Claude Fable 5

Overview

Anthropic released Claude Fable 5 on June 9, 2026 - the first Mythos-class model made publicly available. Where Claude Mythos Preview remained gated to 52 Project Glasswing partners, Fable 5 opens that capability tier to the general API, subscription users, and enterprise customers. The catch: it runs automated safety classifiers that redirect flagged requests to Claude Opus 4.8 rather than answering them directly.

Mythos 5, a less restricted variant for vetted Glasswing partners, launched with Fable 5 on the same date. Mythos 5 posts 78% on ExploitBench (up from 69% for the prior preview version), and Anthropic describes it as targeted at advanced research use cases including genomics and drug design. This page covers Fable 5, the publicly accessible model.

The pricing puts Fable 5 at 2x Claude Opus 4.8's standard rate ($5/$25) and less than half the $25/$125 that Mythos Preview charged. That positions it as an accessible entry point for Mythos-class capability without the restricted-access friction of prior releases.

Key Specifications

Benchmark Performance

Benchmark data comes from Anthropic's release materials and The Decoder's reporting. Opus 4.8 figures are from the Claude Opus 4.8 model card; GPT-5.5 figures come from OpenAI's concurrent disclosure.

The SWE-bench Pro jump is the headline number: a 11-point gain over Opus 4.8 and a 21-point lead over GPT-5.5 on the same split. FrontierCode - Cognition's hardest public coding evaluation - shows an even larger gap: Fable 5's 29.3% is more than double Opus 4.8's 13.4% and roughly 5x GPT-5.5's 5.7%. For context on where these numbers sit in the broader field, see the coding benchmarks leaderboard and the SWE-bench coding agent leaderboard.

Anthropic also reports state-of-the-art scores on finance benchmarks (Hebbia's senior-level finance evaluation), vision tasks (rebuilding web app source code from screenshots, completing Pokémon FireRed from screenshots alone), and gaming benchmarks (Slay the Spire, where it improved 3x more than Opus 4.8). No specific numeric scores are published for those categories yet.

For scientific capabilities, Mythos 5 data is more detailed: 10x acceleration in drug-design workflows per internal protein experts, and genomics outputs that beat a recently published Science paper model despite being 100x smaller. Fable 5's performance on those tasks isn't separately reported, since they fall partly within the safety-restricted domains.

Key Capabilities

Software Engineering at Scale

The Ruby migration task Anthropic highlights is concrete: Fable 5 completed a 50-million-line codebase migration in one day, a job Anthropic estimates would take two months manually. That's an agentic coding result, not a benchmark - it involves planning, execution across a large context window, and handling the edge cases that dominate real migration work.

The model supports long-running, asynchronous execution for complex tasks, and Anthropic added advanced vision capabilities for diagrams, charts, and tables embedded in PDFs and documents. The combination matters for enterprise workflows where structured documents are the primary input format.

Safety Architecture

Three classifier categories can redirect requests to Opus 4.8 instead of Fable 5:

1. Cybersecurity vulnerabilities and attacks
2. Biology and chemistry (Anthropic reports this category sees the most routing)
3. Model distillation attempts

Anthropic says more than 95% of sessions trigger no fallback. When a redirect does occur, customers pay Opus 4.8 pricing for that specific request. Requests blocked mid-conversation incur mixed-tier charges.

The 30-day data retention requirement applies to all Mythos-class traffic. Data is used for abuse detection and human review only, not for training. On Amazon Bedrock, opting into data retention moves traffic outside AWS's standard data boundary, which matters for compliance-sensitive deployments.

Vision and Multimodal Tasks

Fable 5's vision performance stands out enough that Anthropic led with two specific examples rather than a benchmark number: the model reconstructed a web app's source code from screenshots, and completed Pokémon FireRed autonomously using only screenshots as input. Both require spatial reasoning over sequences of frames, not just single-image understanding.

