.
Finally, measure cost per accepted outcome. Token cost alone is misleading. Track total spend, human review time, completion rate, fallback rate, rework, and cycle time. If Fable 5 completes fewer tasks but those tasks are the hardest and highest-value, it may still be the right model. If it runs long without a clean finish, tighten the prompt, task scope, tests, and stop conditions.
Frequently Asked Questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s Mythos-level model for ambitious, long-running AI agent work. It is designed for difficult coding, knowledge work, vision, and enterprise workflows that require planning, tool use, and self-verification.
How much does Claude Fable 5 cost?
Anthropic lists Claude Fable 5 at $10 per million input tokens and $50 per million output tokens. Prompt caching receives a 90% input-token discount, and US-only inference is available at 1.1x input and output pricing.
What is the Claude Fable 5 API model name?
Anthropic says developers can use claude-fable-5 through the Claude API. The model is also available through the Claude Platform, available marketplaces, AWS, Google Cloud, and Microsoft Foundry.
What are Claude Fable 5 safeguards?
Claude Fable 5 includes safeguards for cybersecurity, biology, chemistry, and distillation risk. Flagged requests are routed to Claude Opus 4.8 instead of Fable 5, and Anthropic says rerouted requests are not charged at Fable prices.
Does Claude Fable 5 require data retention?
Yes. Anthropic says using Fable requires 30-day data retention for safety monitoring. Enterprises should review this requirement before sending sensitive workloads to the model.
When should an enterprise use Claude Fable 5?
Use Claude Fable 5 for high-value, long-horizon tasks such as code migrations, complex implementations, deep research, document-heavy analysis, and multi-agent workflows. Use cheaper models for routine, short, high-volume tasks.
Is Claude Fable 5 good for long-running AI agents?
Yes, that is one of its primary use cases. Anthropic says Fable 5 can work for days at a time in an agent harness such as Claude Code or Claude Managed Agents, but teams still need task lists, tests, stop rules, and human approval gates.
Leave a Reply