What Can You Do With Claude Fable 5? New Features

What Can You Do With Claude Fable 5? New Features

Claude Fable 5 is not simply a sharper chatbot. Anthropic built it for the jobs people usually hesitate to hand to an AI: a large code migration, a research project spread across dozens of files, or an analysis that may need to run for hours rather than end after one reply.

Released on June 9, 2026, Fable 5 is the first generally available model from Anthropic's Mythos class. It sits above the familiar Opus, Sonnet, and Haiku tiers in raw capability. After a temporary suspension related to US export controls, Anthropic restored global access on July 1.

The useful question is not whether it tops another benchmark. It is what you can give it on Monday that might actually be finished by Tuesday. Here is where the new model changes the workflow, where it still needs supervision, and what it costs to use.

What Is Claude Fable 5?

Fable 5 is Anthropic's most capable model available to the general public. It uses the same underlying model as Claude Mythos 5, but Fable includes extra safeguards for sensitive cybersecurity, biology, and chemistry requests. Mythos is reserved for approved organizations and specialist programs.

The distinction matters because Fable is designed to bring Mythos-level reasoning to ordinary professional work without opening every high-risk capability. If a prompt triggers one of its safety classifiers, Claude may refuse it or route it to Claude Opus 4.8 instead. Anthropic says these classifiers activate in fewer than 5% of sessions on average, although harmless security work can still produce a false positive.

Fable 5 is best thought of as a project model. Sonnet remains the more sensible choice for quick, frequent tasks. Fable earns its higher price when the work has many stages, an uncertain route, or enough moving parts that a weaker model would need constant correction.

The Biggest Change: Claude Can Stay With a Job for Days

Earlier AI models could be impressive in a single exchange and surprisingly brittle across a long assignment. They would lose a requirement, stop when a tool failed, or ask the user to make a decision that could have been resolved by inspecting the available material.

Fable 5 is built for long-horizon work. In an agent environment such as Claude Code or Claude Managed Agents, it can plan a project in stages, delegate parts of the work to subagents, check intermediate results, and keep going asynchronously. Anthropic describes tasks lasting days, not just longer chat sessions.

That does not mean it should receive unlimited access to a production system. It means the useful unit of work can be larger. Instead of asking for a function, you can ask for a migration. Instead of requesting a summary, you can ask for the evidence to be reviewed, contradictions to be resolved, and a finished report to be produced.

What Can You Do With Claude Fable 5?

Take On Large Coding Projects

Coding is the clearest use case. Fable 5 can inspect a large repository, trace unfamiliar architecture, implement a change across multiple packages, write tests, run them, and repair failures. It is aimed at migrations and complex implementations rather than isolated snippets.

Its stronger visual understanding also changes frontend work. The model can build an interface, look at the rendered result, compare it with a design or screenshot, and revise obvious mismatches. That feedback loop is more valuable than generating plausible CSS once and assuming the page is finished.

In practice, good assignments include upgrading a framework across a monorepo, removing a deprecated internal API, porting software to another platform, or reviewing a pull request that already passed ordinary checks. GitHub says Fable completed equivalent autonomous coding work with fewer tool calls and fewer tokens than previous Opus-tier models in its internal tests.

Run Deep Research Without Losing the Thread

Fable 5 is well suited to research where the answer is scattered across reports, spreadsheets, PDFs, web sources, and internal notes. It can create a research plan, follow promising leads, compare claims, and produce a deliverable rather than a pile of summaries.

The difference should be most visible when the evidence is messy. A market scan may contain inconsistent figures and overlapping definitions. A legal review may require a table in one appendix to be read alongside a clause elsewhere. Fable is designed to preserve those connections over a longer run and to revisit its assumptions when new evidence conflicts with them.

You should still ask for citations and inspect important conclusions. The model's stamina is not a guarantee that every source is correct. Its advantage is that it can carry more of the investigative process before handing the work back.

Understand Charts, Diagrams, Tables, and PDFs

Claude models have supported images and documents for some time, but Fable 5 uses vision more deeply within a task. It can interpret charts and tables embedded inside PDFs, follow technical diagrams, and connect visual information with the surrounding text.

That makes it useful for financial reports, architecture documents, scientific papers, contracts, and analytics decks. You might ask it to compare a chart with the written conclusion, find figures that do not reconcile across a report, or turn a process diagram into an implementation plan.

Vision also acts as a form of self-checking. In software work, Fable can inspect the output of its own code and compare it with the requested result instead of treating a successful build as proof that the job is done.

Produce Finished Knowledge-Work Deliverables

Fable 5 can move from rough material to a report, project plan, analysis, or review that is closer to ready for a human decision. This is especially useful for tasks that mix reading, judgment, calculation, and writing.

Examples include turning meeting notes into a realistic delivery plan, reviewing a complex financial model, preparing an investment memo from a data room, or redlining a contract against a policy. These jobs are not hard because of one clever question. They are hard because dozens of small decisions need to remain consistent.

The model is also more proactive. When it encounters an unfamiliar tool or an incomplete clue, it is more likely to investigate than immediately stop for instructions. That can remove a great deal of back-and-forth, provided the assignment has clear boundaries.

Work as the Lead Agent in a Multi-Agent Setup

Fable 5 can break a project into parallel streams, send focused tasks to subagents, and combine the findings. A coding project might separate database work, interface implementation, and test coverage. A research project might divide sources by market or technical question.

This is not useful for every prompt. Delegation adds cost and coordination. It becomes worthwhile when the subtasks are genuinely independent or when a second pass can challenge the first agent's assumptions.

