What Is Fable 5? Features, Capabilities, and How It Works

Fable 5 is Anthropic's flagship public Claude model in the Mythos class, built for long-horizon autonomous work, advanced coding, multimodal reasoning, and large-context knowledge tasks. Anthropic launched it on June 9, 2026, positioning it above the Opus class while keeping stronger safety controls than its restricted sibling, Claude Mythos 5.

The short version: Fable 5 is the model you reach for when a task is too large for normal chat. Think repo-wide software changes, multi-document research, visual analysis across screenshots or charts, and agent workflows that need to run for hours or days. It is not just a smarter chatbot. It is designed to plan, inspect, test, revise, and hold state across very large projects.

What Is Fable 5?

Claude Fable 5 is Anthropic's most capable widely released Claude model, according to Anthropic's launch materials and early technical analyses. It shares its underlying model with Claude Mythos 5, but the two versions are configured differently.

• Fable 5: Public-facing, safeguarded, and available through approved Claude surfaces and API access.
• Mythos 5: Restricted to high-trust cybersecurity and critical infrastructure deployments, including programs such as Project Glasswing.

That split matters. Fable 5 gives developers and enterprises access to Mythos-class reasoning without exposing the full set of potentially risky capabilities. Mythos 5 is the less constrained version for vetted defensive security use cases. Same core model. Different safety posture.

Fable 5 Availability, Context Window, and Pricing

Fable 5 is available through Claude.ai, Claude Code for web, the Claude Code CLI, Claude Cowork, the Claude API, and supported cloud platforms, subject to regional and user eligibility rules. Anthropic initially offered early access at no extra charge on paid Claude plans until June 22, 2026, after which usage moved to separate billing.

Public pricing reports place Fable 5 at about $10 per million input tokens and $50 per million output tokens. That is not cheap. It is easier to justify when you are running high-value coding, legal review, research, or engineering workflows rather than simple question-and-answer chat.

Key specifications

• Model class: Mythos-class, above Opus 4.8
• Context window: 1,000,000 tokens
• Maximum output: 128,000 tokens
• Knowledge cutoff: January 2026
• Safety fallback: Routes restricted requests to Claude Opus 4.8 in under 5 percent of sessions on average, according to Anthropic

A practical note from API work: the 128,000-token output limit is a ceiling, not a setting you should reach for casually. If your client timeout is still 60 seconds, a giant code migration plan can fail in your application layer even when the model is doing exactly what you asked. Set timeouts, streaming, retries, and cost limits on purpose.

How Fable 5 Works

Fable 5 works as a long-context, agent-capable Claude model. The big difference is not only that it can read more text. It can keep plans, intermediate artifacts, tests, visual observations, and tool outputs inside a single working session.

In agent workflows, reviewers such as CodeRabbit describe Fable 5 as better at exploring vague tasks. It tends to inspect the environment first, identify constraints, understand available tools, then build. That sounds ordinary, but many coding agents still jump straight into editing files before they understand the repo. You have probably seen the result: broken imports, fake helper functions, and tests that pass only because they never exercised the bug.

Fable 5 also supports workflows built on sub-agent delegation, self-written test suites, and self-verification. In software terms, it can split a large job into smaller parts, assign specialized reasoning to subtasks, generate tests, run checks through connected tools, and revise its own output.

Core Features of Fable 5

Long-horizon autonomous work

Fable 5 is built for sustained work, not one-off answers. Its 1 million token context window lets it keep large repositories, long reports, design documents, plans, and prior decisions in scope.

This helps when you need the model to:

• Analyze a full codebase before proposing changes
• Maintain continuity across a multi-day research project
• Compare multiple long documents without losing details
• Draft and revise complex artifacts such as technical specifications or policy reports

My view: this is where Fable 5 earns its cost. If your task fits into a short prompt, a smaller model is the better choice. Use Fable 5 when losing context would create real risk.

Advanced software engineering

Fable 5 is one of the strongest public coding models reported in the Claude family. Benchmark data cited in technical coverage shows Fable 5 scoring 80.3 on SWE-Bench Pro, compared with 69.2 for Claude Opus 4.8. On CursorBench, one evaluation reports 72.9 for Fable 5 versus 64.3 for GPT 5.5.

Anthropic also reports Fable 5 as the highest scoring model on Cognition's FrontierCode evaluation, with strong performance on production-quality coding tasks. Public demos include rebuilding a web application from screenshots, generating playable games from a single prompt, and fixing bugs that prior models failed to resolve.

