Claude Fable 5

Overview

Claude Fable 5 is Anthropic's most capable widely released model, made generally available on June 9, 2026. Anthropic describes it as "a Mythos-class model that we've made safe for general use" — the same underlying system as the restricted Claude Mythos 5, shipped with safety classifiers that make it suitable for broad deployment.

Fable 5 sits above the Opus tier rather than replacing it. Anthropic still recommends Claude Opus 4.8 as the default for complex agentic coding and enterprise work; Fable 5 is for workloads that need the highest available capability and can absorb its higher price and longer response times.

The model is aimed squarely at long-horizon autonomous work. It can operate independently for longer than any previous Claude model, and its API surface is shaped around that: thinking is always on, single requests on hard tasks can run for many minutes, and the context window is one million tokens by default.

Capabilities

• Frontier reasoning: State-of-the-art performance on nearly all benchmarks Anthropic tested, spanning software engineering, knowledge work, vision, and scientific research.
• Long-horizon agentic execution: Sustains autonomous multi-step work over extended runs without human correction. In early testing, Stripe used Fable 5 to complete a codebase-wide migration of a 50-million-line Ruby codebase in a single day — work a team estimated at over two months by hand.
• Software engineering: Achieved the highest scores among frontier models on Cognition's FrontierCode evaluation, at medium effort.
• Knowledge work: Scored highest on Hebbia's Finance Benchmark for senior-level reasoning, and performs strongly on end-to-end enterprise deliverables such as financial analysis, spreadsheets, and documents.
• Vision: Strong on dense and degraded images. Fable 5 played Pokémon FireRed using vision alone, where earlier models required additional tools.
• Sub-agent delegation: Reliably sustains ongoing communication with long-running sub-agents and peer agents, making it well suited to orchestrator roles in multi-agent systems.

Technical Specifications

• API model ID: claude-fable-5
• Context window: 1M tokens by default (~555K words). Fable 5 uses the tokenizer introduced with Claude Opus 4.7 — the same text produces roughly 30% more tokens than on models older than Opus 4.7.
• Max output: 128K tokens per request
• Thinking: Adaptive thinking only, always on
• Effort levels: low, medium, high, xhigh, max
• Training data cutoff: January 2026 (reliable knowledge cutoff: January 2026)
• Latency: Slower than Opus, Sonnet, and Haiku — an intentional trade for depth of reasoning
• Data retention: 30 days, required. Designated a Covered Model; not available under zero data retention.

Supported features

Effort, task budgets (beta), the memory tool, code execution, programmatic tool calling, tool-result clearing via context editing (beta), compaction, prompt caching, and vision.

API Behavior

Fable 5 is not a drop-in swap for an Opus-tier model. Four behaviors differ, and integrations need to account for each:

• Adaptive thinking is always on. It applies whenever the thinking parameter is unset. Sending thinking: {"type": "disabled"} returns a 400 error. Control reasoning depth with the effort parameter instead of a token budget.
• The raw chain of thought is never returned. Responses carry regular thinking blocks; thinking.display: "summarized" fills them with a readable summary, while the default "omitted" leaves the field empty. Thinking blocks must be passed back unchanged in multi-turn conversations on the same model.
• Sampling parameters are rejected. temperature, top_p, and top_k return a 400 error. Steer behavior with prompting.
• Assistant prefill is not supported. Use structured outputs or system prompt instructions to constrain the response format.

Long turns are the biggest structural shift: at higher effort, a single request on a hard task can run for many minutes. Plan timeouts, streaming, and progress indicators accordingly.

Safety & Safeguards

Fable 5's headline difference from other Claude models is that it can decline a request. Anthropic's announcement describes three areas — cybersecurity, biology and chemistry, and distillation — but the API exposes four stop_details.category values, and the fourth is easy to trip over:

• cyber: blocks offensive exploitation and agentic hacking tasks
• bio: broadly blocks dangerous biological and chemical research capabilities
• frontier_llm: the distillation classifier, covering requests that "could assist the development of competing AI models," which is restricted under Anthropic's commercial terms. Anthropic warns that benign machine learning work can also trigger it
• reasoning_extraction: fires when a request asks the model to reproduce its internal reasoning in the response text. To get reasoning in a structured form instead, use adaptive thinking with thinking.display: "summarized"

A declined request is returned as a successful HTTP 200 response with stop_reason: "refusal" — not an error — along with the classifier category that triggered it. Code that reads response content unconditionally will break; check the stop reason first.

