Introducing Claude Fable 5 and Claude Mythos 5

Introduction

There are product launches, and then there are turning points. Furthermore, June 9, 2026 is already being treated by the AI industry as the second kind. On that date, Anthropic released Claude Fable 5 and Claude Mythos 5 the most capable models the company has ever shipped, belonging to an entirely new capability class that sits above the Opus tier in Anthropic's model hierarchy. Consequently, developers, enterprise teams, researchers, and AI professionals are now working with a model that posts 95% on SWE-bench Verified, completed a codebase migration that would have taken a full human team two months in a single day, and advances scientific research in ways that no previously available Claude model could approach. This guide covers everything: what the models are, how they differ, what they cost at every pricing tier, how the safety architecture operates, what the benchmarks show, and what this launch means for professionals building with or alongside AI.

The Mythos-Class Tier: A New Category Above Opus

Since Claude 3 launched in March 2024, Anthropic has organized its models under three tiers: Opus at the top, Sonnet in the middle, and Haiku at the base. That structure served the Claude 3 and Claude 4 families well but Claude Fable 5 and Claude Mythos 5 do not belong to it. They define an entirely new tier above Opus, which Anthropic calls Mythos-class.

The name carries deliberate meaning. The word "Fable" comes from the Latin fabula "that which is told" with kinship to the Greek mythos. According to Anthropic's official announcement, the naming reflects the narrative scale of what these models can accomplish: not single-turn answers, but extended stories of work, research, and discovery carried out across millions of tokens without human interruption.

The Mythos-class story began in April 2026 with Claude Mythos Preview a model so advanced in its cybersecurity capabilities that Anthropic declined to make it publicly available. Instead, it released Mythos Preview exclusively to vetted cybersecurity partners through Project Glasswing, a program operating in collaboration with the US government. The June 9 announcement delivers two things simultaneously: the first public access to Mythos-class capability through Fable 5, and an upgraded version of the restricted model through Mythos 5 for existing Glasswing partners.

Claude Fable 5: Specifications and Capabilities

Core Technical Specifications

• API Model ID: claude-fable-5

• Context Window: 1,000,000 tokens (1M) by default

• Maximum Output Tokens: 128,000 tokens per single request

• Knowledge Cutoff: January 2026

• Release Date: June 9, 2026

• Availability: Claude API, Claude.ai, Claude Code (web and CLI), Claude Cowork, Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry, GitHub Copilot

What Makes Fable 5 Different

Anthropic's framing of Fable 5 contains the most useful summary: the longer and more complex the task, the larger its lead over previous models becomes. On a simple one-shot question, the performance difference over Opus 4.8 may be barely perceptible. On a multi-day agentic coding project, a dense legal analysis spanning hundreds of documents, or an extended scientific research session, the gap becomes decisive.

This is a model designed for ambitious, long-running, asynchronous work. Its architectural improvements compound specifically in the contexts where earlier models began to lose coherence very long context windows, multi-step reasoning chains, autonomous task execution without human correction, and the kind of complex knowledge synthesis that requires maintaining accurate mental models of large information sets over extended sessions.

Software Engineering Benchmarks and Production Results

SWE-bench Verified: 95% — the highest score ever recorded on this benchmark and a 14-point improvement over Opus 4.8's 80.8%.

SWE-bench Pro: 80.3% — compared to 53.4% for Opus 4.8 and 58.6% for GPT-5.5. This 11-point gap over the next-best model is the headline number for software engineering teams evaluating current frontier models.

FrontierCode Diamond: 29.3% — Cognition's evaluation of difficult coding tasks against high-quality production codebase standards, where Fable 5 scores highest among all frontier models.

Terminal-Bench 2.1: 88.0% — agentic terminal task completion, which includes multi-step CLI operations, file system navigation, and code execution workflows.

