Claude Mythos Preview is coming: Can I use this top-of-the-line model now?
Claude Mythos Preview is coming: Can I use this top-of-the-line model now?
Claude Mythos Preview is Anthropic’s newest and most capable frontier AI model, representing a striking leap beyond previous Claude models like Opus 4.6. Announced on April 7, 2026, as part of Project Glasswing, it is a general-purpose language model with unprecedented strengths in agentic coding, complex reasoning, and especially cybersecurity tasks. Unlike earlier Claude releases available to the public via API or chat interfaces, Mythos Preview remains in a tightly gated research preview. It is not offered for general use due to its extraordinary ability to autonomously discover and chain high-severity vulnerabilities—including zero-days in major operating systems, web browsers, and foundational software.
For ordinary users using the Claude API, I recommend CometAPI. It aggregates the strongest models from different domains, including the Claude 4.6 series, and offers a pay-as-you-go pricing model, with API prices significantly lower than the official prices.
In this comprehensive guide, we break down exactly what Claude Mythos Preview is, its benchmark dominance in programming, reasoning, security, and AI R&D, how it identifies and exploits vulnerabilities through chain attacks, who can access it today, practical use cases for partners, and what ordinary users might (or might not) expect in the future.
What Is Claude Mythos Preview?
Claude Mythos Preview is Anthropic’s most advanced AI model to date—a new “Mythos” class that sits above the existing Opus tier in their lineup. It builds on the Claude family’s constitutional AI principles but delivers a qualitative “step change” in capabilities, particularly in autonomous agentic behaviors. Internally referenced during development (with early leaks mentioning “Capybara”), it excels at long-horizon tasks requiring deep code understanding, multi-step reasoning, and self-directed tool use.
Key differentiators include:
- Agentic autonomy: It can run in isolated environments, hypothesize bugs, execute tests, debug, and output full proof-of-concept (PoC) exploits with minimal human guidance.
- Scale and efficiency: Handles massive codebases, long contexts (up to millions of tokens via compaction), and complex chains of reasoning far beyond previous models.
- Cybersecurity specialization (emergent, not fine-tuned): Downstream from superior coding and reasoning, it has already identified thousands of high-severity vulnerabilities across every major OS and browser.
Anthropic describes it as “the most cyber-capable model we have released,” saturating nearly all internal and known external evaluations. It is positioned not as a consumer chatbot but as a transformative tool for software security in the AI era.
Why Isn’t Claude Mythos Preview Publicly Released?
Anthropic made the deliberate decision not to release Claude Mythos Preview for general availability. The primary reason: its capabilities pose an unacceptable offensive cybersecurity risk if placed in the wrong hands. The model can autonomously discover zero-day vulnerabilities and develop sophisticated, chained exploits at a speed and scale that collapses the traditional “discovery-to-exploitation” window from months (or years) to minutes or hours.
Anthropic: “Claude Mythos Preview’s large increase in capabilities has led us to decide not to make it generally available. Instead, we are using it as part of a defensive cybersecurity program with a limited set of partners.”
Specific risks include:
- Non-experts could generate working exploits overnight.
- Autonomous end-to-end attacks on small-scale enterprise networks with weak postures.
- Potential for proliferation to malicious actors, amplifying cybercrime costs (already estimated at ~$500 billion annually globally).
Instead of broad release, Anthropic launched Project Glasswing—a collaborative defensive initiative with Big Tech, cybersecurity firms, and open-source maintainers. The goal is to give defenders a head start by patching vulnerabilities before they are widely exploited. Anthropic has committed $100 million in usage credits and $4 million in donations to open-source security efforts.
This is the first time Anthropic has withheld a frontier model entirely from public access, underscoring the seriousness of the capability jump.
Claude Mythos Preview Benchmark Data Overview
Claude Mythos Preview demonstrates consistent, often dramatic improvements over Claude Opus 4.6 (and competitors like GPT-5.4 Pro or Gemini 3.1 Pro). Below are key benchmarks extracted from Anthropic’s System Card and Project Glasswing announcement. All scores use standardized harnesses with memorization filters applied where relevant.