There is a caveat. CodeRabbit found that Fable 5 did not beat Opus 4.8 on every raw precision metric. It reported 32.8 percent actionable precision and 19.4 percent full precision for Fable 5, against 35.5 percent and 26.5 percent for Opus 4.8. Yet the same evaluation noted that Fable 5 felt stronger in messy, underspecified work because it explored and planned better. That distinction matters. Benchmarks measure slices. Real repos punish poor judgment.

Multimodal vision and document reasoning

Fable 5 can reason across images, scientific figures, charts, tables, screenshots, and long text. Anthropic reports major gains on document-based reasoning and chart interpretation benchmarks such as GDP Val AA and GDP.pdf.

Examples highlighted in demos include extracting numerical values from dense scientific figures, interpreting game screens, and connecting visual evidence to code or simulations. Fable 5 reportedly completed Pokemon Fire Red using only screenshots, with no maps or external navigation tools. It has also been shown playing Factorio, a game that demands understanding layouts, resource flows, bottlenecks, and delayed consequences.

For enterprises, the more practical use is less flashy: financial reports, legal exhibits, laboratory figures, product screenshots, engineering diagrams, and compliance tables. If your team still copies chart values by hand into spreadsheets, this is the kind of model that changes that workflow.

Research and knowledge work

Fable 5 suits long-form synthesis. You can feed it technical documentation, policy material, research papers, meeting notes, and code. The model can then produce structured outputs such as reports, slide outlines, comparisons, implementation plans, or experiment notes.

Tech News Weekly demonstrated Fable 5 generating presentation material that connected AI history, fan fiction, and tech culture. That may sound niche, but it points to a broader strength: the model can hold several distant concepts in context and build a coherent structure around them.

Safety Guardrails: Fable 5 vs Mythos 5

Fable 5 is intentionally more restricted than Mythos 5. Anthropic blocks or heavily controls requests involving offensive cybersecurity techniques, exploit development, and harmful biological protocols. When a request trips safety classifiers, Fable 5 can automatically fall back to Claude Opus 4.8.

Simon Willison noted that the API exposes signals for guardrail-triggered behavior, which matters for developers building production tools. You do not want a user-facing application to silently change model behavior without logging it. If a fallback happens, your app should know.

The downside is overblocking. Tech News Weekly observed that even benign biology explanations, such as the mammalian dive reflex for stress relief, could sometimes trigger stricter handling. That is the trade-off. Anthropic appears to prefer conservative safety filters over wider access to risky capability areas.

Real-World Use Cases for Fable 5

• Software teams: Repo-wide refactors, bug diagnosis, test generation, CI checks, and feature implementation from high-level specs.
• Product teams: Turning screenshots, requirements, and user feedback into working prototypes or technical plans.
• Researchers: Reading long papers, extracting chart values, comparing methods, and producing reproducible analysis steps.
• Legal and finance teams: Reviewing long contracts, filings, tables, and risk documents with traceable summaries.
• Enterprise AI teams: Building agents that maintain project state and coordinate tools across longer workflows.

If you are learning to build with models like Fable 5, start with fundamentals before you jump into autonomous agents. Blockchain Council's Certified Prompt Engineer™, Certified Generative AI Expert™, and Certified Artificial Intelligence (AI) Expert™ are natural learning paths for professionals who want structured AI training.

Limitations and Trade-Offs

Fable 5 is powerful, but it is not the right answer for every task.

• Cost: At reported pricing of $10 per million input tokens and $50 per million output tokens, small tasks should use smaller models.
• Latency: Huge context and large outputs can be slow. Design workflows around streaming and checkpoints.
• Safety restrictions: Cybersecurity and biology queries may trigger fallback, even when user intent is harmless.
• Availability: Export controls and region-based rules limit access to some users and capabilities.
• Agent risk: Long-running agents still need human review, especially before code touches production systems.

To be blunt, do not let Fable 5 directly merge code, change infrastructure, or send external communications without approval gates. Use it as a senior assistant with tools, not as an unsupervised operator.

What Fable 5 Means for AI Professionals

Fable 5 signals where frontier AI is heading: longer context, stronger planning, deeper multimodal reasoning, and tighter access control. The public Fable 5 and restricted Mythos 5 split may become a common pattern for high-capability AI deployment, especially where a system can assist with both beneficial defense and harmful misuse.

Your next step should be practical. Pick one workflow that is painful because of scale: a large repo, a stack of research papers, a long compliance document, or a visual analysis task. Build a controlled Fable 5 test with clear inputs, logging, human approval, and measurable output quality. If you want formal grounding, pair that experiment with Blockchain Council training in prompt engineering, generative AI, or AI development so you understand both the model and the operating risks.