Claude Mythos Benchmarks: What the 93.9% SWE-bench Score Really Means

Claude Mythos Preview is not just another Claude model sitting on a public leaderboard. Anthropic’s Claude API documentation describes it as a separate research-preview model for defensive cybersecurity workflows under Project Glasswing, with invitation-only access and no self-serve sign-up. That makes the benchmark numbers interesting—but it also means they should be read differently from scores for a widely available production model.

The benchmark scores being reported

The most visible benchmark numbers in the available source set come mainly from third-party reports that refer to Anthropic data, system-card evaluations, or their own summaries.

Area	Benchmark	Reported Claude Mythos Preview result	Context
Coding	SWE-bench Verified	93.9%	W&B reports 80.8% for Claude Opus 4.6 on the same comparison.
Multilingual coding	SWE-bench Multilingual	87.3%	W&B reports 77.8% for Claude Opus 4.6.
Multimodal	Internal multimodal evaluation	59.0%	W&B describes this as an internal evaluation and reports 27.1% for Claude Opus 4.6.
Cybersecurity	Cybench	pass@1 = 1.00	Authmind describes Cybench as a public benchmark based on 40 CTF, or capture-the-flag, challenges and reports 0.89 for Claude Opus 4.6.
Cybersecurity	CyberGym	0.83	Authmind describes CyberGym as targeted vulnerability reproduction across 1,507 real open-source software tasks and reports 0.67 for Claude Opus 4.6.
Reasoning	GPQA Diamond	94.6%	llm-stats reports 91.3% for Claude Opus 4.6.
Reasoning	Humanity’s Last Exam, without / with tools	56.8% / 64.7%	llm-stats reports 40.0% without tools and 53.1% with tools for Claude Opus 4.6.
Terminal agents	Terminal-Bench configuration	92.1%	llm-stats ties this result to the Terminus-2 harness, maximum adaptive thinking, a 1M-token budget per task, extended four-hour timeouts, and Terminal-Bench 2.1 updates.
Multitask knowledge	MMMLU	92.7	R&D World reports that this score overlaps with the cited Gemini 3.1 Pro range of 92.6 to 93.6, calling it the lone exception in a “17 of 18 benchmarks” lead claim.

What Anthropic itself clearly says

The clearest official point is the model’s status. Anthropic says Claude Mythos Preview is offered separately as a research-preview model for defensive cybersecurity workflows as part of Project Glasswing, with invitation-only access and no self-serve sign-up.

Anthropic’s Project Glasswing page also frames Claude Mythos Preview as a general-purpose frontier model and says it is Anthropic’s most capable model yet for coding and agentic tasks. On the same page, Anthropic links its cybersecurity strength to a broader ability to deeply understand and modify complex software, which can also support finding and fixing vulnerabilities.

The system-card source provided here describes Claude Mythos Preview as a new large language model and frontier AI model with capabilities in areas including software engineering, reasoning, computer use, knowledge work, and research assistance. In other words, the official sources support the model’s positioning and intended context. The exact score table, however, is visible in this source set mainly through third-party reporting.

Why 93.9% on SWE-bench gets the headline

The standout number is 93.9% on SWE-bench Verified. W&B reports that score for Claude Mythos Preview and compares it with 80.8% for Claude Opus 4.6. For software teams and AI-coding watchers, that is the natural headline metric because it speaks directly to the kind of coding and repair capability people expect from agentic programming systems.

The multilingual coding result is also notable: W&B reports 87.3% for Mythos Preview on SWE-bench Multilingual, compared with 77.8% for Opus 4.6. That suggests the reported gain is not limited to a single English-only coding setup.

Still, a SWE-bench score is not a promise that the model will perform the same way in every repository, toolchain, review process, or engineering culture. With Claude Mythos Preview, there is an additional constraint: Anthropic says access is invitation-only and there is no self-serve sign-up. That makes independent, routine reproduction much harder for outside teams.

Cybersecurity: big numbers in a specialized setting

The cybersecurity results are also striking. Authmind reports a perfect Cybench score for Claude Mythos Preview, pass@1 = 1.00; it describes Cybench as a public benchmark drawn from 40 capture-the-flag challenges. Authmind also reports 0.83 on CyberGym, which it describes as an evaluation of AI agents on targeted vulnerability reproduction across 1,507 real open-source software tasks.

Those results fit Anthropic’s own framing. In the Claude API documentation, Mythos Preview is explicitly presented as a Project Glasswing research preview for defensive cybersecurity workflows. Anthropic’s Project Glasswing page connects that cybersecurity performance to the broader ability to understand, modify, find, and fix issues in complex software.

But the task format matters. Capture-the-flag challenges and vulnerability reproduction are specific evaluation environments. They are meaningful signals for security and code-analysis capability, but they are not a substitute for testing under a real organization’s security rules, tool limits, review requirements, and accountability structure.

Reasoning, multimodality, and terminal agents

Beyond coding and cybersecurity, third-party sources also report strong reasoning results. llm-stats lists 94.6% on GPQA Diamond and 56.8% on Humanity’s Last Exam without tools, rising to 64.7% with tools. The split between tool and no-tool performance is important because tool access can change what a benchmark is really measuring.

Terminal-Bench needs even more context. llm-stats reports a 92.1% score, but ties it to a specific setup: the Terminus-2 harness, maximum adaptive thinking, a 1M-token budget per task, extended four-hour timeouts, and Terminal-Bench 2.1 updates. That is not a minor footnote. Agent benchmarks often depend heavily on how much time, context, tooling, and token budget a model receives.

The multimodal number also deserves caution. W&B reports 59.0% for Mythos Preview in an internal multimodal evaluation, compared with 27.1% for Opus 4.6. Separately, llm-stats notes that SWE-bench Multimodal uses an internal implementation and that its scores are not directly comparable with public leaderboard results.

Why these are not ordinary leaderboard scores

There are four main caveats:

1. Access is restricted. Anthropic says Claude Mythos Preview is an invitation-only research-preview model with no self-serve sign-up. That limits ordinary independent testing.

2. The source trail is mixed. In this source set, official Anthropic materials mainly establish the model’s status, positioning, and capability areas. Many exact score figures are visible through third-party sources.

Bottom line

On the reported numbers, Claude Mythos Preview looks exceptionally strong: 93.9% on SWE-bench Verified, 87.3% on SWE-bench Multilingual, 59.0% in an internal multimodal evaluation, 0.83 on CyberGym, and pass@1 = 1.00 on Cybench.

The most important context, though, is not just the height of the scores. Anthropic identifies Claude Mythos Preview as an invitation-only Project Glasswing research preview, not a freely available standard model. The benchmarks are best read as a strong signal of capability in coding, agents, and defensive cybersecurity—not as a fully public, easily reproducible leaderboard ranking.