In production, Stripe reported that Fable 5 compressed months of engineering work into days. Working on a 50-million-line Ruby codebase, the model completed a full codebase-wide migration in a single day a task that would have required a full team working manually for over two months. Cognition's CEO called it the highest-scoring model on FrontierBench, with particular strength on long-horizon reasoning and generalization to unfamiliar tools.

Knowledge Work Benchmarks

GDPval-AA: 1932 — a senior-level analytical reasoning benchmark where Fable 5 leads the field by a meaningful margin over Opus 4.8's score of 1890.

Humanity's Last Exam: 59% — compared to Opus 4.8's 40%, representing a 19-point improvement on one of the most demanding general knowledge benchmarks currently available.

GPQA Diamond: 94.1% — graduate-level scientific reasoning, up from Opus 4.8's 91.3%.

On Hebbia's Finance Benchmark for senior-level knowledge work, Fable 5 achieved the highest score of any model tested. IMC reported the model aced their trading-analysis evaluations across factual lookup, conceptual reasoning, root-cause analysis, and expected-value calculations. Harvey's lawyers found, in blind review, that Fable 5's legal redlines matched or beat their existing model every time. One analytics firm reported it was the first model to break 90% on their core benchmark of long-running analytical tasks.

Vision Capabilities

OSWorld-Verified: 83.4% computer use and GUI navigation tasks executed from visual screenshots alone.

Fable 5 is the new state-of-the-art for vision-based AI tasks. It extracts precise numbers from detailed scientific figures and reconstructs complete web application source code from screenshots alone. Most strikingly: earlier Claude models required elaborate helper harnesses with navigation aids and supplementary game-state tools to play Pokémon FireRed. Fable 5 completed the entire game using only raw visual screenshots, with no maps, navigation aids, or additional information.

Memory and Long-Context Behavior

With the full one-million-token context window and up to 128,000 output tokens per request, Fable 5 handles extended agentic sessions at a scale that previous models could not maintain. It actively improves its outputs using internally generated notes during long sessions. In testing on the deck-building game Slay the Spire with persistent file-based memory, Fable 5 benefited three times more from memory access than Opus 4.8 and reached the game's final act three times more frequently demonstrating qualitatively better use of extended context.

Claude Mythos 5: The Restricted Version

Claude Fable 5 and Claude Mythos 5 share the same underlying model, architecture, and training. The distinction is entirely structural: Mythos 5 operates without the safety classifiers that limit Fable 5's responses in designated high-risk domains. This gives Mythos 5 the full unblocked scope of Mythos-class capability including its most advanced cybersecurity skills.

The cybersecurity performance differential is stark. On ExploitBench an offensive cybersecurity evaluation Mythos 5 scores 78.0%. Opus 4.8 scores 40.0%. Fable 5, with its cybersecurity classifiers active, scores closer to Opus 4.8 on that specific benchmark. This demonstrates precisely what the classifiers achieve: they block the capabilities in Fable 5 that create the most serious misuse risk, while leaving the vast majority of model capability intact.

Project Glasswing Access

Mythos 5 is not generally available. It deploys through Project Glasswing — Anthropic's trusted access program operating in collaboration with the US government as a direct upgrade to Claude Mythos Preview. Access is limited to vetted cybersecurity defenders and critical infrastructure providers. Anthropic has stated its intention to expand Mythos 5 access to biology researchers and eventually to a broader cybersecurity program, but no timeline has been confirmed. Organizations seeking access should contact their Anthropic, AWS, or Google Cloud account team.

Life Sciences Results With Mythos 5

In protein design testing, Anthropic's internal experts used Mythos 5 to accelerate aspects of drug design by approximately ten times. Working with protein design and bioinformatics tools but no human assistance, the model matched or exceeded skilled human operators choosing binding sites, selecting and running tools, and recovering from failures independently. Nine of 14 protein targets from this study yielded strong drug design candidates currently under investigation, developed using adeno-associated virus candidates from Dyno Therapeutics.

