While Claude Code was the undisputed favorite last year, with many tutorials and side projects, I can't quite understand what is happening with the project in 2026. Judging by the decreasing number of YouTube videos, other people can't either.

In February–March, Anthropic announced and rolled out several features that made Claude Code much more autonomous (agentic). There is an active transition from "a single agent in the terminal" to a managed task system and coordination of background agents (Ctrl+B) with an ecosystem of hot-reloaded MCP integrations, skills, hooks, and plugins. Through /teleport, you can initialize /remote-control sessions that can be managed from a mobile app. /loop was introduced for periodic prompt/command execution, along with in-session cron scheduling tools, etc.

Of the truly useful additions, only Auto Mode is worth noting.

Auto Mode
https://claude.com/blog/auto-mode
Presented as a "middle ground" between two extremes in Claude Code. Previously, you either had to manually approve every file change and bash command (very secure but annoying) or use the --dangerously-skip-permissions flag. The new Auto Mode allows Claude to decide for itself which actions are safe and execute them automatically without approval.

Before each tool call, a separate classifier (based on Sonnet 4.6) quickly checks the action for danger. Safe actions proceed automatically; risky ones are blocked. If the model persistently insists on blocked actions, a user prompt eventually appears anyway.

Claude Mythos Announcement Discussion
https://news.ycombinator.com/item?id=47679258
Anthropic describes the personality, goals, and limitations of the new model in a system card. It is not being released publicly—allegedly due to a sharp jump in capabilities and security risks. They claim Mythos has found thousands of zero-day vulnerabilities in OSs, browsers, virtual machines, etc. (including very old bugs). Many write that this could significantly change cybersecurity—for better or worse.

https://red.anthropic.com/2026/mythos-preview/
They also announced Project Glasswing, providing Mythos access to a limited circle of companies to fix critical software using the model.

Recently, many people paying for subscriptions have found Claude Code becoming practically unusable due to recent changes in Anthropic's policies and restrictions without clear rules. Even just mentioning OpenClaw in the system prompt causes the request to be rejected with an error. The system has also become worse at handling non-coding tasks.

Most likely, due to the launch of the new model, they had to maximize the squeeze on all compute that was previously distributed just to attract people to the infrastructure.

#claudecode #newllmmodel