Claude Fable 5 Explained: Capabilities, Pricing & How to Use It (2026 Guide)

Introduction

Every few months, an AI lab claims its new model is "the most capable yet." Most of the time, the difference is a few benchmark points that nobody outside a research team will ever feel. Claude Fable 5 is not that kind of release.

Launched by Anthropic on June 9, 2026, Fable 5 is the first model from a new tier the company calls Mythos-class — a class of models that sits above the familiar Opus / Sonnet / Haiku ladder. Anthropic's own framing is unusually blunt: Fable 5's capabilities exceed those of any model the company has ever made generally available, and it is state-of-the-art on nearly every benchmark it was tested on — software engineering, financial analysis, vision, scientific research, and long-running autonomous work.

What makes the release genuinely different, though, is the shape of it. Fable 5 ships as a pair:

• Claude Fable 5 — the version anyone can use, with built-in safety classifiers for high-risk domains.
• Claude Mythos 5 — the same underlying model without those classifiers, available only to vetted organizations through a government-partnered program called Project Glasswing.

That split — one brain, two doors — is a first for a frontier AI release, and it changes how you use the model, how you're billed, and even how your code handles responses.

This guide explains everything: what Fable 5 actually is, what "Mythos-class" means, how the safety system works, what it costs, how it compares to Claude Opus 4.8 and Sonnet 5, how to access it step by step, and what developers need to change in their integrations. Whether you're a curious beginner, a developer, a business decision-maker, a student, or an agency evaluating it for clients — this is written for you. (New to Claude entirely? Start with our Claude tool profile for the basics, then come back.)

1. What Is Claude Fable 5?

Claude Fable 5 is Anthropic's most capable generally available AI model — a large language model built for the hardest reasoning problems and for long-horizon agentic work: tasks that take hours or days, involve hundreds of steps, and require the model to plan, delegate, verify its own output, and keep going without a human nudging it along.

The essentials at a glance:

Spec	Claude Fable 5
Developer	Anthropic
Released	June 9, 2026
Model tier	Mythos-class (above Opus)
API model ID	claude-fable-5
Context window	1 million tokens (default)
Max output	128,000 tokens per request
Pricing	$10 / million input tokens · $50 / million output tokens
Thinking	Adaptive thinking, always on
Safety	Built-in classifiers with automatic Opus 4.8 fallback
Availability	Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, Microsoft Foundry

The single most useful way to understand Fable 5 is this: the longer and more complex the task, the bigger its lead over every other Claude model. On a quick question, you may barely notice a difference from Opus 4.8. Ask it to migrate an entire codebase, build a financial model from raw filings, or run a multi-day research loop with sub-agents, and the gap becomes dramatic. Anthropic reports that when run in an agent harness like Claude Code or Claude Managed Agents, Fable 5 can work productively for days at a time — planning across stages, delegating to sub-agents, and checking its own work as it goes.

2. What Does "Mythos-Class" Mean?

For three years, the Claude family had a stable three-tier structure: Haiku (fast and cheap), Sonnet (balanced), and Opus (most capable). Mythos-class is a fourth tier above all of them — but it isn't just "a bigger Opus."

Anthropic introduced the tier because its frontier research models had reached capability levels the company wasn't comfortable releasing without extra guardrails, particularly in cybersecurity and biology. The same skills that let a model find and fix vulnerabilities across a huge codebase could, in the wrong hands, help exploit one. The same biology reasoning that accelerates drug design could assist with harmful work.

The Mythos-class answer is a two-model structure:

1. Train one frontier model (the shared Fable 5 / Mythos 5 brain).
2. Wrap it in AI safety classifiers that watch for misuse in a small set of high-risk domains.
3. Release the wrapped version broadly (Fable 5), and offer the unwrapped version (Mythos 5) only to organizations with a verified legitimate need.

The precursor to all this was Claude Mythos Preview, an invitation-only model Anthropic ran earlier in 2026 through Project Glasswing — a collaboration with the US government that gives qualified cybersecurity professionals and critical-infrastructure operators access to unrestricted capabilities. Mythos 5 succeeds that preview model, and Fable 5 is the first Mythos-class model to reach general availability. Notably, both launched at less than half the price of Mythos Preview.

