Claude Fable 5 and Mythos 5: Anthropic's Mythos-Class Models — Technical Analysis
On June 9, 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5 — two configurations of the same underlying model that define a new capability tier above the Opus family. Fable 5 is the general-release version, gated by safety classifiers that route high-risk queries to Opus 4.8. Mythos 5 is the unrestricted variant, available only to vetted partners through Project Glasswing.
Within 72 hours, the US government issued the first-ever export control directive on a commercially deployed AI model, forcing a global suspension. By July 1, Fable 5 was back online with improved classifiers. The episode marked a watershed moment for frontier AI deployment: the point at which model capability intersected national security regulation in real time.
This article covers what engineers and decision-makers need to know: the architectural profile, the safety system design, benchmark positioning against GPT-5.5 and Gemini 3.1 Pro, pricing economics, the export-control incident, and practical guidance on where these models fit.
The Mythos-Class Tier
Anthropic's model hierarchy now has four tiers. From highest to lowest capability:
| Tier | Models | Availability |
|---|---|---|
| Mythos-class | Claude Mythos 5 | Glasswing partners only |
| Mythos-class | Claude Fable 5 | General API / subscription |
| Opus-class | Opus 4.8, 4.7, 4.6 | General availability |
| Sonnet-class | Sonnet 5, 4.6, 4.5 | General availability |
| Haiku-class | Haiku 4.5, 3.5 | General availability |
The Mythos-class sits above Opus. The first model at this tier was Claude Mythos Preview, released in April 2026 through Project Glasswing to roughly 50 partner organizations for defensive cybersecurity work. Fable 5 and Mythos 5 represent the second generation of this tier, and the first time Mythos-level capability is accessible outside Glasswing's controlled environment.
Why the names? Anthropic notes that "Fable" derives from the Latin fabula ("that which is told"), akin to the Greek mythos. The two names exist because the safeguards differ. Same model weights. Different safety posture.
Architecture Profile
Anthropic has not disclosed parameter counts, architecture details, or training methodology for the Mythos-class models. What is known from the documentation and system card:
| Specification | Value |
|---|---|
| Context window | 1,000,000 tokens |
| Max output tokens | 128,000 tokens |
| Modalities | Text, image (vision) |
| Adaptive thinking | Always enabled |
| Raw reasoning traces | Not returned to user |
| Knowledge cutoff | Not separately disclosed |
The 1M-token context window matches Opus 4.8. The 128K output ceiling is new and significantly larger than prior Claude models. Adaptive thinking — Anthropic's internal mechanism that lets the model allocate variable compute per token during inference — is always on and cannot be disabled. Raw reasoning chains (chain-of-thought traces) are consumed during generation but not surfaced to the API, which matters for anyone building agentic workflows where observability of the model's internal decision process is required.
The Safety Classifier System
The defining architectural innovation in this release is not a training advance but a deployment mechanism. Fable 5 ships with a suite of classifiers — smaller AI systems that run alongside the main model during inference, inspecting each request before it reaches the model.
How the Classifiers Work
When a user sends a request to Fable 5, the classifiers evaluate whether it falls into one of three restricted categories:
- Cybersecurity: Exploit development, vulnerability discovery, offensive cyber operations, agentic hacking (reconnaissance, lateral movement, exploitation)
- Biology and chemistry: Bioweapons-related queries, dual-use biological research capabilities, viral engineering
- Distillation: Large-scale extraction of model capabilities to train competing models, particularly by actors in authoritarian countries
If a classifier triggers, the request is silently rerouted to Claude Opus 4.8. The user receives a notification that the fallback occurred. The model continues to respond — just from Opus 4.8 rather than Fable 5.
Anthropic tuned the classifiers with a deliberately large safety margin. This means many benign requests are caught. The company states fallback occurs in fewer than 5% of sessions on average. In the remaining 95%+, Fable 5 performs effectively identically to Mythos 5.
Classifier Robustness
Anthropic's public testing shows the classifiers are strong but not impenetrable:
- An external bug bounty produced no universal jailbreaks in over 1,000 hours of testing
- External red-teaming organizations failed to find universal jailbreaks on long-form agentic tasks
- The UK AI Safety Institute (AISI) made progress toward a universal jailbreak within a brief initial testing window
- On single-turn harmful cybersecurity requests, Fable 5 complied with zero out of tested cases, even when 30 different public jailbreak techniques were applied
A universal jailbreak — a prompt or harness that bypasses safeguards entirely — is likely impossible to prevent completely, per Anthropic's own assessment. The design goal is to make jailbreaks sufficiently slow and costly that they can be detected and patched before they are used at scale.
