Korva is a scroll of the internet’s best ideas.
Why developers don't upgrade models
Not breaking what works
One reason is that the app works and they don't want to break it. If a support bot starts saying something unexpected, that is a headache, and the optimization and guardrails around the current model are already done.
The price curve
Another is price. Prices on current models fall over time, and a new advance brings a jump in intelligence and a jump in the price curve, which then starts falling. Shifting to the newest model, even among open weights, is not necessarily the most price-effective move.
Trust in the outputs
The third is trust in the outputs. Many people have a personal eval, a random test they give every model. If the new model doesn't look good on it, they keep the one they like, such as Kimi K2.6.
Why do prices fall after a new model arrives?
Epoch AI analyzed model pricing in 2025 and found that the price of reaching a fixed level of capability fell somewhere between 9x and 900x per year, depending on the capability threshold. New frontier models launch at a premium, and then cheaper models reach the same level of capability. Better chips and serving software push the price down further.
Why can swapping models break an app?
Prompts are often tuned to the quirks of one model, so a new model can read the same wording differently. A 2023 study by Melanie Sclar and colleagues found that changing only the formatting of a prompt, such as separators and spacing, swung accuracy by tens of points on some tasks. Output formats can also shift, which breaks code that expects a specific structure.
Why do public benchmarks mislead individual users?
Benchmark questions are published, so they can leak into training data, which researchers call contamination. A model can then score well by having seen the answers. A private test that a user wrote for their own needs cannot leak that way.
More from 20VC with Harry Stebbings
