On March 26, 2026, security researchers discovered nearly 3,000 internal Anthropic files in an unsecured CMS cache. Among them: a draft blog post describing the most powerful AI model Anthropic has ever built. The model is called Claude Mythos. Its internal codename is Capybara. And within hours, Anthropic confirmed it was real — calling it a "step change" in AI capabilities.

On April 7, 2026, Anthropic officially announced Claude Mythos Preview alongside Project Glasswing, a cybersecurity initiative involving Apple, Google, and 45+ other organizations. The benchmark numbers are staggering: 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, and 94.6% on GPQA Diamond. This guide covers everything developers need to know about Claude Mythos — architecture, benchmarks, API preparation, and migration strategy.

1What Is Claude Mythos?

Claude Mythos is Anthropic's next-generation flagship model. The name was chosen to "evoke the deep connective tissue that links together knowledge and ideas" — signaling a model designed for synthesizing complex, multi-domain reasoning at an unprecedented level. The community widely refers to it as Mythos 5 or Claude Mythos, positioning it as the generational leap beyond the Claude 4.x series.

The model was first exposed through a Sanity CMS misconfiguration on March 26, 2026. Security researchers Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge independently discovered the exposed data store. Fortune broke the story, and Anthropic confirmed the model's existence the same day, attributing the leak to "human error."

A second leak followed days later: Anthropic accidentally published Claude Code's full source code to NPM instead of only the compiled version, exposing roughly 500,000 lines of code across 1,900 files. This provided additional corroboration that the Capybara model was actively in preparation.

Training for Mythos has been completed. As of April 8, 2026, the model is in early access testing with selected cybersecurity defense organizations via Project Glasswing. Anthropic has emphasized that Mythos is computationally intensive and expensive to run, and is taking a cautious, phased approach to broader availability.

2The Capybara Tier: A New Level in the Claude Family

Before Mythos, Anthropic's model lineup had three tiers: Haiku (fastest, cheapest), Sonnet (balanced), and Opus (most capable). Mythos introduces a fourth tier — Capybara — that sits above Opus. This is the first expansion of the Claude tier structure since its original design.

Leaked internal documents describe Mythos as "larger and more intelligent than our Opus models," with dramatically higher benchmark scores across software coding, academic reasoning, and cybersecurity. A novel feature highlighted in the leak: Mythos can identify and correct its own errors recursively, without intermediate human input.

3Benchmark Results: Coding, Reasoning & Agentic Tasks

The benchmark numbers from Anthropic's official April 7, 2026 announcement paint a clear picture: Mythos Preview doesn't just beat Opus 4.6 — it laps it on several key tests. Here are the headline results (source: Anthropic):

Coding Benchmarks

The SWE-bench Multimodal result is the most striking: 59.0% vs 27.1% — more than double. This benchmark tests AI's ability to understand visual context alongside code, which is increasingly important as AI agents work directly with GUIs and interfaces.

Reasoning & Knowledge Benchmarks

Anthropic notes that Mythos still performs well at low effort on Humanity's Last Exam, which they flag as a possible sign of some memorization. Worth keeping in mind when interpreting those numbers. That said, the 93.9% SWE-bench Verified score sits more than 13 points above any publicly available model as of April 2026.

4Cybersecurity Capabilities & Project Glasswing

This is where Mythos gets unprecedented. During internal testing, Anthropic found that Mythos Preview can identify and exploit zero-day vulnerabilities in every major operating system and every major web browser. The vulnerabilities it finds are often subtle, with the oldest being a now-patched 27-year-old bug in OpenBSD — an OS known primarily for its security.

In one case, Mythos wrote a browser exploit that chained four vulnerabilities together, including a JIT heap spray that escaped both renderer and OS sandboxes. It autonomously obtained local privilege escalation exploits on Linux by exploiting race conditions and KASLR-bypasses. It wrote a remote code execution exploit on FreeBSD's NFS server using a 20-gadget ROP chain split over multiple packets.

In response, Anthropic launched Project Glasswing on April 7, 2026 — a consortium of 45+ organizations including Apple and Google that will use Mythos Preview to analyze critical software, spot high-stakes vulnerabilities, and help patch them. Access is restricted to keep adversaries from using the same capabilities offensively.

For a deeper dive into the cybersecurity implications, see our dedicated guide: Claude Mythos & Project Glasswing: AI Cybersecurity Guide.

5Architecture & Key Innovations

While Anthropic has not published a full technical paper for Mythos, the leaked documents and official announcements reveal several key architectural advances:

• Recursive self-correction: Mythos can identify and correct its own errors recursively without intermediate human input. This is a significant leap for agentic workflows where the model operates autonomously over multiple steps.
• Token efficiency: On BrowseComp, Mythos uses 4.9x fewer tokens than Opus 4.6 for comparable or better results. This suggests architectural improvements in how the model processes and retrieves information.
• Multimodal code understanding: The 2x improvement on SWE-bench Multimodal (59.0% vs 27.1%) indicates substantially better visual-code integration, critical for GUI-based agent tasks.
• Agentic consistency: Leaked documents describe improved consistency in autonomous multi-step task execution — fewer hallucinations and off-track behaviors during long-running agent sessions.