The Mythos 5 Exception
Mythos 5 is the identical model with safety classifiers lifted in specific domains. It is gated by organizational vetting through Project Glasswing, run in coordination with the US government and major infrastructure providers. Glasswing partners — including AWS, Apple, Google, Microsoft, Cisco, CrowdStrike, JPMorgan Chase, NVIDIA, Palo Alto Networks, Broadcom, and the Linux Foundation — can use Mythos 5 for defensive cybersecurity work with the cyber safeguards removed.
Anthropic has announced plans to open a separate trusted access program for biology researchers, where the biology and chemistry safeguards would be lifted while cyber safeguards remain in place. No public timeline has been announced.
The 30-Day Data Retention Requirement
All Mythos-class model traffic is subject to a mandatory 30-day data retention policy, even on surfaces that otherwise support zero data retention. Anthropic states this data is not used for training and is retained solely to monitor patterns of misuse, including multi-request jailbreak attempts that cannot be detected from a single prompt. Human access to the data is logged, and deletion occurs after 30 days in nearly all cases.
Benchmark Analysis
The following table compiles scores from Anthropic's official launch comparison (June 9, 2026). Anthropic reports the higher score of Mythos 5 and Fable 5 for each benchmark; the two are within 1-3 points of each other except where noted. Starred entries reflect Mythos 5 scores where Fable 5's classifiers redirect to Opus 4.8.
| Benchmark | Fable 5 / Mythos 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Agentic Coding | ||||
| SWE-bench Pro | 80.3% | 69.2% | 58.6% | 54.2% |
| FrontierCode Diamond (xhigh) | 29.3% | 13.4% | 5.7% | — |
| Terminal-Bench 2.1 | 88.0% | 82.7% | — | — |
| SWE-bench Verified | 95.5% | — | — | — |
| Knowledge Work | ||||
| GDPval-AA (Elo) | 1932 | 1890 | 1769 | 1314 |
| GDP.pdf vision (no tools) | 29.8% | 22.5% | 24.9% | 16.7% |
| Reasoning | ||||
| HLE (no tools) | 59.0% | 49.8% | 41.4% | 44.4% |
| HLE (with tools) | 64.5% | 57.9% | 52.2% | 51.4% |
| GPQA Diamond | 92.6% | — | 93.6% | 94.3% |
| ARC-AGI-2 | — | 75.8% | 85.0% | 77.1% |
| Vision & Spatial | ||||
| Blueprint-Bench 2 | 38.6% | 14.5% | 36.2% | 26.5% |
| OSWorld-Verified | 85.0% | 83.4% | 78.7% | 76.2% |
| Tool Use | ||||
| AutomationBench | 17.4% | 15.5% | 12.9% | 9.6% |
| Legal Agent Benchmark | 13.3% | 10.4% | 2.1% | 0.0% |
| Cybersecurity* | ||||
| ExploitBench | 78.0% | 40.0% | 34.0% | — |
| Biology* | ||||
| BioMysteryBench (hard) | 46.1% | 40.0% | — | — |
| BioMysteryBench (human-tier) | 83.9% | 80.4% | — | — |
| Healthcare* | ||||
| HealthBench Professional | 66.0% | 56.9% | 51.8% | — |
| Long Context | ||||
| MRCR v2 (128K, 8-needle) | — | 32.2% | 74.0% | 84.9% |
Interpreting the Benchmarks
Coding is the headline story. The 80.3% on SWE-bench Pro represents an 11.1-point jump over Opus 4.8 and a 21.7-point lead over GPT-5.5. On FrontierCode Diamond — the hardest split of Cognition's production-grade coding eval — Fable 5 scores more than double Opus 4.8 (29.3% vs 13.4%) and roughly five times GPT-5.5 (5.7%). The gap widens with task complexity and length. Stripe's reported migration of a 50-million-line Ruby codebase in one day that would have taken a team over two months is the sort of qualitative datapoint that makes the benchmark numbers concrete.
Knowledge work is competitive but not dominant. On GDPval-AA, Fable 5 leads the field at 1932 Elo, but the gap over Opus 4.8 (1890) is narrower than the gap over GPT-5.5 (1769). On GPQA Diamond, Fable 5 scores 92.6%, slightly behind GPT-5.5 (93.6%) and Gemini 3.1 Pro (94.3%) — though all three are near saturation on this benchmark.