Pricing and Availability

Fable 5 launched at $10/M input tokens and $50/M output tokens across all API access. Subscription plans get a grace period: Pro, Max, Team, and Enterprise plans include Fable 5 at no extra cost through June 22, 2026. Starting June 23, access on those plans shifts to usage credits pending capacity expansion.

Platforms at launch:

• Claude API (direct)
• GitHub Copilot
• Amazon Bedrock (US East N. Virginia, Europe Stockholm; expanding)

Bedrock deployment requires opt-in to data retention via the Data Retention API before the model is accessible. Google Cloud Vertex AI access hasn't been confirmed for Fable 5 at launch (Mythos Preview was available on Vertex; Fable 5's status there differs by region and rollout phase).

Pricing comparison against the existing Claude lineup:

Strengths and Weaknesses

Strengths

• SWE-bench Pro at 80.3% is the highest published score in the Claude family and leads GPT-5.5 by 21 points
• FrontierCode at 29.3% is more than double the next closest competitor on that evaluation
• Safety fallback is transparent: customers know when a redirect happens and pay Opus 4.8 rates for it
• Available on GitHub Copilot from day one - broader developer reach than Mythos Preview had
• 1M token context window at standard Fable pricing (no surcharge)

Weaknesses

• 2x the per-token cost of Claude Opus 4.8 - for most general-purpose tasks, Opus 4.8 is still the better price-performance option
• Safety classifiers reduce effective capability in cybersecurity and biology, the domains where Mythos-class performance matters most
• GPQA Diamond, Chatbot Arena Elo, and most science benchmarks not yet reported for Fable 5 specifically
• 30-day mandatory data retention is a non-starter for some regulated industries, especially on Bedrock
• No Vertex AI availability confirmed at launch; limits Google Cloud deployments

Related Coverage

• Claude Mythos Preview - the restricted predecessor with no safety fallback
• Claude Opus 4.8 - the current public flagship that serves as Fable 5's safety fallback model
• Coding Benchmarks Leaderboard - full rankings context for SWE-bench and FrontierCode
• SWE-bench Coding Agent Leaderboard - detailed SWE-bench breakdown by model

FAQ

Is Claude Fable 5 the same as Claude Mythos?

No. Fable 5 is a Mythos-class model with additional safety classifiers that fall back to Opus 4.8 for high-risk requests. Mythos 5 - released the same day - is a separate model for vetted Glasswing partners with fewer restrictions and higher scores on cybersecurity benchmarks like ExploitBench.

How does the safety fallback affect pricing?

When a request triggers a redirect to Opus 4.8, you pay Opus 4.8 rates for that request. If a request is blocked mid-conversation, mixed-tier charges apply. Anthropic reports fewer than 5% of sessions trigger any fallback.

Can I use Claude Fable 5 on Amazon Bedrock?

Yes, but you must opt into data retention via Bedrock's Data Retention API first. Opting in moves your data outside AWS's standard security boundary for safety monitoring purposes.

What happens to Fable 5 pricing after June 22?

Free access on Pro, Max, Team, and Enterprise plans ends June 23. After that date, those plans shift to usage credits for Fable 5, subject to capacity. API pricing stays at $10/$50 per million tokens.

Where does Fable 5 rank on SWE-bench?

At 80.3% on SWE-bench Pro at launch, Fable 5 leads all publicly disclosed scores on that benchmark. Full rankings are tracked on the SWE-bench coding agent leaderboard.

Sources:

• Anthropic releases Claude Fable 5 - TechCrunch
• Claude Fable 5 and Mythos 5 with major gains in coding and science - The Decoder
• Anthropic launches Claude Fable 5, most powerful public model with gated safeguards - TechTimes
• Anthropic Claude Mythos launches today - Yahoo Finance
• Claude Fable 5 on AWS - Amazon Web Services Blog
• Anthropic launches Claude Fable 5: Mythos-class AI with cybersecurity guardrails - SecurityWeek