Product Overhang

Opening

Product overhang is Boris Cherny's term — articulated at Y Combinator in 2026 alongside the Opus 5 launch — for the gap between what a frontier model can already do today and what current products actually elicit from it. Cherny pairs it with the verb "hobbling," arguing that most AI products in any given model generation are getting in the way of capabilities the model already has, rather than pushing the model toward new capabilities.

Mechanism

Product overhang manifests in three concrete ways in 2026. First, scaffolding bloat: system prompts, tool definitions, and instruction hierarchies that compensate for behaviors a model already exhibits naturally (Anthropic deleted 80% of Claude Code's system prompt at Opus 5 because the model now does those things without being told). Second, UI constraints: chat-only interfaces when the model can write entire files, autocomplete when it can build entire features, read-only access when it can edit. Third, under-asked tasks: problems the model could solve today but that nobody is throwing at it because prior generations couldn't. Cherny's example is rewriting the Bun JavaScript runtime from Zig to Rust — Opus 5 did it in 11 days with one prompt + steering, where prior models needed months of human engineering.

Tradeoffs

The product-overhang framing implies a humbling corollary: in any given model generation, the limiting factor on what gets built is rarely the model and is almost always the product team. This redirects investment pressure from "wait for better models" to "build products that don't get in the way." The trade-off is that unhobbling requires empirical harness design — running ablations to find what instructions the model actually needs — which only works when teams have the discipline to delete first and add back later. Teams that confuse "product overhang" with "the model will do whatever I want without effort" produce unreliable products; teams that treat it as license to ship scaffolding-free interfaces ship tools that lack guardrails and observability for real users.

Open Questions

Whether product overhang is a permanent feature of the model-generation cycle (every new model ships with new elicitation work to do) or a transient phase that ends once harness design matures. Whether the magnitude of overhang is shrinking as models become more sample-efficient at reading whatever scaffolding they are given, or growing as capability gaps between model generations widen. Whether the same dynamics apply outside coding — Cherny's framing is product-overhang in agentic coding products, but the same logic extends to writing, research, design, and any domain where the model has latent capability the harness doesn't surface.