Reasoning benchmarks reveal model specialization. GPT-5.5 leads on ARC-AGI-2 (85.0%) — a benchmark that measures abstract visual pattern recognition and adaptation. Fable 5 does not have a published ARC-AGI-2 score. On Humanity's Last Exam, Fable 5 leads both with and without tools. The profile is consistent: Fable 5 excels on open-ended, multi-step tasks where sustained reasoning and tool use are required, while GPT-5.5 leads on pattern-recognition tasks that benefit from its different training distribution.
Long-context retrieval is GPT-5.5's differentiator. On MRCR v2 at 128K context, GPT-5.5 scores 74.0% while Opus 4.8 scores 32.2%. Fable 5's MRCR score is not part of the published comparison. If your workload requires retrieving specific information from very long documents, GPT-5.5 is the current leader, and Gemini 3.1 Pro extends that lead at 2M context.
Cybersecurity and biology scores are Mythos 5. The ExploitBench 78.0%, BioMysteryBench 46.1%, and HealthBench Professional 66.0% scores all reflect the unrestricted Mythos 5. Fable 5's effective scores in these areas sit closer to Opus 4.8 due to classifier fallback.
Pricing and Cost Engineering
Fable 5 and Mythos 5 share identical pricing:
| Component | Price per MTok |
|---|---|
| Standard input | $10.00 |
| Standard output | $50.00 |
| Cached input (hit) | $1.00 |
| Cache write (5-minute TTL) | $12.50 |
| Cache write (1-hour TTL) | $20.00 |
| Batch API input | $5.00 |
| Batch API output | $25.00 |
At 50, Fable 5 costs exactly twice Opus 4.8 (25). This premium is justified on hard coding and knowledge-work tasks where the capability gap is largest, but it makes Opus 4.8 the better default for routine traffic.
Prompt caching is the primary cost lever. Cache reads bill at 12.50/MTok, 1.25x base) is recouped after four cache reads.
The Batch API provides a 50% discount on both input and output for asynchronous workloads. Critically, the multipliers stack: a cache hit served through the Batch API prices input at approximately 10 list price. For batchable, cache-heavy workloads, effective pricing approaches commodity levels.
The Export-Control Suspension: A Timeline
The launch of Fable 5 and Mythos 5 became an inflection point for AI governance. The timeline:
June 9: Anthropic launches Fable 5 (general availability) and Mythos 5 (Glasswing partners). Both models priced at 50.
June 12, 5:21 PM ET: The US Commerce Department issues an export control directive requiring Anthropic to suspend access to both models for foreign nationals. The order cites national security concerns. Because Anthropic cannot verify citizenship in real time across all deployment channels, it disables both models globally.
The catalyst: Amazon researchers had discovered a prompt technique that bypassed Fable 5's classifiers to identify software vulnerabilities and, in one case, produce code demonstrating exploitation. Amazon CEO Andy Jassy reportedly raised the findings with Treasury Secretary Scott Bessent and other senior administration officials. Anthropic contested the severity, demonstrating that every model it tested — including Opus 4.8, Sonnet 4.6, Haiku 4.5, GPT-5.4, GPT-5.5, and Kimi K2.7 — could produce the same outputs.
June 16-30: Anthropic dispatches technical staff to Washington. Working with the government and Amazon, it trains an improved safety classifier that blocks the reported technique in over 99% of cases. The US Commerce Department's Center for AI Standards and Innovation (CAISI) tests both the prior and new classifiers and deems them "extraordinarily strong."
June 26: The US government approves restored access to Mythos 5 for US organizations.
June 30: Export controls are formally lifted.
July 1: Fable 5 is redeployed globally with the improved classifier. Availability is restored on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Cloud platform rollouts (AWS Bedrock, Google Vertex AI, Microsoft Foundry) follow shortly after.
The incident marks the first time the US government has used export controls to halt access to a commercially deployed AI model. It also produced a joint industry effort — Anthropic, Amazon, Microsoft, Google, and other Glasswing partners are developing a consensus framework for assessing jailbreak severity, modeled loosely on CVSS for software vulnerabilities.
Comparative Positioning
Fable 5 vs GPT-5.5
GPT-5.5's strengths are in long-context retrieval (MRCR v2 74.0% vs Opus 4.8's 32.2%), abstract reasoning (ARC-AGI-2 85.0%), and native multimodal support including audio and video. On agentic coding benchmarks, Fable 5 leads decisively: SWE-bench Pro 80.3% vs 58.6%, FrontierCode Diamond 29.3% vs 5.7%. At 30 per MTok, GPT-5.5 is cheaper on output than Fable 5 ($50) but comparable on input.
