Two separate data leaks in early April 2026 exposed Anthropic's internal development plans, revealing that the company is building Claude Mythos, a new generation of AI models, alongside a fourth pricing tier called Capybara that sits above the current Opus tier. The leaks also showed that Anthropic is simultaneously deploying Claude Mythos through Project Glasswing, a defensive cybersecurity initiative with 11 major partners including AWS, Apple, Cisco, Google, Microsoft, and NVIDIA, to identify thousands of previously unknown software vulnerabilities .

What Are Mythos, Capybara, and the New Claude Roadmap?

The leaked information reveals a carefully structured product strategy that distinguishes between generation names and tier names. Mythos is a generation name, representing the next family of Claude models that will succeed the current Claude 4.x lineup. Within the Mythos generation, Capybara is a new tier name, introducing a fourth capability level above the existing Haiku (fast and cheap), Sonnet (balanced), and Opus (powerful and expensive) tiers .

Simultaneously, Anthropic is developing incremental updates to its current Claude 4.x line, including Claude Opus 4.7 and Claude Sonnet 4.8. This two-track strategy means the company is improving existing models while building an entirely new tier for customers with extreme performance requirements. The approach mirrors what competitors like OpenAI have done with their reasoning models, but Anthropic is formalizing it as a permanent product tier rather than an experimental parallel line .

How Did Anthropic's Data Leaks Expose These Plans?

Two separate incidents in early April 2026 exposed sensitive internal information. The first leak occurred when Claude Code's npm package accidentally shipped with source code containing version strings for unreleased models. Roughly 500,000 lines of code across 1,900 files went live before Anthropic issued takedown requests, but the code had already been mirrored and analyzed by developers across multiple continents .

The second leak came from a misconfiguration in Anthropic's content management system, which made close to 3,000 unpublished documents publicly accessible. These documents included draft blog posts, benchmark comparisons, and architecture descriptions for the Mythos generation. According to Fortune's reporting, this second incident occurred just days after the npm leak. Anthropic's response was swift but minimal, with the company confirming that Mythos exists and early access customers are testing it, but declining to confirm specific timelines, benchmarks, or release plans .

What Do the Leaked Benchmarks Claim About Claude's Performance?

The leaked draft documents contain performance claims that remain unverified and should be treated as internal projections rather than confirmed results. According to these leaked but unconfirmed documents, Claude Opus 4.7 is claimed to outperform the current Opus 4.6 by large margins on multi-step reasoning, complex code understanding, debugging, and planning tasks .

For context, Opus 4.6 already achieves 80.8% on SWE-bench Verified, a widely used software engineering benchmark that measures how well AI models can understand and write code. It also scores 65.4% on SE-bench, a harder software engineering benchmark, and 91.3% on GPQA Diamond for scientific reasoning. These are frontier-class numbers, and improvements would be significant .

The Mythos preview, which appears to correspond to the Capybara tier, reportedly achieved 93.9% on SWE-bench Verified during early testing. Based on this pattern, one analyst estimates that Opus 4.7 could land in the 85 to 90 percent range on the same benchmark, potentially making it the second most capable coding model Anthropic has ever tested, behind only the restricted Capybara tier . However, it is crucial to note that benchmark scores do not always translate directly to real-world performance on actual coding tasks.

Ways to Understand Anthropic's Product Naming and Strategy

• Generation Names: Mythos represents the next generation of Claude models, similar to how Apple uses "iPhone 16" to denote a product family. It succeeds the current Claude 4.x family and introduces new architectural improvements across all tiers within that generation.
• Tier Names: Capybara is a new fourth tier sitting above the existing Haiku, Sonnet, and Opus tiers. This marks the first time Anthropic has expanded beyond three pricing and capability tiers, signaling a strategy to serve customers with extreme performance requirements who are willing to pay premium prices.
• Version Numbers: Opus 4.7 and Sonnet 4.8 are incremental updates within the current Claude 4.x generation, following the pattern of previous releases like Opus 4.5 and Opus 4.6. These represent continuous improvement rather than generational leaps or new tier introductions.

Why Is Project Glasswing Significant for AI Security?

Project Glasswing demonstrates a fundamental shift in how AI companies approach security responsibility. Rather than building AI systems and hoping they are used safely, Anthropic is proactively deploying Claude Mythos to identify vulnerabilities in the software that underpins global infrastructure. The partnership with 11 major technology companies, including AWS, Apple, Cisco, Google, Microsoft, and NVIDIA, suggests broad industry recognition that this approach is necessary and valuable .

According to Anthropic, Claude Mythos has already discovered thousands of previously unknown vulnerabilities in common operating systems and browsers, some of which have existed undetected for up to 20 years . This discovery rate highlights a critical gap in current security practices. Traditional security testing relies on human researchers and automated tools with limited capabilities. AI models can analyze code at scale and identify patterns that humans might miss, potentially preventing breaches before they occur.

The irony is notable, however. Anthropic built its entire brand on safety and careful control, yet the company experienced two significant data leaks within days of each other. The leaks exposed internal development plans, benchmark data, and architectural details that the company clearly intended to keep confidential. This raises important questions about whether companies developing powerful AI systems can adequately protect their own sensitive information .

What Should Developers Know About Claude's Future?

For developers and organizations currently using Claude's API, the leaked roadmap suggests that more capable models are coming, with particular strength in coding and reasoning tasks. The expansion to a fourth tier indicates that Anthropic is betting on a market for extremely powerful, expensive AI models alongside its existing product line. Whether these models will be available through the standard Claude API or through specialized channels remains unclear .

As of April 2026, Anthropic has confirmed that Mythos exists and that early access customers are testing it, but the company has not announced a public launch date or confirmed the specific performance claims found in the leaked documents. The company's official position is to acknowledge the leaks, issue takedown requests, and continue development without confirming specific release dates or benchmark numbers. Developers should expect announcements about Opus 4.7 and the Capybara tier in coming months, but timelines remain uncertain.