New Features for Developers

Fable 5 introduces a few changes that matter beyond model quality.

Adaptive Thinking Is Always On

Developers no longer switch reasoning on and off for this model. Fable 5 always uses adaptive thinking, then adjusts the depth according to the task and the configured effort level. The API does not return raw chain-of-thought. Applications can request a readable reasoning summary or omit thinking output.

This makes effort the practical control. A lower setting can handle a straightforward task without spending heavily, while higher effort gives a difficult project more room to check alternatives and validate the result.

Task Budgets Put a Ceiling on Long Jobs

Task budgets, currently in beta, let developers control how much work an agent can spend on an assignment. This is important when a model is capable of operating for hours or days. Autonomy without a budget is an unpredictable invoice.

A sensible agent can therefore have both a goal and a stop condition: finish the migration, but do not exceed the approved budget; research the market, but return the best supported answer within the available allowance.

Memory and Context Management Support Longer Runs

Fable 5 supports Anthropic's memory tool, context compaction, and tool-result clearing. Together, these features help an agent preserve useful facts while removing bulky intermediate output that no longer needs to occupy the active context.

This is less glamorous than a benchmark win, but it is essential for long projects. A model that remembers every terminal log verbatim eventually drowns in noise. A model that forgets everything cannot maintain a plan. Better context management helps it retain decisions, constraints, and progress without carrying every scrap of the process.

Programmatic Tool Calling Reduces Chatter

With programmatic tool calling, Claude can write code that coordinates tools and processes their results before returning to the model. Instead of making a long series of separate calls and repeatedly moving large payloads through the conversation, it can handle intermediate steps more efficiently.

For developers building agents around APIs, databases, or internal systems, this can reduce tool-call overhead and token use. It also makes more complex workflows practical because the model can calculate, filter, and combine results inside the tool process.

Availability, Pricing, and the Data-Retention Catch

Claude Fable 5 is available in Claude.ai, Claude Code, Claude Cowork, and the Claude API. Anthropic lists access for Pro, Max, Team, and Enterprise customers. It is also offered through supported cloud marketplaces, including AWS, Google Cloud, and Microsoft Foundry, and it has returned to GitHub Copilot for eligible plans.

API pricing is $10 per million input tokens and $50 per million output tokens. Prompt-cache reads receive the existing 90% input-token discount. US-only inference costs 1.1 times the standard input and output rates.

That makes Fable much more expensive than a model intended for routine chat. The right strategy is not to make it the default for everything. Use a faster model for classification, rewriting, and simple code changes, then route the assignments that repeatedly defeat those models to Fable.

There is also a privacy detail teams should review before enabling it. Fable 5 requires up to 30 days of prompt and output retention so Anthropic can operate its safety classifiers. Anthropic says retained data is deleted after that period and is not used to train its models. Organizations that require zero data retention may need to keep Fable disabled or remove sensitive information before use.

Limits You Should Know Before Using It

Fable 5 can work longer, but it can also spend longer pursuing a weak interpretation. A vague brief does not become a good brief because the model is more capable. Define what success looks like, what it may change, which sources are authoritative, and when it should ask for approval.

The cyber and biology safeguards may also affect legitimate work. Flagged requests can fall back to Opus 4.8, so the behavior and cost profile may change during a session. API developers need to configure fallback handling rather than assume it happens automatically in every integration.

Finally, self-checking is helpful but not independent verification. Tests written and passed by the same agent can share the same blind spot. For important software, finance, legal, or scientific work, keep a human review and use external checks wherever possible.

Is Claude Fable 5 Worth It?

For everyday prompts, probably not. Fable 5 is expensive, retains data for safety monitoring, and brings little advantage to a two-paragraph rewrite.

For a project that would otherwise take a skilled person several days, the calculation changes. Its appeal is not a slightly better answer. It is the possibility of handing over a larger block of work and receiving something coherent enough to review, test, and use.

The best early use cases are large code changes, document-heavy research, visual analysis, and multi-stage professional work. Start with a bounded project, watch where the model spends its time, and compare the finished result with the total cost—not just the first response.

Claude Fable 5 feels less like a new chat model and more like Anthropic's attempt to build an AI project worker. Whether it succeeds depends on the assignment, the tools around it, and the quality of the review at the end.

Frequently Asked Questions

When was Claude Fable 5 released?

Anthropic launched Claude Fable 5 on June 9, 2026. Access was suspended on June 12 after new US export controls and restored globally on July 1 after those restrictions were lifted.

Is Claude Fable 5 better than Claude Opus 4.8?

Anthropic positions Fable 5 above Opus 4.8 for long-horizon coding, knowledge work, vision, and complex agent tasks. Opus may still be the more practical choice when cost, latency, or zero-data-retention requirements matter.

Can Claude Fable 5 work for days?

Yes, when it runs inside an agent environment such as Claude Code or Claude Managed Agents. It can plan stages, use tools, delegate subtasks, and continue asynchronously. Clear permissions, budgets, and review points are still essential.

How much does the Claude Fable 5 API cost?

The API costs $10 per million input tokens and $50 per million output tokens. US-only inference is available at 1.1 times the standard price.

Does Claude Fable 5 retain my data?

Yes. Anthropic requires up to 30 days of prompt and output retention for Fable 5 safety monitoring. The company says this retained data is not used for model training and is deleted after the retention period.

Sources

Anthropic: Claude Fable 5 and Claude Mythos 5

Anthropic: Claude Fable 5 model overview

Anthropic: Redeploying Claude Fable 5

Claude Platform Docs: Working with Fable 5

GitHub: Claude Fable 5 in GitHub Copilot