💡 Expert Tip: Don't read "Mythos-class" as marketing. It has a concrete technical meaning: the model is capable enough that Anthropic gates part of its skillset. If you never touch cybersecurity or life-science topics, you will almost never see the gate — Anthropic tuned it to trigger in under 5% of sessions on average.

3. Claude Fable 5 vs Claude Mythos 5

Because the two models share one architecture, one training run, one price, and one API surface, the comparison comes down to access and safeguards:

Aspect	Claude Fable 5	Claude Mythos 5
Underlying model	Identical	Identical
API model ID	claude-fable-5	claude-mythos-5
Pricing	$10 / $50 per million tokens	$10 / $50 per million tokens
Cybersecurity safeguards	Active	Lifted
Biology / chemistry safeguards	Active	Lifted for approved researchers
Behavior on flagged requests	Falls back to Claude Opus 4.8	No fallback — full capability
Who can use it	Anyone	Project Glasswing partners and trusted-access programs
Alignment profile	Same as Mythos 5	Low misaligned-behavior rates, comparable to Opus 4.8

Who actually gets Mythos 5? At launch: qualified cybersecurity organizations and critical-infrastructure providers working with Anthropic through Project Glasswing, via Anthropic, AWS, or Google Cloud account teams. Anthropic has said it plans to expand this into a systematic trusted-access program for cybersecurity firms, plus a dedicated program for biomedical researchers (with biology safeguards lifted but cyber safeguards retained).

Winner for almost everyone: Fable 5. Unless your organization does legitimate offensive-security research or advanced life-science work that trips the classifiers, Fable 5 is Mythos 5 for practical purposes — same intelligence, same price, no application process.

✅ Recommendation: If you're evaluating the model for a business, start with Fable 5 and measure how often (if ever) you hit the safety fallback. Only pursue Glasswing access if flagged sessions materially affect your workload — for most companies the answer is that they never will.

4. Capabilities: What Fable 5 Can Actually Do

Benchmarks age quickly, so this section focuses on reported, named results from Anthropic's launch materials and early-access customers — the kind of evidence you can verify against the sources at the end of this guide.

4.1 Software Engineering

This is where the headline numbers live:

• Stripe reported that Fable 5 "compressed months of engineering into days," completing a codebase-wide Ruby migration in a single day that would otherwise have taken a team more than two months by hand.
• Cognition measured it as the highest-scoring model on its frontier coding benchmark — even when run at medium effort, i.e. with the model deliberately spending fewer reasoning tokens.
• Cursor called it state-of-the-art on CursorBench and said it "opened up a class of long-horizon problems that were out of reach."
• GitHub reported that it exceeded benchmarks on complex, long-horizon coding tasks "with a level of autonomy and reliability" beyond prior models.

The pattern across all four: not better autocomplete — better sustained engineering. Multi-stage refactors, overnight runs, self-review before handoff.

4.2 Knowledge Work and Finance

Fable 5 posts the highest score of any model on Hebbia's Finance Benchmark, which tests senior-analyst-level reasoning. Trading firm IMC said it "aced their trading-analysis evaluations nearly across the board." Analytics company Thasos reported it was the first model to break 90% on its core analytics benchmark — a 10-point jump over Opus. Quill called it "the strongest finance-first model." Anthropic also reports 25–30% faster completion on spreadsheet tasks compared to Opus 4.8.

For businesses, the practical translation: end-to-end deliverables. Fable 5 is markedly better at producing the finished artifact — the financial model, the formatted spreadsheet, the slide deck, the redlined contract — rather than a draft a human then has to assemble.

4.3 Vision

Fable 5 is Anthropic's new state-of-the-art vision model. Two launch demonstrations stand out:

• It completed Pokémon FireRed using vision alone — reading raw game frames, no helper tools — where earlier Claude models needed scaffolding.
• It can extract precise numbers from dense scientific figures and rebuild a working web app's source code from screenshots of its interface.

