Over the past two weeks, two separate incidents have given the public an unfiltered look at what Anthropic is building next. References to Claude Opus 4.7, Sonnet 4.8, an entirely new model tier called Capybara, and a next-generation family called Mythos all surfaced from source code and internal documents that were never meant to be public.
If you build on the Claude API, this matters. Not because leaked codenames are guaranteed to ship, but because the direction they reveal tells you where the platform is heading and how to build systems that will not break when the next update lands. Here is everything we know, separated cleanly from what is still speculation.
Two leaks, one week apart
Leak one: the npm package (31 March 2026). A faulty update to Claude Code's npm package shipped with source maps that should have been stripped before publication. Roughly 500,000 lines of TypeScript source code across 1,900 files became publicly accessible. Inside: internal version strings, feature flags, testing configurations, and references to unreleased models.
Leak two: the CMS dump (1 to 2 April 2026). A configuration error in Anthropic's content management system exposed nearly 3,000 unpublished documents. These included draft benchmark comparisons, internal architecture notes, and documentation for a model generation codenamed Capiara (likely an early spelling of what became Capybara).
Neither leak was a hack. Both were configuration errors caught and patched within days. But the information was already out, and the AI community has been picking through it since.
What we know about Opus 4.7
The leaked source code contained "Opus 4.7" and "Sonnet 4.8" as forbidden version strings in what appears to be an internal testing mode. The CMS documents included draft benchmark comparisons suggesting Opus 4.7 targets significant improvements over 4.6 on multi-step reasoning and code tasks. Predicted performance ranges from the leaked drafts: SWE-bench Verified in the 85 to 90% range, and SE-bench potentially exceeding 75%.
Architecture details from the leaks suggest Opus 4.7 uses a dense decoder transformer (not a sparse mixture of experts), maintains or exceeds the 1 million token context window introduced in Opus 4.6, and has a larger parameter count than its predecessor. Pricing is expected to sit at or above Opus 4.6's current rates of $5 per million input tokens and $25 per million output tokens.
On 14 April, The Information reported that Anthropic is preparing to launch Opus 4.7 "as soon as this week", citing a person familiar with the plans. As of today, Anthropic has not confirmed or denied this.
Capybara: a fourth tier above Opus
This is the most structurally significant detail from the leaks. Since Claude 3, Anthropic has maintained three tiers: Haiku (fast and cheap), Sonnet (balanced), and Opus (powerful and expensive). The leaked code references a fourth tier called Capybara, positioned above Opus.
If this ships, it changes the model selection calculus for production systems. Today, Opus is the ceiling: you pick it when you need maximum capability and accept the cost. A Capybara tier would mean Opus becomes the mid-to-high option, not the top, and developers would need to decide whether the premium tier justifies its cost for specific workloads.
Whether Capybara ships as a public tier or remains an internal research tool is still unknown. Anthropic regularly tests models internally that never reach production.
Mythos: the next generation
The CMS leak referenced a model generation called Mythos, distinct from the current 4.x family. Leaked benchmark drafts suggest Mythos-class models demonstrate capabilities "significantly above" Opus 4.6 and operate under ASL-3 (AI Safety Level 3) protocols with a controlled rollout: security teams first, then broader developer access, consumer access last.
Mythos appears to be a product generation name using the Capybara tier. As of April 2026, it remains in restricted early access with select partners. Think of it as Anthropic's next major generation, the way Claude 3 was to Claude 2, but with a new tier structure rather than just faster versions of the existing tiers.
44 feature flags and what they suggest
The npm leak exposed 44 internal feature flags. Most are cryptic internal names, but several point to capabilities that have been widely anticipated:
- Karios: a background agent system. This aligns with the Routines feature that shipped on 14 April, suggesting Karios was the internal name for what became the cloud execution layer.
- Undercover Mode: functionality that strips Anthropic branding for employee external projects. A development convenience, not a consumer feature.
