Claude Mythos 5
Claude Mythos 5 is the full release of Anthropic's restricted Mythos family - same weights as Fable 5 but without safety classifiers for cybersecurity and biology, at $10/M input and $50/M output tokens.
10 min read
Overview
Claude Mythos 5 is Anthropic's full release of the Mythos model family, launched June 9, 2026, alongside Claude Fable 5. Where Fable 5 is the public-facing model with automated safety classifiers that redirect certain queries to Claude Opus 4.8, Mythos 5 removes those classifiers for cybersecurity and biology requests. Access remains gated to Project Glasswing partners and approved biology researchers - the same restricted program that governed Claude Mythos Preview since April 2026.
TL;DR
- Full release of the Mythos family - same weights as Fable 5, without safety classifiers for offensive security and biology
- $10/M input, $50/M output - half the $25/$125 pricing of Mythos Preview; restricted to ~100 Glasswing partners
- Access was suspended globally June 12 by US export order, then restored for ~100 trusted US partners on June 26; Fable 5 remains blocked
Mythos 5 shares its weights with Fable 5 - both are built on the same underlying model. The distinction is operational: Fable 5 routes flagged requests to Opus 4.8 as a fallback, while Mythos 5 answers them directly. On ExploitBench, a cybersecurity evaluation that measures autonomous exploit development and vulnerability discovery, Mythos 5 scores 78% - up from 69% on the prior Mythos Preview. The Preview was released in April 2026 at $25/$125 per million tokens. Mythos 5 cuts that to $10/$50, making it meaningfully cheaper for the organizations that can access it.
The model's short public life has been complicated by US government action. On June 12, Commerce Secretary Howard Lutnick signed an export control directive that took both Fable 5 and Mythos 5 offline globally. On June 26, Mythos 5 access was restored for roughly 100 vetted US partners under updated export control conditions. Fable 5 remains suspended as of June 27, 2026.
Key Specifications
| Specification | Details |
|---|---|
| Provider | Anthropic |
| Model Family | Claude |
| Parameters | Not disclosed |
| Context Window | 1M tokens |
| Input Price | $10.00/M tokens |
| Output Price | $50.00/M tokens |
| Release Date | June 9, 2026 |
| License | Proprietary - restricted access |
| Availability | Project Glasswing partners, approved biology researchers (~100 US partners as of June 26) |
| Safety Fallback | None (Fable 5 has fallback; Mythos 5 doesn't) |
| Data Retention | 30-day mandatory for safety monitoring |
The pricing drop from Mythos Preview's $25/$125 to $10/$50 is sizable. For organizations that have been running Preview under Glasswing, Mythos 5 cuts input costs by 60% and output costs by 60%, with identical architecture and improved benchmark scores.
Benchmark Performance
Benchmark data comes from Anthropic's June 9 release materials and the Claude Fable 5 model card which covers the shared benchmark set. Mythos-specific figures (ExploitBench, CyberGym) are where the two variants diverge meaningfully.
| Benchmark | Mythos 5 / Fable 5 | Mythos Preview | Claude Opus 4.8 | GPT-5.5 |
|---|---|---|---|---|
| SWE-bench Pro | 80.3% | Not reported | 69.2% | 58.6% |
| SWE-bench Verified | 95.0% | 93.9% | 88.6% | - |
| FrontierCode Diamond | 29.3% | Not reported | 13.4% | 5.7% |
| GPQA Diamond | 92.6% | 94.6% | - | - |
| HLE (with tools) | 64.5% | 64.7% | ~53% | - |
| ExploitBench | 78% | 69% (CyberGym equiv.) | - | - |
| Intelligence Index | 64.9 | Not reported | ~59.0 | ~59.9 |
The ExploitBench jump from 69% (Mythos Preview's CyberGym equivalent) to 78% is the clearest capability signal specific to Mythos 5. That benchmark measures end-to-end autonomous exploit development: finding vulnerabilities, writing working exploits, and chaining them into multi-step attacks. The 9-point improvement reflects better tool use and longer-horizon planning across attack simulations, not just improved static analysis.
On shared coding metrics, Fable 5 and Mythos 5 post identical scores since the weights are the same. The SWE-bench Verified improvement from 93.9% (Preview) to 95.0% is the headline coding number. FrontierCode Diamond at 29.3% is the most significant competitive signal: Opus 4.8 scores 13.4% on that evaluation, and GPT-5.5 scores 5.7%. A gap that large doesn't close with prompting or scaffolding.