Because benign adjacent work (security tooling, life-sciences research) can occasionally trigger a false positive, Anthropic provides three ways to retry a refused request on a model with broader availability, typically Claude Opus 4.8:

1. Server-side fallback — pass the fallbacks parameter and the API retries within the same request (beta).
2. Client-side fallback — SDK middleware retries from the client on any platform.
3. Manual retry — build the retry yourself, using fallback credit to avoid paying the prompt-cache cost twice.

You are not billed for a request that is refused before any output is generated. All Mythos-class model traffic carries a 30-day retention policy to support safety monitoring.

Availability & Export Control History

Fable 5 launched on June 9, 2026, generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry, with a phased rollout across Claude subscription plans.

On June 12, 2026, the U.S. government imposed export controls on Fable 5 and Mythos 5, requiring that access be restricted from foreign nationals whether inside or outside the United States. Anthropic could not verify nationality in real time and suspended access to both models for all users.

The export controls were lifted on June 30, 2026, and access resumed on July 1. On redeployment, Anthropic shipped an improved safety classifier targeting the specific bypass technique that Amazon researchers had disclosed; the company reports it blocks the reported technique in over 99% of cases, redirecting blocked requests to Claude Opus 4.8. Anthropic also restored Mythos 5 access for approved U.S. organizations and committed to pre-release testing and jailbreak information-sharing protocols with the government.

Fable 5 is now available globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork.

Claude Mythos 5

Claude Mythos 5 (claude-mythos-5) is the same underlying model as Fable 5, with safeguards lifted in some areas. It shares Fable 5's specs, pricing, and API behavior. Anthropic's API documentation states flatly that "Claude Mythos 5 does not include these classifiers," so refusal handling does not apply.

Access is invitation-only through Project Glasswing, Anthropic's collaboration with the U.S. government. The launch announcement frames the lifting of safeguards per cohort, not wholesale: Mythos 5 "is restricted to Glasswing partners (with cyber safeguards lifted) and soon to select biology researchers (with biology and chemistry safeguards lifted) only." Anthropic does not state which safeguards remain active for each cohort, and we do not infer it. There is no self-serve sign-up. Customers without access should use Fable 5, which offers the same capabilities.

Use Cases

• Overnight autonomous engineering: Large-scale refactors and migrations that run unattended and complete without human correction.
• Long-horizon research: Multi-hour investigations that gather context, build, and self-verify before reporting.
• Multi-agent orchestration: Coordinating long-running sub-agents, delegating independent subtasks and intervening only when one goes off track.
• Enterprise deliverables: End-to-end financial analysis, spreadsheets, slide decks, and documents produced from a single well-specified brief.
• Code review and debugging: Finding real bugs with high recall and precision, and correctly identifying intermittent flakes rather than declaring a fix after one clean run — outside security-focused analysis, where the cyber classifiers apply.
• Vision on hard inputs: Reading dense, flipped, blurry, or noisy images, including screenshots, charts, and scanned documents.

Performance / Benchmarks

Anthropic's launch announcement publishes its benchmark table as an image ("Benchmark table showing Claude Fable and Mythos compared to other leading models"), and the post's text carries no scores at all — only claims. Fable 5 is "state-of-the-art on nearly all tested benchmarks of AI capability"; on Cognition's FrontierCode it "scores highest among frontier models, even at medium effort"; on Hebbia's Finance Benchmark it "has the highest score of any model." None of those come with a number.

The figures below are transcribed from Section 8, "Capabilities," of the Claude Fable 5 and Claude Mythos 5 System Card, which does publish them as text.

Anthropic's caption for the summary table gives the standard configuration — "adaptive thinking at max effort, default sampling settings (temperature, top_p), averaged over 5 trials" — but attaches it to Mythos 5, and adds a warning that colors every Fable 5 row: "Fable's scores reflect its production safeguards, including fallback to Opus 4.8, which is why certain benchmarks score slightly lower on Fable compared to Mythos."