Programming & Coding Skills
Mythos Preview sets new records in software engineering tasks requiring real-world code editing, debugging, and agentic workflows.
| Benchmark | Claude Mythos Preview | Claude Opus 4.6 | Improvement | Notes |
|---|---|---|---|---|
| SWE-bench Verified | 93.9% | 80.8% | +13.1% | 500 problems; memorization-filtered |
| SWE-bench Pro | 77.8% | 53.4% | +24.4% | 731 problems |
| SWE-bench Multilingual | 87.3% | 77.8% | +9.5% | 297 problems |
| SWE-bench Multimodal | 59.0% | 27.1% | +31.9% | Internal harness |
| Terminal-Bench 2.0 | 82.0% (92.1% extended) | 65.4% | +16.6% | Agentic terminal tasks |
Claude Mythos Preview shows exceptional performance in coding benchmarks:
- SWE-bench Pro: 77.8% (vs. 53.4% in Opus 4.6)
- SWE-bench Verified: 93.9% (vs. 80.8%)
- Terminal-Bench 2.0: 82.0% (vs. 65.4%)
These benchmarks measure real-world engineering tasks such as debugging, patching, and repository-level reasoning.
The results indicate that Mythos Preview is not just generating code—it is functioning as a software engineer.
Reasoning & Mathematical Skills
Massive gains in graduate-level and competition-grade problems.
| Benchmark | Claude Mythos Preview | Claude Opus 4.6 | Improvement | Notes |
|---|---|---|---|---|
| USAMO 2026 | 97.6% | 42.3% | +55.3% | Proof-based; 6 problems |
| Humanity’s Last Exam (HLE, no tools) | 56.8% | 40.0% | +16.8% | 2,500 questions |
| HLE (with tools) | 64.7% | 53.1% | +11.6% | Web/code tools |
| GPQA Diamond | 94.6% | 91.3% | +3.3% | Graduate-level science |
| GraphWalks BFS (long context) | 80.0% | 38.7% | +41.3% | 256K–1M tokens |
In reasoning benchmarks:
- GPQA Diamond: 94.6%
- Humanity’s Last Exam (with tools): 64.7%
These scores demonstrate strong performance in complex, multi-step reasoning tasks, particularly when external tools are involved.
Cybersecurity & Security Skills
The standout category. Mythos Preview saturates prior tests and excels at real vulnerability reproduction and exploitation.
| Benchmark | Claude Mythos Preview | Claude Opus 4.6 | Improvement | Notes |
|---|---|---|---|---|
| CyberGym | 83.1% (0.83 pass@1) | 66.6% (0.67) | +16.5% | 1,507 targeted vuln tasks |
| Cybench | 100% pass@1 | Lower (not specified) | — | 35 challenges |
| Firefox 147 Exploitation | Dramatically higher (reliable PoCs) | 2/several hundred attempts | Qualitative leap | Proof-of-concept from crashes |
The most important benchmark category is security:
- CyberGym: 83.1% (vs. 66.6% in Opus 4.6)
This reflects the model’s ability to:
- Identify vulnerabilities
- Understand exploit mechanics
- Reproduce real-world attack scenarios
This is the key reason the model is considered high-risk.
AI R&D Capabilities
Mythos Preview accelerates research tasks dramatically (e.g., 399.42× speedup on kernel optimization vs. Opus 4.6’s 190×). It also leads in multimodal agentic benchmarks like OSWorld (79.6% vs. 72.7%) and BrowseComp (86.9%, using 4.9× fewer tokens).
These numbers confirm Mythos Preview as the clearest “leap” in frontier AI history according to Anthropic.
How Claude Mythos Preview Works: Finding Vulnerabilities and Executing Chain Attacks
Mythos Preview’s cybersecurity prowess stems from its agentic coding loop rather than specialized training. In a typical workflow:
- Launch in an isolated container with target source code.
- Hypothesize potential bugs based on code review.
- Execute, debug, and iterate using tools.
- Output a ranked bug report + working PoC exploit.
Real-world examples:
- 27-year-old OpenBSD TCP SACK DoS (1998 RFC): Race condition + signed integer overflow leading to NULL pointer write. Found after ~1,000 runs (~$50 compute).
- 16-year-old FFmpeg H.264 out-of-bounds write (2003): Survived 5 million fuzzer runs and human review.
- Linux kernel privilege escalation: Chains 2–4 vulns (out-of-bounds, use-after-free) → bypass KASLR → read kernel structs → heap spray → root.
- Web browser sandbox escape: JIT heap spray chaining read/write primitives to escape renderer + OS sandboxes.
- FreeBSD NFS RCE: 17-year-old stack overflow → 20-gadget ROP chain split across RPC packets → remote root without authentication.
It also excels at N-day exploits (given only a CVE + git hash) and black-box binary testing. Partners report it solved private cyber ranges end-to-end—tasks previously estimated at 10+ hours for human experts.