The practical recommendation: for repository-scale coding, long-horizon agentic tasks, and knowledge work requiring sustained reasoning over tool use, Fable 5 is the stronger choice. For applications centered on long-document retrieval, abstract visual reasoning, or multimodal input beyond text and images, GPT-5.5 is preferable.
Fable 5 vs Gemini 3.1 Pro
Gemini 3.1 Pro's differentiator is context: 2M tokens at 12 per MTok — five times Fable 5's context for one-fifth the input price. It leads on GPQA Diamond (94.3%) and ARC-AGI-2 (77.1%). On coding and agentic benchmarks, it trails significantly: SWE-bench Pro 54.2% vs 80.3%, Legal Agent Benchmark 0.0% vs 13.3%.
The use case for Gemini 3.1 Pro is cost-sensitive deployment at very long context lengths where coding capability is not the primary requirement. For agentic coding workloads, Fable 5's premium is readily justified by the performance delta.
Fable 5 vs Opus 4.8
Opus 4.8 at 25 remains the sensible default for most production traffic. The 11.1-point SWE-bench Pro gap is meaningful but only on hard coding tasks. For chat, drafting, classification, and routine knowledge work, Opus 4.8 delivers comparable results at half the price. The rule of thumb: start with Opus 4.8, profile where it falls short, and escalate to Fable 5 only for the tasks that genuinely require it.
Agentic Coding Implications
The coding benchmark results translate to real differences in agentic workflow capability. The SWE-bench Pro score of 80.3% means Fable 5 can correctly resolve over four out of five real GitHub issues in a production-like setting, compared to roughly three out of five for GPT-5.5 (58.6%) and just over half for Gemini (54.2%).
Cursor CEO Michael Truell called Fable 5 "the state of the art model on CursorBench," saying it has "opened up a class of long-horizon problems that were out of reach for earlier models." Cognition CEO Scott Wu described it as "the highest-scoring model on FrontierBench."
The pattern across user reports is consistent: Fable 5 stays on task longer, validates its own outputs before declaring completion, and handles repository-scale migrations that previous models could not sustain. The 128K output ceiling is material here — it allows the model to produce entire file migrations, multi-file refactors, and comprehensive test suites in a single generation.
Practical Adoption Guidance
When to use Fable 5: Repository-scale code migrations, multi-file refactoring, long-horizon agentic coding tasks, complex knowledge work requiring sustained reasoning over large contexts, document-intensive professional workflows (legal, financial analysis, research), and vision-centric tasks where screenshots or diagrams must be interpreted with high accuracy.
When to use Opus 4.8: Chat, drafting, content generation, classification and extraction pipelines, routine customer-facing interactions, code review and explanation, and any high-volume production workload where per-token cost matters. Opus 4.8 is the cost-efficient default.
When to use GPT-5.5: Applications centered on long-document retrieval, abstract visual reasoning (ARC-AGI-2 type tasks), native audio/video processing, and tasks where the reasoning trace does not need to be the model's internal chain-of-thought.
When to use Gemini 3.1 Pro: Cost-sensitive deployments at very long context lengths (approaching 2M tokens), high-throughput classification workloads where coding capability is secondary, GPQA-domain scientific reasoning, and scenarios where $2/MTok input pricing changes the unit economics.
Safety classifier awareness: If your application touches cybersecurity (including routine vulnerability research or defensive security tooling), biology or chemistry, or involves large-scale prompt engineering workflows, expect classifier fallback. Anthropic's safety margin is deliberately conservative. The improved classifier deployed in July 2026 introduced additional false positives on routine coding and debugging tasks. Plan for fallback handling in your application logic.
Cost optimization: Cache aggressively. For agentic workloads with repeated context, cache reads at 0.50/MTok effective input pricing makes high-volume agentic workloads economically viable.
Related Articles
- AI Agents That Find Zero-Days: The FFmpeg Case Study -- "AI Agents That Find Zero-Days: The FFmpeg Case Study"
- Agentic AI in Practice — "Agentic AI in Practice"
- Prompt Injection in LLM Applications — "Prompt Injection in LLM Applications"
- LLM Runtime Monitoring with OpenTelemetry — "LLM Runtime Monitoring with OpenTelemetry"