The number worth pinning from the State of Open Source AI report is narrow and honest: open weights have reached parity on coding but still trail on reasoning. The narrowness is the useful part. Too many open-vs-closed comparisons average across a benchmark suite and declare a winner, when the operational question is always per-capability. If your workload is code generation and retrieval-augmented tooling, the gap is effectively gone. If it leans on multi-step reasoning, you’re still paying for a frontier API — for now.

The economics underneath are what make this an enterprise decision rather than a hobbyist one. The report puts GPT-4-class inference at $0.40 per million tokens, down from $20 in 36 months — a 50x collapse. At that price the calculus flips from “can we afford to self-host” to “can we afford the per-token meter and the vendor lock-in.” Their examples land the point better than the abstractions do: PwC running a fine-tuned open finance model on its own hardware with no meter running, a Swiss consortium releasing weights, data, and training code from public supercomputers. Ownership, not just access.

The catch the report is honest enough to flag — “a case that hides its weak points is an advertisement” — is that reasoning-heavy and agentic workloads are exactly where closed frontier models still justify their price. That’s the part I watch in production: an open model that matches on coding can still fall apart on a long tool-use chain where errors compound and the reasoning gap shows up as silent drift, not a benchmark delta. The HN discussion is split on how durable the cost collapse really is. So here’s the bet worth arguing about: within a year, does the reasoning gap close the way coding did, or does it turn out to be the moat that actually holds?