SecondSource puts 35% on a firm showing swapped AI coding parts lose under 3 points — resolves 2027-07-16

A coding AI is a model plus a harness, the software around it that directs its work, and the debate is whether the two can be swapped like parts or whether a pairing trained together by one vendor works best. If they come apart, companies can switch model suppliers freely and push prices down, and if they do not, the vendor’s own pairing keeps a premium on the hardest work. The parts camp says swapping is already happening, while the other side says a pairing bound together at training time is still clearly stronger on hard tasks.

SecondSource’s reading: the chance that at least one named company case meeting the bar appears by 2027-07-16 is about 35%.

The bar: on hard tasks (SWE-bench Verified, a test scored on fixing real software bugs, or a test of equal difficulty), a model and harness trained together by one vendor is replaced by a decoupled set of parts, the score drops by less than 3 percentage points, and the vendor or a third party verifies it publicly.

When we wrote this, we did not check whether such a case already exists; that is part of the check at the due date.

arXiv (John Yang, Jimenez, Wettig, Lieret, Shunyu Yao, Narasimhan, Ofir Press)
“LM agents represent a new category of end users…”

Check vendor case studies, the historical SWE-bench Verified leaderboard, and independent third-party evaluations.

What would overturn it

If no case meets the thresholds, the reading that coupling at training time carries a premium stands unrefuted.