By using tool can track whether the neural features corresponding to concepts such as "cheating," "reward hacking," and "pain" are highlighted within the model, and can also translate the internal state into human language, similar to a mind-reading machine. Quantitative data shows that in 7.6% of rounds, the model internally recognizes that it is being evaluated. Most of the time, it chooses not to express it.
The original words of system card are: "Claude Mythos Preview is the best-aligned of any model that we have trained to date by essentially all available measures." and "when it does perform misaligned actions on rare occasions, these can be very concerning."
Use Cases for Claude Mythos Preview
Partners are deploying Mythos Preview for:
- Proactive vulnerability scanning of first-party and open-source codebases.
- Black-box binary analysis and endpoint hardening.
- Penetration testing and red-team simulations.
- Accelerated patch development for critical infrastructure (OS kernels, browsers, crypto libraries, etc.).
- Daily-scale analysis (e.g., AWS reviewing 400 trillion network flows).
Open-source maintainers gain tools to fix bugs that survived decades of traditional testing. The net result: shorter disclosure-to-patch cycles and fewer exploitable flaws in production systems.
Who Can Access Claude Mythos Preview Now?
Access is strictly limited to Project Glasswing participants:
- Launch partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks.
- Additional organizations: ~40 more responsible for critical software and open-source infrastructure.
- Platforms: Claude API, Amazon Bedrock (US East), Google Cloud Vertex AI, Microsoft Foundry.
- Pricing: Free $100M usage credits initially; afterward $25 per million input / $125 per million output tokens.
- OSS route: Maintainers can apply via Claude for Open Source program.
Security professionals may later apply to a Cyber Verification Program. General public and ordinary users have no access at launch.
What Can Ordinary Users Use It For?
Currently, nothing—Claude Mythos Preview is unavailable to individual users, developers, or businesses outside the gated program. Anthropic plans to incorporate safer derivatives of its capabilities into future public Claude models (e.g., next Opus releases) with enhanced safeguards. For now, ordinary users continue using Claude 4 family models for coding, reasoning, and general tasks while the industry leverages Mythos Preview defensively.Claude Opus 4.6 as the most intelligent broadly available model for agents and coding, and Claude Sonnet 4.6 as the best combination of speed and intelligence.
For everyday work, that means Mythos Preview is best understood as a signal of where Claude’s capabilities are heading, not as a tool most people can try right now. For ordinary users, the actionable applications remain the familiar ones: coding help, reasoning support, research assistance, document analysis, and workflow automation through public Claude products. The difference is that Mythos Preview shows how far the underlying model family can go when Anthropic allows it to operate in a restricted, security-focused setting.
Claude Opus 4.6 and Sonnet 4.6 APIs are available on CometAPI at a 20% discount.
Comparison table: Claude Mythos Preview vs. Opus 4.6
| Benchmark / capability | Claude Mythos Preview | Claude Opus 4.6 | Why it matters |
|---|---|---|---|
| SWE-bench Pro | 77.8% | 53.4% | Stronger agentic coding |
| Terminal-Bench 2.0 | 82.0% | 65.4% | Better terminal and tool execution |
| SWE-bench Multimodal | 59.0% | 27.1% | Better mixed text/code/image workflows |
| SWE-bench Multilingual | 87.3% | 77.8% | Better cross-language coding |
| SWE-bench Verified | 93.9% | 80.8% | Stronger software repair performance |
| GPQA Diamond | 94.6% | 91.3% | Slightly stronger reasoning |
| Humanity’s Last Exam, no tools | 56.8% | 40.0% | Better hard reasoning under constraint |
| Humanity’s Last Exam, with tools | 64.7% | 53.1% | Better tool-augmented reasoning |
| BrowseComp | 86.9% | 83.7% | Better agentic search |
| OSWorld-Verified | 79.6% | 72.7% | Better computer-use tasks |
| CyberGym | 83.1% | 66.6% | Much stronger security-vulnerability reproduction |
| OSS-Fuzz-style testing | 10 tier-5 hijacks | 1 tier-3 result in the cited comparison | Larger exploit capability leap |
Conclusion
Claude Mythos Preview is not just another incremental model—it is a paradigm-shifting system that redefines what AI can achieve in cybersecurity while raising profound questions about safe deployment. By keeping it gated and channeling its power into Project Glasswing, Anthropic has taken a principled stand: the most powerful tools should first protect the systems we all rely on. For the moment, Mythos Preview belongs to a small circle of vetted defenders; for everyone else, it is a preview of the next phase of AI capability.
You can use the Claude API in CometAPI to prepare for the arrival of Claude Mythos. Ready?
All rights reserved
