Anthropic's next-generation AI model, codenamed Claude Mythos or "Claude Capybara," was accidentally exposed this week through a content management system misconfiguration, revealing capabilities that could reshape the competitive landscape between Claude and OpenAI's GPT-5.4. The leaked documents describe Mythos as a tier above Anthropic's current flagship Claude Opus 4.6, with dramatically higher performance in software coding, academic reasoning, and cybersecurity tasks. Security researchers discovered the exposure on March 26, 2026, when nearly 3,000 internal unpublished assets became searchable online .

What Is Claude Mythos and How Does It Compare to Current Models?

Claude Mythos represents a significant architectural leap for Anthropic. Unlike the current model hierarchy of Haiku, Sonnet, and Opus, Mythos would sit as an entirely new tier above all three, making it the most capable AI model Anthropic has ever built. Anthropic confirmed the model's existence, stating: "We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. We consider this model a step change and the most capable we've built to date" .

Anthropic

The leaked benchmarks paint a striking picture of Mythos's capabilities. On coding tasks, the model achieves "dramatically higher" scores than Claude Opus 4.6. For academic reasoning, it exceeds Opus's 91.3% performance on the GPQA Diamond benchmark. Most notably, the documents describe Claude Mythos as "currently far ahead of any other AI model in cyber capabilities," a distinction that carries significant implications for cybersecurity professionals and enterprises .

How Does Claude Mythos Stack Up Against GPT-5.4?

The timing of the leak is particularly noteworthy because it arrives just weeks after OpenAI launched GPT-5.4, its latest flagship model. GPT-5.4 ships in three versions: standard, a reasoning-focused "Thinking" variant, and a performance-optimized "Pro" tier. The model boasts a context window of up to 1 million tokens, meaning it can process roughly 100,000 words in a single session. OpenAI claims GPT-5.4 is 33% less likely to make factual errors than its predecessor and achieved an 83% score on the GDPval benchmark, matching or exceeding industry professionals in 83% of tasks across 44 occupations .

However, GPT-5.4 also introduces Computer Use, allowing the AI to move a mouse, click buttons, and complete tasks in Excel or web browsers autonomously. It scored 75% on the OSWorld benchmark for real-world computer use. Claude Sonnet 4.6, by contrast, reached 72.5% on the same benchmark, approaching functional parity with human performance .

Key Capability Differences Between the Models

• Coding Performance: Claude Opus 4.6 leads on standard coding benchmarks with 90.4% on HumanEval and 80.8% on SWE-bench Verified, while GPT-5.4 dominates hard coding problems with 57.7% on SWE-bench Pro compared to Opus's 45.9%
• Reasoning Tasks: Claude Opus 4.6 achieves 87.4% on GPQA Diamond reasoning benchmarks, outperforming GPT-5.4's 83.9%, while GPT-5.4 leads on science knowledge with 88.5% on MMLU-Pro versus Claude's 85.1%
• Real-World Professional Work: GPT-5.4 published a GDPval score of 83%, demonstrating performance parity with human professionals across 44 occupations, while Claude Opus 4.6 has not published equivalent metrics
• Native Capabilities: GPT-5.4 includes native image generation via DALL-E and video generation via Sora, while Claude currently lacks these integrated features
• Context Window: Both models support 1 million token context windows at maximum, though Claude's standard tier offers 200,000 tokens compared to GPT-5.4's 128,000 tokens

For most developers, GPT-5.4 represents the better default choice at a fraction of the API cost. Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens, while GPT-5.4 costs $2.50 and $15 respectively. However, Claude's unique Agent Teams feature, which allows splitting tasks across multiple Claude agents with different specializations, remains a capability GPT-5.4 doesn't match yet .

Why the Cybersecurity Implications Matter Most

The most alarming detail in the leaked documents concerns cybersecurity. The draft warns that Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." Because of these risks, Anthropic is restricting early access to cyber defense organizations only, giving them time to harden their systems before broader release .

This cautious approach reflects a broader tension in AI development. While Anthropic's CEO Dario Amodei has publicly refused to allow Claude to be used for lethal autonomous military operations or mass surveillance, the company still faces the challenge of releasing increasingly powerful models responsibly. The decision to limit Mythos access to cybersecurity defenders first suggests Anthropic is taking the risks seriously.

How to Evaluate AI Models for Your Organization

• Benchmark Alignment: Identify which performance metrics matter most for your use case. If you need strong coding capabilities, Claude Opus 4.6 leads on standard benchmarks. If you need real-world professional task performance, GPT-5.4's 83% GDPval score is the most comprehensive metric available
• Cost Efficiency: Calculate total cost of ownership by comparing API pricing and subscription tiers. GPT-5.4 costs roughly 50% less per token than Claude Opus 4.6, which matters significantly for high-volume applications
• Values Alignment: Consider organizational values. Anthropic has publicly refused military contracts and committed to not monetizing user conversations with ads, while OpenAI signed a Department of Defense deal. This difference drove a wave of new Claude subscriptions among users prioritizing ethical AI partnerships
• Feature Availability: Assess whether integrated capabilities like image generation (DALL-E), video generation (Sora), or specialized agent teams matter for your workflow. Claude's Agent Teams feature remains unique for multi-agent task orchestration

The leak also reveals that Claude Mythos is expensive to run and not yet ready for general release, suggesting Anthropic is still optimizing the model for production deployment. The company's decision to restrict early access to cybersecurity organizations indicates a deliberate, measured approach to releasing increasingly powerful AI systems .

What This Means for the AI Competition in 2026

The accidental exposure of Claude Mythos underscores how rapidly the AI landscape is evolving. OpenAI shipped GPT-5, 5.2, 5.3, and 5.4 within roughly eight months, demonstrating an aggressive iteration schedule. Anthropic's development of Mythos as a tier above its current flagship suggests the company is pursuing a different strategy, focusing on fewer but more powerful models rather than rapid incremental releases .

Brand trust has become as important as raw capability in the Claude versus ChatGPT competition. Anthropic's Super Bowl ads mocking ChatGPT's decision to show ads to users, combined with the company's public refusal of military contracts, sparked a surge in Claude subscriptions. According to credit card data from approximately 28 million U.S. consumers, Anthropic's Claude more than doubled its paid subscriber base in 2026, with record growth between January and February .

The cybersecurity implications of Mythos also signal a shift in how AI labs think about responsible release. Rather than launching powerful models immediately to the public, Anthropic is giving defenders a head start. This approach acknowledges that increasingly capable AI systems pose real risks that require coordination between developers and security professionals.

As the AI race intensifies, the choice between Claude and ChatGPT increasingly comes down to values alignment, specific capability needs, and cost considerations rather than raw benchmark dominance. The leaked Claude Mythos documents suggest that Anthropic is betting on a strategy of fewer, more carefully vetted releases, while OpenAI continues its rapid iteration approach. For organizations evaluating AI partnerships in 2026, both strategies offer distinct advantages depending on your priorities.