It's also explicitly trained to work on hard images: when a chart is blurry, flipped, or noisy, it will reach for cropping and image-processing tools on its own rather than guessing.

4.4 Long-Context, Memory, and Self-Verification

The 1M-token context window isn't new to the Claude family — but Fable 5's ability to stay coherent across it is. Anthropic's most striking internal demonstration: in the roguelike game Slay the Spire, giving Fable 5 a persistent file-based memory tripled its performance compared to Opus 4.8, and it reached the final act three times more often. The model is unusually good at writing notes to itself and actually using them later — the core skill behind multi-day agentic work.

It also self-verifies. Tyson (early-access customer) put it well: the model "reflects on and validates its own work... that's what makes highly autonomous operations possible."

4.5 Scientific Research

Most of the published science results come from the unrestricted Mythos 5 (the safeguarded domains overlap with biology), but they show what the shared brain can do:

• Protein-design experts reported roughly 10× acceleration on aspects of drug development, with strong candidates for 9 of 14 protein targets.
• In blind comparisons, scientists preferred Mythos 5's novel molecular-biology hypotheses ~80% of the time over Opus-class output — and one hypothesis about an E. coli protein was later independently corroborated in published research.
• In an autonomous genomics project, it assembled single-cell data covering millions of cells across 138 animal species and designed a custom ML model that outperformed a Science-published model while being 100× smaller.
• In physics testing, Fable 5 nearly reached in 36 hours what GPT-5.5 achieved in four days — using about a third of the reasoning tokens.

🚀 Pro Tip: The common thread in every category is duration. If your workload fits in one prompt and one answer, cheaper models are usually fine. Fable 5 earns its price when the work spans many steps, many files, or many hours.

5. How the Safety System Works

This is the part of Fable 5 that has no precedent in earlier Claude releases, so it's worth understanding precisely.

The mechanism. Every Fable 5 request passes through AI classifiers that screen for a small set of high-risk categories: offensive cybersecurity, biology and chemistry misuse, and distillation attempts (trying to extract the model's capabilities to train a competitor). If a request is flagged, Fable 5 doesn't answer it — instead, on consumer surfaces the query is served by Claude Opus 4.8, a still-excellent model without Mythos-class capabilities in those domains. On the API, a flagged request returns a refusal stop reason, and you choose the fallback behavior (more on that in Section 8).

How often it triggers. Anthropic tuned the classifiers conservatively: under 5% of sessions on average see a fallback, and for users who never touch security or life-science topics the practical rate is near zero. False positives do happen — benign security tooling and legitimate biology questions occasionally trip the filter — which is exactly why the fallback path exists.

How robust it is. Anthropic ran an external bug bounty totaling more than 1,000 hours of testing: no universal jailbreaks were found. External red-teaming on long-form agentic tasks reached the same result (the UK AI Security Institute reported partial progress in initial testing). Zero harmful single-turn cybersecurity requests succeeded even when testers applied 30 different public jailbreak techniques.

The data-retention trade-off. Mythos-class traffic carries a mandatory 30-day data retention requirement — Fable 5 is not available under zero-data-retention agreements. Anthropic states the retained data is used only for safety (defending against sophisticated attacks, identifying false positives), never for model training, with human access logged and near-universal deletion after 30 days. On AWS Bedrock this surfaces as an explicit provider_data_share opt-in.

⚠ Common Mistake: Assuming the classifiers make Fable 5 "more censored" than other Claude models for everyday work. They target three narrow domains. Regular coding, writing, analysis, math, and even defensive security discussion at a normal level are unaffected — and when a false positive does occur, you get an Opus 4.8 answer rather than nothing.

