Anthropic stripped 80 % of its system-prompt instructions and rolled out Opus 5, a new Claude Code model that can run autonomously for days or weeks and scores 30 % on the Arc AGI benchmark. The “press-delete” redesign, which the company calls an “unhobbling” strategy, lets the AI’s raw capabilities surface without the software shackles that normally keep large language models in a chat-only mode.

Why the change matters

Anthropic’s engineers used to pad Claude with a dense set of guardrails—explicit prompts that tell the model what to say, what not to say, and how to stay on task. Those prompts act like a leash, preventing the model from wandering into untested behavior but also capping what it can accomplish. When Opus 5 arrived, the team deleted the entire prompt stack, then re-added instructions one by one, measuring the impact of each. Because the new model is more capable, they dropped most of the original wording and still got reliable results. The system now understands a high-level goal and generates complete files and features.

The technical payoff

  • Benchmark leap – Opus 5 hit a 30 % score on the Arc AGI benchmark, a metric that gauges a model’s ability to solve novel problems. Earlier Claude versions scored far below that mark, indicating a substantial jump in raw reasoning power.
  • Extended autonomy – The “Auto Mode” feature lets the model operate for extended periods, persisting its own context and making progress without constant human nudges. In practice, a developer can hand the model a specification and let it iterate for days, checking in only when a milestone is reached.
  • Hardening against prompt injection – Prompt injection tricks a model into ignoring its safety rules by embedding malicious instructions in user input. Anthropic says Opus 5 resists this attack better than its predecessors, thanks to internal research and specialized classifiers that flag suspicious patterns. They also recommend defining exit criteria and guardrails before launch.

Stakes for developers and enterprises

For startups building AI-powered tools, the change widens the “product overhang”—the gap between what a model can do and what surrounding software permits it to do. By lifting that overhang, Claude Code can turn a simple chat interface into a full-stack development assistant that writes, tests, and refactors code autonomously. Enterprises that adopt the model could cut the time spent on repetitive coding tasks, freeing engineers for higher-level design work.

What “unhobbling” really means

Cherny coined “unhobbling” to describe freeing a capable model from artificial constraints. In many AI products, developers deliberately limit what the model can attempt because the UI or policy framework isn’t ready for the model’s full scope. The result is a “product overhang” where the AI sits on a pedestal of unused potential. Claude Code brings that potential down to the ground, letting the model act as a co-author of software rather than a mere question-answering bot.

Advice for founders and students

  • Spot the overhang – Look for tasks where existing AI tools stumble just beyond their current limits. Those are low-hanging fruits where an “unhobbling” approach could yield outsized returns.
  • Stretch the model – Give the AI assignments that feel marginally beyond its advertised capabilities. The model’s internal reasoning may surprise you, especially when you reduce prompt friction.
  • Define clear guardrails – Before letting the model run unsupervised, lay out exit conditions (e.g., maximum runtime, resource caps) and safety checks. This lets the AI explore while keeping risks bounded.

Students should blend theory with practice. Mastering algorithms is valuable, but pairing that knowledge with product thinking and data-driven experimentation will enable the next wave of AI-first builders.

What to watch next

Takeaway: By deleting most of its system prompts, Anthropic let a smarter model act with far less hand-holding, turning Claude Code into an autonomous coder that can tackle weeks-long projects.