Anthropic's Leaked Forbidden-Strings File Names Opus 4.8, Sonnet 4.8 and a New Tier Above Opus Called Capybara

The detail that turns the March 31 @anthropic-ai/claude-code source-map leak from gossip into a roadmap document is mundane: a TypeScript array. Inside a module the leaked code calls “Undercover Mode” sits a list of forbidden strings — names Claude is hard-instructed never to surface in public commits or chat output. The list, as quoted by WaveSpeed AI and decodethefuture.org, names opus-4-7, opus-4-8, sonnet-4-8, animal codenames including Capybara and Tengu, and the umbrella label Mythos. Boris Cherny, who runs Claude Code at Anthropic, has acknowledged the bundle slip as “a plain developer error” — a missing .npmignore rule plus a Bun bundler quirk.

The Opus piece of that list is the part that has moved fastest. Independent watchers spotted a claude-opus-4.8 model ID active in Google Vertex’s backend earlier this month — not a marketing page, just the routing table — which is the same fingerprint that preceded the Opus 4.7 public launch on April 16, 2026. Anthropic has not announced 4.8, but the combination of the Vertex entry, the forbidden-strings list, and the cadence of the 4.x line is what tipped multiple independent trackers into treating Opus 4.8 as a matter of weeks, not months.

The more interesting reveal is what sits above Opus. Capybara, originally treated as just another codename, turns out — per internal docs paraphrased in the leak coverage — to be the label for an entirely new tier: “a new name for a new tier of model: larger and more intelligent than our Opus models.” In other words, Anthropic is preparing to break out of its three-tier Haiku / Sonnet / Opus structure with a fourth, top-end class. Capybara already appears in version 2 internally, with a reported 1-million-token context window and both a “fast” and a regular “thinking” variant. The umbrella name for the tier, used in product copy fragments inside the bundle, is Mythos.

The other codenames in the leak fill out the surrounding stack. KAIROS appears to be a persistent agent for autonomous monitoring; ULTRAPLAN is a long-horizon task system; a separate Claude Jupiter v1 shows up in test harnesses; and Tengu is paired with Capybara in the same forbidden-strings entry, suggesting a sibling experimental line. None of these names is on a public model card, and Anthropic has not commented on any of them by name.

The structural takeaway is sharper than any single model. Anthropic’s public face is three tiers and a steady drumbeat of point releases; the leaked tree shows a four-tier hierarchy, an Opus refresh already wired into a hyperscaler, a Sonnet jump that skips a number, and a meta-system whose entire job is to keep all of this off the wire. That meta-system shipped to npm in plaintext. As of May 24, 2026, neither Sonnet 4.8 nor Opus 4.8 has a public model card, an API ID or pricing — but the names, the ladder and the failure mode are now public record, and the next Anthropic launch event will be read against them.