6. Pricing: What Fable 5 Costs

API pricing

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context
Claude Fable 5	$10.00	$50.00	1M
Claude Mythos 5	$10.00	$50.00	1M
Claude Opus 4.8	$5.00	$25.00	1M
Claude Sonnet 5	$3.00 (intro $2.00)	$15.00 (intro $10.00)	1M
Claude Haiku 4.5	$1.00	$5.00	200K

Fable 5 costs exactly 2× Opus 4.8 per token. Three billing rules soften that in practice:

1. Refused-before-output requests are free. If the classifier declines a request before any output is generated, you're not billed at all.
2. Fallback credit. When a refused request is retried on another model, a credit mechanism refunds the prompt-cache cost of the switch, so you don't pay cache-write costs twice.
3. Fallback traffic bills at the fallback model's rate. If Opus 4.8 serves a flagged query, you pay Opus rates for that answer, not Fable rates.

Subscription plans

On the consumer side (claude.ai), Fable 5 was included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans from June 9 to June 22, 2026. Since June 23, using it requires usage credits on those plans, while API and consumption-based Enterprise customers simply pay per token. Anthropic has said it intends to bring it back to standard plans as capacity allows.

📌 Best Practice: For cost control on the API, use the effort parameter rather than downgrading models. Fable 5 at low or medium effort often beats older models at their maximum settings while spending far fewer output tokens — Cognition's benchmark result at medium effort is the proof point.

7. How to Access Claude Fable 5 (Step by Step)

There are five doors into Fable 5. Pick the one matching how you work.

7.1 In the Claude apps (easiest)

1. Open the model picker. In claude.ai (or the desktop/mobile apps), click the model name shown in the message composer — it displays the current model, e.g. "Opus 4.8," with a small chevron.
2. Select Fable 5. In the dropdown, choose Fable 5 — it sits at the top of the list, tagged as the most capable model. If your plan requires usage credits, a small credits note appears next to the entry.
3. Start chatting. Every message in the thread now runs on Fable 5. If a message trips the safety classifier, the app quietly serves an Opus 4.8 answer instead — you always get a response.

7.2 Via the Claude API

Change one string. If your code already calls the Messages API, set the model to claude-fable-5:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    messages=[{"role": "user", "content": "Audit this repository structure..."}],
)

if response.stop_reason == "refusal":
    # Classifier declined — retry on Opus 4.8 or use server-side fallbacks
    ...
else:
    print(response.content[0].text)

Section 8 covers the four integration changes hiding behind that one-line swap.

7.3 On Amazon Bedrock and Claude Platform on AWS

Fable 5 is on Amazon Bedrock (model ID anthropic.claude-fable-5, or global.anthropic.claude-fable-5 via bedrock-runtime), launched in US East (N. Virginia) and Europe (Stockholm) with more regions planned, and on the Anthropic-operated Claude Platform on AWS with the bare claude-fable-5 ID. Bedrock access requires opting into the 30-day provider_data_share retention mode.

7.4 On Google Cloud (Vertex AI) and Microsoft Foundry

Both platforms carry Fable 5 as of the June 9 launch — use your existing Claude-on-Vertex or Foundry setup and swap the model ID. Feature support varies slightly by platform (server-side fallbacks, for instance, are a Claude API / Claude Platform on AWS beta), so check the platform docs for the current matrix.

7.5 Claude Code and agent harnesses

If you use Claude Code, Fable 5 is selectable as a model, and it's where the model's long-horizon strengths show most: multi-hour refactors, sub-agent delegation, and memory-backed overnight runs.

🚀 Pro Tip: In Claude Code, pair Fable 5 with a clearly stated goal up front (one well-specified instruction beats ten incremental corrections) and let it run at high effort. The model is tuned for exactly this shape of work — under-specified, drip-fed prompts waste its main advantage.

8. Using Fable 5 in the API: What Developers Must Change

Fable 5 keeps the standard Messages API, but four behaviors differ from every previous Claude model. Budget an hour for integration changes, not a week.

1. Thinking is always on — delete your thinking config. Adaptive thinking is the only mode. Omit the thinking parameter entirely (or pass {"type": "adaptive"}). An explicit {"type": "disabled"} or the old budget_tokens form returns a 400 error. To control how deeply it reasons, use the effort parameter (low / medium / high / xhigh / max) instead.

