Anthropic's next AI model Mythos can hack you...

Anthropic accidentally published a Mythos pre-release page, confirmed by Fortune, that sits above Opus and is being shipped with a 52-day cyber-security head start.

Mythos positioning

The leaked page, saved before deletion, describes a model tier above Opus — not a Claude 4.x increment. Anthropic is explicitly telling users to "patch up holes" before release, which the host reads as a signal the model can hack live systems. The same page also names "V1 Mythos" and "V2 Capara" as the same model, with Capara being the new internal name.

Claude rate limits cut without notice

Anthropic quietly reduced Claude plan rate limits last week, then confirmed it. Unlike MiniMax and other Chinese providers, Claude's usage screen does not show requests per 5-hour window — users only see a vague "usage" bar. A researcher quoted on the channel's Chinese-language mirror said their workflow "got destroyed" overnight with no warning.

GLM 5.1 jumps coding tasks

GLM 5.1 dropped Friday night. Z.AI's official pitch markets the $10/month coding plan as "3x the usage of Cloud Pro" — they benchmark only against Claude, not their own prior models. Internal WildClawBench runs are at 36% completion (vs GLM 5.0's 33%), with skipped tests caused by server congestion, not model failure. The host moved GLM 5.1 out of D-tier and is retesting.

Hermes as the open alternative

The Hermes agent repo is at 17k GitHub stars and is positioned as the "better open claw." The team migrated their existing Stark agent memory over with no changes and got a more verbose trace, auto-improvement every ~15 prompts, and explicit context usage and memory-deletion logs. Recommendation: if you can pay $200/mo for Claude Code Max, stay; everyone else should run Hermes and bring their own model.

Watch list

DeepSeek has shipped only papers for ~3 months and is expected to return after academic protocols clear.