Anthropic’s New Claude Model Triggers Leak, Creators Alarmed Over Usage Caps and Swelling

Anthropic’s internal scramble over the Claude Mythos leak has turned into a public lesson in how “too powerful” can become a product‑development roadblock. The company confirmed that draft documents describing the unreleased model—dubbed Claude Mythos—were unintentionally posted online on March 26‑27, 2026, and quickly discovered by journalists and independent researchers (CoreProse). The files, which originated from Anthropic’s own systems rather than a third‑party breach, label Mythos as the firm’s most capable LLM to date and assign it a risk tier the company has never used before, explicitly calling it “too powerful” for broad release (CoreProse). That self‑assessment, not external criticism, is what has alarmed Anthropic’s safety and policy teams, who now face the paradox of having built a system they fear to deploy.

The leak also shone a spotlight on a separate, more immediate pain point for Anthropic’s customers: the rapid depletion of Claude Code usage limits. Lydia Hallie, Anthropic’s product lead, explained that the surge in “peak‑hour caps” and the ballooning of context windows—some sessions now stretching to a million tokens—are the primary culprits behind users hitting their quotas far sooner than expected (The Decoder). Hallie noted that while the company has patched several bugs, none of those fixes caused incorrect billing. Instead, the company rolled out efficiency improvements and in‑product pop‑ups to keep developers aware of their consumption in real time.

For developers wrestling with the new limits, Hallie’s advice is pragmatic: switch to Sonnet 4.6 instead of the more voracious Opus model, which “burns through limits roughly twice as fast” (The Decoder). She also recommends disabling the “Extended Thinking” feature when it isn’t needed, starting fresh sessions rather than extending old ones, and deliberately trimming the context window to avoid the runaway token counts that trigger caps. Users who still notice anomalously high usage are urged to report the issue through Anthropic’s feedback channel, a move that signals the company’s willingness to fine‑tune its metering in response to real‑world stress tests (The Decoder).

The broader implications of the Mythos documents extend well beyond Anthropic’s internal product roadmap. The leaked drafts outline capabilities that could reshape enterprise AI workflows, but they also hint at applications that intersect with cybersecurity, bio‑risk, and even national‑security concerns (CoreProse). For security teams, the prospect of an LLM with “frontier‑scale” reasoning power raises the specter of adversaries weaponizing similar technology. Regulators, meanwhile, are reminded how quickly existing governance frameworks can be outpaced when labs push the envelope of model capability (CoreProse). The leak therefore serves as a cautionary preview: the next generation of LLMs may arrive with a double‑edged sword of unprecedented utility and equally unprecedented risk.

Anthropic’s response—tightening usage caps, adding pop‑ups, and urging developers toward more efficient model variants—reflects a broader industry trend of throttling access to powerful AI while the safety apparatus catches up. As Hallie puts it, the company is “fixing some bugs” and “shipping efficiency improvements,” but the underlying tension remains: how to monetize a model that the creators themselves deem too risky for unrestricted deployment (The Decoder). The Claude Mythos episode underscores that the race to build ever‑larger LLMs is now as much about internal governance and responsible rollout as it is about raw performance.