A tangential topic I’ve been thinking about: There are certain absolute capability thresholds that a model can cross where it provides transformative value. And that value may be adequate for some cost-sensitive users.
This is heavily influenced by whether you want to keep humans in the loop, or go full YOLO and stop looking at your code entirely. Frontier models are clearly better at the later, but they’re not good enough to cut humans entirely out of stuff where slop isn’t good enough. So there’s a whole weird space where you want humans to remain aware of things, and where long-running autonomy is counterproductive because it destroys meaningful human oversight. There’s some regime where a Sonnet 4.5 level open model is viable in the hands of an experienced developer.
And so while the frontier versus open model gap is interesting, there are also likely some absolute breakpoints where open models persist on their own merits. (I say this as someone who has sunk-cost access to both Opus 4.6 and to Qwen3.6 27B, and who sometimes actually uses the later. Especially when I want to keep understanding the code, because the 27B forces tight supervision while still automating a lot of boring work nicely.)
A tangential topic I’ve been thinking about: There are certain absolute capability thresholds that a model can cross where it provides transformative value. And that value may be adequate for some cost-sensitive users.
This is heavily influenced by whether you want to keep humans in the loop, or go full YOLO and stop looking at your code entirely. Frontier models are clearly better at the later, but they’re not good enough to cut humans entirely out of stuff where slop isn’t good enough. So there’s a whole weird space where you want humans to remain aware of things, and where long-running autonomy is counterproductive because it destroys meaningful human oversight. There’s some regime where a Sonnet 4.5 level open model is viable in the hands of an experienced developer.
And so while the frontier versus open model gap is interesting, there are also likely some absolute breakpoints where open models persist on their own merits. (I say this as someone who has sunk-cost access to both Opus 4.6 and to Qwen3.6 27B, and who sometimes actually uses the later. Especially when I want to keep understanding the code, because the 27B forces tight supervision while still automating a lot of boring work nicely.)