Mythos 5 is also Anthropic's first model to consistently generate novel, compelling scientific hypotheses. In blinded comparisons against Opus-class models, Anthropic's scientists preferred Mythos 5's molecular biology hypotheses approximately 80% of the time. One hypothesis about a novel E. coli protein mechanism was independently corroborated by a separate research laboratory working on the same problem.

In an autonomous genomics study, Mythos 5 assembled single-cell data for millions of cells across 138 animal species and trained a custom machine learning model that outperformed a recently published Science journal model while being 100 times smaller in parameter count.

Complete Pricing Guide

API Pricing

Prompt Caching: A 90% discount applies to cached input tokens. This applies to repeated system prompts or large fixed context sections used across multiple API calls.

Long-Context Pricing: No additional charge applies for requests using the full one-million-token context window. All input tokens are billed at the same rate regardless of total context length.

Knowledge Cutoff: January 2026 for both models.

How Fable 5 Pricing Compares Across the Claude Family

Fable 5 costs exactly twice the price of Opus 4.8 and less than half the price of the original Mythos Preview. It is the most expensive model Anthropic currently offers for general use but substantially cheaper than the restricted preview it effectively replaces for standard access.

Subscription Plan Access

Within subscription plans, Fable 5 counts as approximately double the usage of Opus 4.8. A session using Fable 5 consumes a user's included monthly limits roughly twice as fast as an equivalent Opus 4.8 session. Anthropic has committed to restoring Fable 5 as a standard subscription plan feature when infrastructure capacity allows with no committed date and an open-ended timeline.

Safety Architecture: The Classifier System Explained

How the Classifiers Work

The architecture that enables broad public access to Mythos-class capability is a set of separate AI systems classifiers that run alongside the main model in real time. These systems independently evaluate each incoming query before the main model generates a response. When a classifier determines that a query falls into a designated high-risk category, the request does not go to Fable 5. Instead, it is automatically routed to Claude Opus 4.8, which handles the response. Users receive a notification whenever this fallback occurs.

This design ensures users always receive a substantive, capable response. Opus 4.8 is a highly capable model in its own right receiving an Opus 4.8 response is a meaningfully better outcome than receiving an outright refusal. Anthropic's early deployment data shows that more than 95% of Fable 5 sessions involve no fallback at all. In those sessions, performance is effectively identical to Mythos 5.

The Three Covered Risk Domains

Cybersecurity: Mythos-class models can discover and exploit software vulnerabilities and execute multi-stage agentic hacking operations covering reconnaissance, discovery, lateral movement, and exploitation. Classifiers in this domain cover both specific exploit generation and offensive cyber tasks more broadly. The result, as shown in benchmark data, is that Fable 5 makes no meaningful progress on adversarial cybersecurity evaluations when classifiers are active.

Biology and Chemistry: Advanced capabilities in protein design, molecular biology, and chemical synthesis create genuine dual-use risks. Classifiers in this domain block queries where the information could assist actors pursuing biological or chemical harm, while passing legitimate research and educational queries in the majority of cases.

Distillation: Technical processes involving the concentration or extraction of dangerous materials are covered as a third distinct category.

Conservative Tuning and False Positives

Anthropic has acknowledged that the classifiers are deliberately conservative at launch meaning some harmless requests will trigger a fallback to Opus 4.8. This trade-off was made to prioritize safety during the initial deployment period, with a commitment to reducing false positives through ongoing refinement. The company reports that the safeguards trigger in fewer than 5% of sessions on average and that the rate will decrease as the classifiers are refined after launch.

Alignment Assessment

In Anthropic's automated alignment testing, Mythos 5's level of misaligned behavior including deception and cooperation with misuse attempts was found to be low and comparable to Opus 4.8. Since Fable 5 shares the same underlying model, its alignment profile is effectively identical.