2. Raw chain of thought is never returned. Thinking blocks arrive either as readable summaries (display: "summarized") or as empty placeholders (display: "omitted", the default). If your UI streams reasoning to users, set summarized explicitly — otherwise users see a long silent pause before output. Pass thinking blocks back unchanged on multi-turn conversations.

3. Handle the refusal stop reason — and opt into fallbacks. A classifier decline is a successful HTTP 200 with stop_reason: "refusal", not an error. Code that reads response.content[0] unconditionally will break. You have three retry options: the server-side fallbacks parameter (beta — the API transparently re-serves the request on Opus 4.8 in the same call), SDK middleware for client-side retries on any platform, or a manual retry using fallback credit to avoid paying prompt-cache costs twice. Anthropic's own consumer apps use the first pattern; API integrations should ship one of the three from day one.

4. Plan for 30-day retention and longer turns. Organizations on zero-data-retention agreements get a 400 on every Fable 5 request — check your retention configuration before debugging payloads. And because single requests on hard tasks can legitimately run many minutes, use streaming, generous timeouts, and progress UX rather than blocking HTTP calls.

Integration item	Old habit	Fable 5
Thinking config	budget_tokens: 10000	Omit thinking entirely; use effort
Reasoning display	Raw thinking text	display: "summarized" or omitted
Assistant prefill	Last-turn prefill for JSON	Not supported — use structured outputs
Refusals	N/A	Check stop_reason before reading content
Fallbacks	N/A	fallbacks param / middleware / fallback credit
Data retention	ZDR allowed	30-day retention required

⚠ Common Mistake: Migrating to Fable 5 by copying an Opus 4.6-era request with thinking: {"type": "enabled", "budget_tokens": 8000} — that's an instant 400. Strip the thinking block, strip temperature / top_p / top_k (also rejected), and let adaptive thinking do its job.

9. Fable 5 vs Opus 4.8 vs Sonnet 5

The three models most teams will actually choose between:

	Claude Fable 5	Claude Opus 4.8	Claude Sonnet 5
Tier	Mythos-class	Opus	Sonnet
Price (in/out per 1M)	$10 / $50	$5 / $25	$3 / $15
Context / max output	1M / 128K	1M / 128K	1M / 128K
Thinking	Always on (adaptive)	Optional adaptive	Adaptive by default
Long-horizon agentic work	Best available	Excellent	Very good
Typical turn speed	Slowest (deliberate)	Fast for its tier	Fastest of the three
Safety classifiers	Yes (cyber/bio/distillation)	No	No
ZDR compatible	No (30-day retention)	Yes	Yes
Best for	Hardest, longest work	Daily driver for demanding work	High-volume production workloads

Pros and cons in brief:

Fable 5 — Pros: highest ceiling of any available model; advantage compounds on long tasks; superb self-verification and memory use; state-of-the-art vision. Cons: 2× Opus pricing; minutes-long turns on hard problems; classifier false positives possible; no ZDR.

Opus 4.8 — Pros: excellent across the board; half Fable's price; no classifier layer; fast mode available; ZDR-compatible. Cons: ceiling is measurably below Fable 5 on frontier work; less durable on multi-day autonomous runs.

Sonnet 5 — Pros: near-Opus quality on coding at a third of Fable's input price; fastest; intro pricing through August 2026. Cons: the first to lose coherence as task length grows; not the right tool for frontier reasoning.

Winner by use case:

• Multi-day agentic runs, hardest engineering, frontier research → Fable 5
• Demanding daily work, security-sensitive orgs needing ZDR → Opus 4.8
• Production APIs, high-volume features, cost-sensitive scale → Sonnet 5

Two related reads if you're weighing Claude against other ecosystems: our ChatGPT vs Claude and Claude vs Gemini comparisons. And for an outside data point: Sakana AI reports that its multi-agent orchestrator Fugu Ultra matches or beats Fable 5 on several benchmarks by coordinating other frontier models — we break that claim down in our Sakana Fugu guide.