That is not a footnote to skim past. On Terminal-Bench 2.1 the system card is explicit: "Claude Fable 5: achieved 84.3% mean reward—with 20.9% of trials hitting a safety refusal and falling back to Claude Opus 4.8 for the rest of the trajectory, at high effort." Roughly one trial in five was finished by a different model.

What Anthropic does not publish for Fable 5

The summary table reports Mythos 5 and Claude Opus 4.8 on several benchmarks while leaving the Fable 5 column blank. There is no published Fable 5 figure for GPQA Diamond (Mythos 5: 94.1%), Humanity's Last Exam, BrowseComp, RiemannBench, USAMO 2026, ArxivMath, GraphWalks, CharXiv Reasoning, or HealthBench.

ProgramBench is omitted deliberately, and the reason is worth quoting: "We do not report separate ProgramBench results for Claude Fable 5, given that ProgramBench's core task, reconstructing the behavior of a compiled binary, falls within that category of tasks blocked by the cyber classifiers."

Anthropic also publishes no number for the Hebbia Finance Benchmark result it cites in the announcement.

Artificial Analysis

Fable 5's Artificial Analysis page reads: "Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) scores 60 on the Artificial Analysis Intelligence Index." The page labels that index v4.1. Note that the configuration string itself names the Opus 4.8 fallback — the same safeguard effect the system card warns about. Scores are not comparable across index versions.

How to read these numbers

These are vendor-reported figures. Anthropic chose the benchmarks, the harnesses, and the effort levels, and ran the evaluations itself except CursorBench (measured and reported independently by Cursor) and FrontierCode (Cognition's benchmark). Effort varies row to row — max on SWE-bench, high on Terminal-Bench, xhigh on FrontierCode — so the rows are not comparable with one another, let alone across vendors. And because Fable 5's safety classifiers can hand a trajectory to Opus 4.8 mid-run, part of what these scores measure is Opus 4.8. Treat them as an upper bound and validate on your own evaluation set.

Limitations

• Cost: At $10/$50 per million tokens, Fable 5 is Anthropic's most expensive generally available model — double the Opus tier. Reserve it for work that genuinely needs the extra capability.
• Latency: Comparative latency is "slower" by Anthropic's own classification, and hard tasks at high effort can occupy a single request for many minutes.
• Refusals: Safety classifiers can decline requests across four categories — cyber, bio, frontier_llm, and reasoning_extraction — including false positives on legitimate adjacent work. Anthropic explicitly warns that benign machine learning work can trip frontier_llm. Integrations must handle the refusal stop reason and should configure a fallback.
• No zero data retention: A 30-day retention requirement makes it unavailable to organizations with ZDR configured.
• Reduced API surface: No thinking-disabled mode, no sampling parameters, no assistant prefill, and no raw chain of thought.
• Over-prescriptive prompts hurt: Prompts and skills written for earlier models are often too prescriptive for Fable 5 and can reduce output quality. State the goal and constraints rather than enumerating steps.
• No internet access: Cannot browse the web unless integrated through tool use or web search.

Pricing & Access

API Pricing

Available via the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry:

• Input: $10 / MTok
• Output: $50 / MTok
• Prompt caching: Available for cost optimization on repeated context, with cache reads at roughly 10% of base input price.

This is less than half the price of Claude Mythos Preview, the invitation-only model Fable 5 succeeds, but twice the cost of the Opus tier.

Consumer Plans

Fable 5 is available on Claude.ai, Claude Code, and Claude Cowork. Subscription access rolled out in phases after launch, with usage credits required beyond the included allowance.

Ecosystem & Tools

Official SDKs

• Python SDK: Full-featured library with async support
• TypeScript SDK: Type-safe integration for Node.js and browsers
• REST API: Direct HTTP access for any programming language

Refusal-fallback middleware ships in the TypeScript, Python, Go, Java, and C# SDKs.

Developer Tools

• Anthropic Console: Web-based playground for testing and prompt development
• Claude Code: Agentic coding in the terminal, IDE, and web
• Claude.ai: Browser-based interface for direct interaction

Community & Resources

• Official Announcement: Claude Fable 5 and Mythos 5
• Redeploying Claude Fable 5 - Statement on the export-control suspension and restoration
• Fable 5 Cyber Safeguards and Jailbreak Framework
• Introducing Claude Fable 5 and Claude Mythos 5 - Developer documentation
• Models Overview
• Project Glasswing
• Pricing Page