Theo Browne compares Kimi K3's reported benchmarks with his own day of coding work. He discusses model scale, native vision and hosting requirements, while distinguishing planned open weights from the hosted access available at the video's publication. Benchmark rankings and cost comparisons are his account, not independent catalogue measurements.
The coding examples include an extended codebase migration, a 3D game and interface redesigns for T3 Code. Browne praises visual iteration and the darker application redesign, but reports context-window limits, browser-opening mistakes and an overly broad process cleanup that the agent acknowledged. These are mixed practical experiences, not a claim that every task completed flawlessly.
Browne also describes a defensive audit of his own project and raises concerns about the security implications of broadly accessible capable models. He notes slow or unnecessary reasoning, awkward completion messages and the need to preserve thinking history. The conclusion is optimistic about visual and long-horizon coding while recognizing a user-experience gap; sponsor material and precise purchase recommendations are omitted.
Watch on YouTube