10. When Should You Use Fable 5? A Decision Framework

Price makes this decision simpler than it looks. Ask three questions in order:

1. Does the task span many steps, files, or hours? If it's a single prompt-and-answer, stop — Sonnet 5 or Opus 4.8 will serve you at a fraction of the cost.
2. Does the outcome justify a premium? A migration that saves two engineer-months justifies frontier pricing a thousand times over. A chatbot answering shipping questions does not.
3. Can you tolerate 30-day retention and occasional classifier fallbacks? If your compliance posture demands zero data retention, Fable 5 is off the table today — use Opus 4.8.

Three yes answers → Fable 5. Anything else → step down a tier and bank the savings.

11. Prompting Claude Fable 5: Best Practices

Fable 5 rewards a different prompting style than earlier models. The short version: specify the destination, not the route.

• Front-load the full task. One well-specified opening prompt — goal, constraints, definition of done — outperforms a drip of incremental instructions. The model is built to run autonomously from a clear spec.
• De-prescribe old prompts. Step-by-step scaffolding written for older models ("first do X, then Y, then Z...") often reduces Fable 5's output quality. State the goal and constraints; let it plan.
• Give it a memory surface. Even a plain markdown notes file it can read and write measurably improves long-run performance — Anthropic's game-playing tests showed a 3× improvement with file-based memory.
• Let it delegate. Sub-agent delegation is dependable on Fable 5. Instead of suppressing it (a common guardrail with older models), tell it when delegation is desirable — parallel, independent workstreams.
• Tune effort, not the model. Start at high for serious work, xhigh for the hardest coding and agentic tasks, and drop to medium/low for routine steps — lower effort on Fable 5 still often exceeds older models at their maximum.
• Explain the why. The model performs better when it knows the intent behind a request — who the output is for and what it enables — because it connects the task to relevant context instead of guessing.

💡 Expert Tip: If Fable 5's answers arrive slower than you'd like on routine work, don't switch models reflexively — drop effort to medium first. You keep the intelligence profile and often halve the wait.

12. Limitations and Things to Consider

An honest scorecard, because no model deserves an uncritical review:

• Price. $10/$50 per million tokens is double Opus 4.8. For high-volume workloads the difference compounds fast; use Fable 5 surgically.
• Latency by design. Single turns on hard problems can run many minutes. That's the correct behavior for frontier work — and the wrong tool for a snappy chat UI.
• Classifier false positives. Legitimate security engineers and life-science researchers will occasionally get Opus 4.8 fallbacks on benign requests. The fallback keeps you unblocked, but if flagged sessions are frequent in your domain, look into Anthropic's trusted-access programs.
• No zero data retention. The 30-day requirement is non-negotiable and excludes some regulated deployments today.
• Availability wobbles. Demand at launch was intense — subscription-plan access moved to usage credits after June 22, and Anthropic briefly paused and then restored availability in late June (see its "redeploying Fable 5" statement). Capacity is stabilizing, but plan for the possibility of throttles on consumer plans.
• Overkill for simple tasks. At higher effort on trivial work it can deliberate more than the task needs. Match effort to the job.

Conclusion

Claude Fable 5 is the first model that forced Anthropic to invent a new tier name — and having examined the evidence, the name feels earned. The benchmark story (state-of-the-art nearly everywhere), the customer story (Stripe's two-month migration in a day, Thasos's 10-point analytics jump), and the structural story (a frontier model made publicly available because of its safety layer, not despite it) all point the same direction: the capability ceiling for generally available AI moved in June 2026.

The practical guidance is simpler than the technology. Use Sonnet 5 for scale, Opus 4.8 as the demanding daily driver, and reserve Fable 5 for the work that was previously impossible: the multi-day agentic runs, the codebase-wide migrations, the senior-analyst deliverables, the research loops. That's where the 2× price buys 10× outcomes — and where every cheaper alternative simply doesn't finish the job.

Sources

• Claude Fable 5 and Claude Mythos 5 — Anthropic announcement
• Introducing Claude Fable 5 and Claude Mythos 5 — Claude Platform Docs
• Models overview — Claude Platform Docs
• Claude Fable — Anthropic product page
• Claude Fable 5 on AWS — AWS News Blog
• Claude Fable 5 — Amazon Bedrock model card