6API Access & Pricing Expectations

As of April 8, 2026, Claude Mythos Preview is not available through the public Claude API. Access is restricted to Project Glasswing partners for cybersecurity defense work. Anthropic has not announced a timeline for broader availability.

Current Claude API pricing for reference (source: Anthropic):

• Opus 4.6: $5 / $25 per million input/output tokens
• Sonnet 4.6: $3 / $15 per million input/output tokens
• Haiku 4.5: $1 / $5 per million input/output tokens

Capybara-tier pricing has not been announced. Given that Anthropic describes Mythos as "computationally intensive and expensive to run," expect a significant premium above Opus pricing. A reasonable estimate based on the tier structure would be $8–$15 / $40–$75 per million tokens, though this is speculative.

7How to Prepare Your Codebase Now

Anthropic's unified API means your existing Claude integration will carry forward when Mythos becomes available. But there are concrete steps you can take now to be ready:

Example: Model Router Pattern

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

type ModelTier = "haiku" | "sonnet" | "opus" | "capybara";

const MODEL_MAP: Record<ModelTier, string> = {
  haiku: "claude-haiku-4-5-20250210",
  sonnet: "claude-sonnet-4-6-20260210",
  opus: "claude-opus-4-6-20260210",
  // Update when Capybara becomes available:
  capybara: "claude-opus-4-6-20260210", // fallback to Opus
};

function selectTier(taskComplexity: number): ModelTier {
  if (taskComplexity >= 0.9) return "capybara";
  if (taskComplexity >= 0.6) return "opus";
  if (taskComplexity >= 0.3) return "sonnet";
  return "haiku";
}

async function routedCompletion(
  prompt: string,
  complexity: number
) {
  const tier = selectTier(complexity);
  const model = MODEL_MAP[tier];

  return client.messages.create({
    model,
    max_tokens: 4096,
    messages: [{ role: "user", content: prompt }],
  });
}

8Migration Strategy: Opus 4.6 → Mythos

When Capybara-tier access opens up, migration from Opus 4.6 should be straightforward if you've followed the preparation steps above. Here's the expected migration path:

1. Update model identifier: Change the model string in your API calls from the Opus model ID to the Capybara model ID. Anthropic's API contract remains the same across tiers.
2. Adjust token budgets: Mythos uses fewer tokens for equivalent tasks (4.9x fewer on BrowseComp). You may be able to reduce max_tokens settings while getting better results.
3. Re-evaluate prompt complexity: Tasks that required elaborate chain-of-thought prompting with Opus may work with simpler prompts on Mythos, thanks to its recursive self-correction.
4. Test agentic loops: If you run multi-step agent workflows, test them with Mythos to see if you can reduce the number of retry/correction steps in your orchestration layer.
5. Monitor costs closely: Run A/B tests comparing Opus vs Capybara on your actual workloads. The higher per-token cost may be offset by fewer tokens needed and fewer retries.

9Limitations & What to Watch

Despite the impressive benchmarks, there are important caveats:

• No public access yet: Mythos Preview is restricted to Project Glasswing partners. There is no public release date. Building your entire strategy around Mythos availability would be premature.
• Potential memorization: Anthropic flagged that Mythos performs well at low effort on Humanity's Last Exam, which could indicate some benchmark memorization. Real-world performance may differ from benchmark scores.
• Cost uncertainty: Capybara-tier pricing is unknown. If it's significantly more expensive than Opus, the cost-per-quality tradeoff may not justify it for all workloads.
• Cybersecurity dual-use risk: The same capabilities that make Mythos excellent for defense also make it dangerous for offense. Anthropic's cautious rollout suggests they're still working through the safety implications.
• Competitive landscape is moving fast: GPT-5.4 leads on computer use (75% OSWorld), Gemini 3.1 Pro leads on reasoning (94.3% GPQA Diamond before Mythos). By the time Mythos is publicly available, competitors may have closed the gap. See our full comparison guide.

10Why Lushbinary for Claude Integration

At Lushbinary, we've been building on the Claude API since the early days of Claude 3. We've shipped production systems using every tier — from Haiku-powered classification pipelines to Opus-driven agentic workflows. When Capybara-tier access opens up, we'll be among the first to integrate it into client projects.

• Multi-model routing architectures that optimize cost across Claude tiers
• Agentic workflow design with Claude Code Agent Teams
• Production AI security hardening (see our AI Agent Security Guide)
• AWS deployment and cost optimization for AI workloads

❓ Frequently Asked Questions

📚 Sources

• Anthropic — Claude Mythos Preview (April 7, 2026)
• Anthropic — Project Glasswing
• Claude API Pricing
• OfficeChai — Claude Mythos Preview Benchmark Analysis

Content was rephrased for compliance with licensing restrictions. Benchmark data sourced from official Anthropic publications as of April 8, 2026. Pricing and availability may change — always verify on Anthropic's website.