- Voice mode: a long-anticipated capability that would bring Claude closer to feature parity with ChatGPT's voice interface.
- Chyros: advanced agent primitives enabling persistent multi-session operations and task delegation. This would be a significant upgrade to the current agent tooling.
- Autodream: a reference suggesting autonomous planning capabilities for agent workflows, where the model designs its own execution plan before running it.
Feature flags are not shipping commitments. Companies maintain flags for experiments that never reach production. But the pattern here suggests Anthropic is investing heavily in autonomous agent capabilities, persistent execution, and voice, all areas where Claude currently trails ChatGPT.
The AI design tool
Alongside the Opus 4.7 reporting, The Information also revealed that Anthropic is building an AI-powered tool for creating websites and presentations. The report was significant enough to move Figma and Wix stock prices on 14 April, as investors weighed the competitive implications.
Details are thin. We do not know if this is a standalone product, a Claude feature, or an enterprise offering. What we do know is that Anthropic has been expanding beyond the API and chat product into developer tools (Claude Code), cloud execution (Routines), and now potentially design tooling. The company is building a full-stack AI platform, not just a model provider.
What is confirmed vs speculation
To keep things honest, here is a clear split:
Confirmed (from source code and documents that were publicly accessible):
- Internal references to Opus 4.7 and Sonnet 4.8 exist in Claude Code source
- A fourth tier called Capybara appeared in internal code
- A next-gen family called Mythos appeared in internal documents
- 44 feature flags were exposed, including agent primitives and voice mode
- The Information reported Opus 4.7 launch "as soon as this week" from a named source familiar with the plans
Speculation (plausible but unconfirmed):
- Exact benchmark numbers for Opus 4.7 (leaked drafts may not reflect final performance)
- Whether Capybara will ship as a public tier or remain internal
- Whether Mythos is months away or a year away
- Pricing for any unreleased model
- The scope and format of the design tool
Leaked roadmaps from misconfigured packages carry inherent unreliability. Codenames may represent internal experiments that never ship. Benchmark drafts may reflect aspirational targets, not measured performance. Treat everything here as directional, not definitive.
What this means if you build on Claude
If you run production systems on the Claude API, the practical takeaways are straightforward:
Do not hardcode model names. Any system that has "claude-opus-4-6" as a string literal is going to need manual updates every time Anthropic ships a new version. Build an abstraction layer. Use a config file or environment variable for the model identifier so you can swap models without touching production code.
Expect the tier structure to change. A fourth tier above Opus means the cost-capability curve shifts. Systems that currently default to Opus for everything may want to drop to Sonnet for simpler tasks and reserve the new top tier for the hardest reasoning work. Design your architecture to route different tasks to different model tiers.
Agent capabilities are expanding fast. Routines shipped yesterday. Persistent multi-session agents, autonomous planning, and voice mode appear to be in the pipeline. If you are building multi-agent systems, the platform is moving toward you. The custom orchestration you build today may become a native feature in six months, which means building for flexibility rather than permanent infrastructure.
The upgrade path from 4.6 to 4.7 should be smooth. Anthropic has maintained backward compatibility across the 4.x series so far. When Opus 4.6 shipped, systems built for 4.5 worked without changes. Expect the same for 4.7, but test thoroughly before switching production workloads.
Watch the deprecation schedule. Anthropic announced that Claude Opus 4 and Sonnet 4 retire on the API on 15 June 2026, with migration to 4.6 recommended. If you are still on older models, that deadline is two months away.
For clients running systems I have built, none of this requires immediate action. The systems use config-driven model selection and are designed to swap models without code changes. When Opus 4.7 lands, we test, validate output quality, and update the config. That is the benefit of building systems properly from the start.
If you are building on Claude and want your architecture to handle model transitions cleanly, the AI Automation Systems service covers exactly this kind of production-grade design. And if you want to understand which Claude model tier fits your specific workload before you build anything, the AI Consulting and Roadmapping service starts with that conversation.