The GPQA Diamond score (92.6%) is slightly below Mythos Preview's 94.6%. That's not a regression in the usual sense - both models score well above expert PhD level - but it's worth flagging that scientific reasoning didn't improve monotonically between Preview and the full release. Full context on ranking positions is available on the coding benchmarks leaderboard and the reasoning benchmarks leaderboard.
Mythos 5's ExploitBench score of 78% represents end-to-end autonomous exploit development - finding vulnerabilities, writing exploits, and chaining multi-step attacks without safety redirection.
Source: pexels.com
Key Capabilities
Cybersecurity - No Classifier Redirection
The core capability difference between Mythos 5 and Fable 5 is what happens when a prompt involves offensive security. Fable 5 redirects those requests to Opus 4.8. Mythos 5 answers them directly with Mythos-class capability.
Anthropic's internal testing, documented during the Mythos Preview launch, produced thousands of zero-day discoveries across major operating systems and browsers, including a 27-year-old TCP vulnerability in OpenBSD and a 16-year-old memory corruption bug in FFmpeg. Mythos 5 improves on that baseline. The 78% ExploitBench score puts it above Mythos Preview's reported CyberGym performance and above GPT-5.5-Cyber on comparable offensive security evaluations. The UK AI Safety Institute's evaluation of the Preview - 73% success on expert-level CTF challenges, the first AI to hit that tier - established the capability floor. Mythos 5 is trained on updated data and refines the agentic scaffolding.
The model isn't jailbreak-proof. A narrow bypass was demonstrated during the June 2026 export control dispute, involving prompting the model to review a codebase and identify software flaws. Anthropic described this as a "narrow, non-universal jailbreak." The access suspension and the subsequent deal allowing 100 US partners back on the model came with updated usage controls addressing that specific vector.
Biology and Genomics Research
Anthropic describes Mythos 5 as targeted at "advanced research use cases including genomics and drug design" - language that doesn't appear in Fable 5's description. Internal testing showed 10x acceleration in drug-design workflows per biology domain experts, and genomics output that beat a recently published Science paper model while running at 100x smaller scale.
Fable 5 routes many biology requests to Opus 4.8 due to classifier triggering. Mythos 5 handles those requests with full capability. For approved biology researchers - the second access track with Glasswing partners - that distinction is the whole reason to request Mythos 5 over Fable 5.
Anthropic restricts Mythos 5 biology access to approved researchers working on genomics and drug design - the same domains where Fable 5's safety classifiers most frequently redirect to Opus 4.8.
Source: pexels.com
Agentic Coding
The shared benchmarks with Fable 5 mean Mythos 5 inherits the full coding capability tier. SWE-bench Pro at 80.3% leads all published scores. FrontierCode Diamond at 29.3% is more than double Opus 4.8 and roughly five times GPT-5.5. Long-horizon autonomous work - the model supports sessions up to 12 hours with maintained context - is qualitatively different from what Opus 4.8 can sustain.
For Glasswing partners whose workflows involve security-adjacent coding (penetration testing tools, exploit frameworks, defensive security systems), Mythos 5 is the only model that doesn't redirect those tasks mid-session. That operational continuity has real value in agentic pipelines where a mid-task fallback to a weaker model breaks the output.
Pricing and Availability
Mythos 5 launched at $10/M input and $50/M output - a 60% reduction from Mythos Preview's $25/$125 pricing. The model isn't on the standard API and has no free tier or waitlist.
April 7, 2026 - Mythos Preview launches under Project Glasswing at $25/$125. Access limited to 52 organizations.
June 9, 2026 - Mythos 5 launches alongside Fable 5 at $10/$50. Access expands to ~100 Glasswing partners plus approved biology researchers.
June 12, 2026 - US Commerce Department directive suspends both Fable 5 and Mythos 5 globally. Trigger: a jailbreak disclosure and unauthorized Glasswing roster expansion including SK Telecom.
June 26, 2026 - Mythos 5 access restored for ~100 vetted US partners under updated export control conditions. Fable 5 remains suspended.
The June 26 partial restoration is significant: Mythos 5 - the more restricted, less publicly accessible model - was cleared before Fable 5. The US government's concern was specifically about uncontrolled access to the model's offensive security capabilities. Restoring Mythos 5 to a vetted list of US partners is consistent with that framing. Fable 5's broader public API surface creates a different compliance challenge, and no restoration date has been given.