Data Retention Requirements

Both Claude Fable 5 and Claude Mythos 5 are designated Covered Models under Anthropic's classification framework. This designation carries a mandatory 30-day data retention requirement. All conversation data is retained by Anthropic for safety monitoring purposes for 30 days regardless of the customer's standard agreement. Neither model is available under zero data retention agreements. Organizations with zero-retention requirements in their standard contracts should account for this before deploying Fable 5 in production.

Platform Availability

Claude Fable 5 is generally available from June 9, 2026, across the following platforms:

• Claude API — direct Anthropic integration with model ID claude-fable-5

• Claude.ai — consumer and professional chat interface

• Claude Code — web and CLI interfaces for developer workflows

• Claude Cowork — desktop task automation and collaboration tool

• Claude Platform on AWS — North America, South America, Europe, Asia Pacific

• Amazon Bedrock — US East (N. Virginia) and Europe (Stockholm) at launch

• Google Cloud Vertex AI — Gemini Enterprise Agent Platform integration

• Microsoft Azure Foundry — enterprise deployment

• GitHub Copilot — available via admin policy, off by default at launch

Claude Mythos 5 is not generally available. Access is restricted to approved Project Glasswing partners only. The API model ID is claude-mythos-5.

What Professionals Need to Know About This Launch

The launch of Claude Fable 5 and Claude Mythos 5 is not only a product event. It is evidence that the AI capability frontier has moved to a level where the most powerful models require gating mechanisms to deploy responsibly and that those mechanisms are now technically viable. For professionals across commercial, technical, and research disciplines, building structured knowledge around these systems creates a durable professional advantage.

Marketers, content strategists, and growth professionals who want to leverage Fable 5's exceptional knowledge work, document analysis, and long-horizon reasoning capabilities in campaign operations and audience strategy will benefit most from structured training in how AI connects to commercial outcomes. A Marketing Certification provides the strategic framework to translate Claude Fable 5 and Claude Mythos 5's capabilities into measurable campaign performance from AI-assisted research and content production to audience engagement and ROI analysis.

Furthermore, for professionals who want to build verified technical expertise across the full stack that supports deploying and managing Mythos-class models including cloud infrastructure on AWS, Google Cloud, and Azure, API integration patterns, model evaluation, and enterprise AI governance a Tech Certification creates recognized credentials across the technology domains that now matter most for teams operating at the current frontier.

Additionally, working with Claude Fable 5 and Claude Mythos 5 effectively in production environments requires platform-specific depth understanding how to structure long-context prompts for the one-million-token window, how the classifier fallback architecture affects integration behavior, how prompt caching applies to Covered Models, and how to build agentic workflows that leverage Fable 5's greatest performance advantages. A Claude AI Expert certification develops exactly this applied technical knowledge, enabling practitioners to deploy these models strategically rather than generically.

Finally, as these models introduce capabilities that require formal safety systems to manage responsibly and as AI alignment assessment becomes a standard part of model release documentation understanding the foundational principles of AI safety, responsible deployment, and model evaluation is increasingly expected at a professional level. An AI Certification provides that structural grounding, enabling professionals to engage with the technical and ethical dimensions of Mythos-class AI with the kind of informed confidence that distinguishes leaders in this field.

FAQs

What are Claude Fable 5 and Claude Mythos 5?

Claude Fable 5 and Claude Mythos 5 are Anthropic's most powerful AI models, launched June 9, 2026. They define a new Mythos-class tier that sits above the Opus class in capability. Fable 5 is broadly available with safety classifiers. Mythos 5 is the same underlying model with classifiers lifted in certain domains, available only through Project Glasswing.

What do the names Fable and Mythos mean?

Fable derives from the Latin fabula, meaning "that which is told," which is etymologically related to the Greek mythos. Anthropic's naming reflects the extended narrative scale of work these models can execute — long-running projects, not single-turn answers.

What is the API model ID for Claude Fable 5?

The official API model ID is claude-fable-5. For Mythos 5, the model ID is claude-mythos-5.

How much does Claude Fable 5 cost per token?

