A 41x improvement in 16 months. That's the pace Anthropic says it's seeing in AI-assisted software tasks, and it's forcing the company to rethink how product teams actually work.
Cat Wu, Head of Product for Claude Code at Anthropic, published a detailed account of how exponentially improving models have broken a core assumption in product management: that technical constraints stay roughly stable from the start of a project to the end. When the AI you're building on top of gets dramatically better every few months, rigid quarterly roadmaps become liabilities.
The numbers back up the argument. Wu cites METR research showing that Opus 4.6 can now complete software tasks that take humans about 12 hours, up from Sonnet 3.5's baseline of 21-minute-equivalent tasks. She tested this herself by asking successive Claude models to add table tools to Excalidraw, a whiteboard app. Early models failed unpredictably. The latest one nails it in a single attempt.
Afternoon Experiments Replace Quarterly Plans
Wu describes replacing traditional roadmap planning with what she calls "side quests" - self-directed afternoon experiments where a PM takes an idea from concept to working prototype in hours. Several shipped Claude Code features, including its desktop app integration and todo list tools, originated this way.
The key shift: when a product manager can build a functional prototype with Claude Code in an afternoon, the distance between "what if" and "try this" shrinks to almost nothing. Wu says this has flipped her team's workflow from documentation-first to demo-first. Instead of writing a spec and circulating it for feedback, she hands a rough spec to Claude Code, gets a working prototype, and uses that as the starting point for discussion.
Simpler Code Wins When Models Keep Changing
One practical takeaway stands out. Wu argues that simplicity isn't just good engineering taste - it's a survival strategy when model capabilities shift under your feet. Complex workarounds and over-engineered scaffolding become technical debt the moment a better model drops. Her team cut system prompting for Opus 4.6 by 20% compared to earlier versions, not because they simplified on principle, but because the model no longer needed the hand-holding.
This applies directly to anyone building AI-powered products or workflows. Every hack you write to compensate for a model's weaknesses is code you'll probably rip out in six months. The simpler your implementation, the easier it is to swap in new capabilities.
The Org-Wide Ripple Effect
Wu's most interesting observation isn't about product teams at all. She describes data science, finance, marketing, legal, and design teams at Anthropic all independently adopting AI tools for their own work. The result: handoff bottlenecks disappear because everyone moves at roughly the same speed. A data scientist doesn't wait three days for engineering to build an internal tool when they can prototype it themselves.
Two external PMs echo similar patterns. Bihan Jiang from Decagon describes the PM craft shifting toward rapid discovery rather than upfront certainty. Kai Xin Tai from Datadog frames modern PM work as part creative, part academic - studying model strengths through evaluation loops rather than treating AI as a static feature.
The honest question this raises: how much of traditional product management process exists because building things used to be slow? If prototyping drops from weeks to hours, do you still need the same approval chains, spec reviews, and prioritization frameworks? Wu's answer, at least at Anthropic, is no. Ship fast, revisit every feature when a new model drops, and treat your roadmap as a suggestion rather than a contract.
That's an easy philosophy to hold when you're the company making the models. For teams building on top of third-party AI, the advice still applies, but the uncertainty cuts both ways - the model that makes your workaround obsolete might also make your product obsolete.