Pricing comparison:
| Model | Input | Output | Access |
|---|---|---|---|
| Claude Mythos 5 | $10.00/M | $50.00/M | ~100 vetted US Glasswing partners |
| Claude Fable 5 | $10.00/M | $50.00/M | Public API (suspended as of June 27) |
| Claude Mythos Preview | $25.00/M | $125.00/M | 52 Glasswing partners (superseded) |
| Claude Opus 4.8 | $5.00/M | $25.00/M | Public API |
The batch API is available at 50% discount, bringing Mythos 5 to $5/M input and $25/M output for async jobs - identical to Opus 4.8's standard rate. For Glasswing partners running large-scale vulnerability scanning or drug design pipelines where real-time latency isn't required, batch pricing changes the economics much.
Strengths and Weaknesses
Strengths
- ExploitBench at 78% leads all published scores on autonomous exploit development, up 9 points from Mythos Preview
- No classifier fallback for cybersecurity and biology - full Mythos-class capability for those domains without mid-session redirects
- 60% cheaper than Mythos Preview at $10/$50 vs $25/$125 per million tokens
- 95.0% SWE-bench Verified improves on Preview's 93.9%, maintaining the field lead on software engineering
- 30-day data retention uses a safety monitoring purpose only, not for training
Weaknesses
- Access restricted to ~100 vetted US partners after June 26; not available outside that list
- US export controls make international deployment legally constrained even for Glasswing partners
- No published scores for Chatbot Arena Elo, GDPval, or ARC-AGI-2, leaving some competitive comparisons incomplete
- Parameters and architecture not disclosed
- GPQA Diamond at 92.6% is slightly lower than Mythos Preview's 94.6% - not a regression in practical terms but a metric worth noting
- A demonstrated jailbreak was the proximate cause of the June 12 shutdown; updated controls don't guarantee the vector is fully closed
Related Coverage
- Claude Mythos Preview - the April 2026 predecessor with complete benchmark data and the Project Glasswing backstory
- Claude Fable 5 - the public-access model sharing Mythos 5's weights, with safety classifiers
- US Export Order Forces Global Fable 5, Mythos 5 Shutdown - the June 12 directive and its immediate impact
- How Amazon CEO Triggered the Fable 5 Shutdown - the chain of events that produced the global recall
- Anthropic Ships $100M AI Cyber Defense to 12 Rivals - Project Glasswing founding and structure
- Coding Benchmarks Leaderboard - full SWE-bench and FrontierCode ranking context
FAQ
Is Claude Mythos 5 available to the public?
No. Mythos 5 is restricted to roughly 100 vetted US partners under Project Glasswing and approved biology researchers. There's no waitlist and no self-serve sign-up. As of June 26, 2026, access is further constrained to US partners by export control conditions.
How does Mythos 5 differ from Fable 5?
They share the same underlying weights. The difference is the safety classifier layer: Fable 5 routes cybersecurity, biology, and chemistry requests to Claude Opus 4.8 as a fallback. Mythos 5 answers those requests directly with full Mythos-class capability. The ExploitBench score (78% for Mythos 5) reflects what the model can do without that redirect.
What happened with the June 12 shutdown?
The US Commerce Department ordered Anthropic to disable Fable 5 and Mythos 5 globally after a jailbreak demonstration and the discovery that Anthropic had expanded Glasswing access to new organizations - including SK Telecom - without prior government clearance. Mythos 5 access was restored for ~100 vetted US partners on June 26. Fable 5 remains suspended as of June 27.
How does Mythos 5 pricing compare to Mythos Preview?
Mythos 5 costs $10/M input and $50/M output. Mythos Preview cost $25/$125. That's a 60% reduction in both input and output pricing. Benchmark performance improved as well: ExploitBench went from 69% (Preview's CyberGym equivalent) to 78%, and SWE-bench Verified improved from 93.9% to 95.0%.
Can organizations outside the US access Mythos 5?
The June 26 restoration was specifically for vetted US partners. Export control conditions apply. International Glasswing partners that had access before June 12 do not have access restored as of June 27. No timeline has been given for restoring international access.
Sources:
- Anthropic: Claude Fable 5 and Claude Mythos 5 announcement
- Claude Fable 5 and Mythos 5 with major gains in coding and science - The Decoder
- Anthropic releases Claude Fable 5 - TechCrunch
- Anthropic: Statement on the US government directive to suspend access to Fable 5 and Mythos 5
- TechCrunch: Anthropic's safety warnings may have just backfired
- Anthropic Claude Mythos launches today - Yahoo Finance
- Claude Fable 5 Benchmark Scores - Weights and Biases
- Claude Mythos Preview - Anthropic Research
- AISI Evaluation of Claude Mythos Preview Cyber Capabilities
✓ Last verified June 27, 2026