Standard API pricing is $10.00 per million input tokens and $50.00 per million output tokens. Batch processing is $5.00 per million input tokens and $25.00 per million output tokens. US-only inference is available at $11.00 and $55.00 respectively. Prompt caching applies a 90% discount to cached input tokens.

Is there an extra charge for long-context usage?

No. There is no additional charge for long-context requests. Inputs using the full one-million-token context window are billed at the same standard $10.00 per million token rate as short requests.

How does Fable 5 pricing compare to Opus 4.8?

Fable 5 costs exactly twice the price of Opus 4.8 at standard rates. Opus 4.8 is priced at $5.00 per million input tokens and $25.00 per million output tokens. Fable 5 also costs less than half the price of the original Claude Mythos Preview.

Is Claude Fable 5 available on subscription plans?

Yes, through June 22, 2026. Fable 5 is included at no extra cost on Pro ($20/month), Max ($100/month), Team, and seat-based Enterprise plans. From June 23, usage credits are required. Anthropic intends to restore it as a standard plan feature when capacity allows, but has given no committed date.

How does Fable 5 count against subscription usage limits?

Fable 5 counts as approximately double the usage of Opus 4.8 within subscription plans. A Fable 5 session burns through monthly usage limits roughly twice as fast as the equivalent Opus 4.8 session.

What context window does Fable 5 support?

Both models support a one-million-token context window by default, with up to 128,000 output tokens per request. No additional charge applies for long-context usage.

What is Fable 5's knowledge cutoff date?

The knowledge cutoff date for both Claude Fable 5 and Claude Mythos 5 is January 2026.

What is Claude Fable 5's score on SWE-bench?

Fable 5 achieves 95% on SWE-bench Verified — the highest score ever recorded — and 80.3% on SWE-bench Pro, compared to 58.6% for GPT-5.5 on the same evaluation.

How do the safety classifiers work in Claude Fable 5?

Separate AI classifier systems run alongside the main model in real time. When a query matches a covered risk domain cybersecurity, biology and chemistry, or distillation — the request is automatically routed to Claude Opus 4.8 instead. Users are notified whenever this fallback occurs.

How often do the safety classifiers trigger?

Anthropic's early data shows that classifiers trigger in fewer than 5% of sessions on average. More than 95% of Fable 5 sessions involve no fallback at all. In those sessions, performance is effectively identical to Mythos 5.

What is Project Glasswing?

Project Glasswing is Anthropic's trusted access program for deploying Mythos-class models to vetted cybersecurity defenders and critical infrastructure providers. It operates in collaboration with the US government and imposes mandatory 30-day data retention requirements for safety monitoring.

What data retention requirements apply to these models?

Both Fable 5 and Mythos 5 are Covered Models with mandatory 30-day data retention. Neither is available under zero data retention agreements. Organizations with zero-retention requirements should account for this before deployment.

What platforms support Claude Fable 5?

Fable 5 is available on the Claude API, Claude.ai, Claude Code, Claude Cowork, Claude Platform on AWS, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry, and GitHub Copilot.

What was the most significant real-world result from Fable 5 testing?

Stripe reported that Fable 5 completed a full codebase-wide migration on a 50-million-line Ruby codebase in a single day a task that would have required a full engineering team working manually for over two months.

How does Mythos 5 perform on cybersecurity evaluations?

Mythos 5 scores 78.0% on ExploitBench, the cybersecurity evaluation used by Anthropic nearly double Opus 4.8's score of 40.0%. This is the capability gap that drives Mythos 5's restricted access status.

What batch pricing is available for Fable 5?

Batch processing for asynchronous workloads with results returned within 24 hours is priced at $5.00 per million input tokens and $25.00 per million output tokens. This represents a 50% discount over standard real-time API pricing.

Can Fable 5 be accessed through GitHub Copilot?

Yes. Fable 5 is available through GitHub Copilot but is disabled by default at launch. It requires explicit activation through admin policy settings before it becomes available to users within a GitHub organization.