Using GPT-6.1 Sol only as the planner and letting a local 27B write the code cut my API bill by 77%
retained text
Sol 6.1 came out today and the first thing I wanted to know was whether the architect/editor split still pays off when the smart model is already cheap. Short version: on API spend, yes. I built the same three small 3D games (pool, bowling, foosball) three ways: |Game|Sol alone|Sol plans, local Qwen 3.8 27B codes|Qwen 27B alone| |:-|:-|:-|:-| |Pool|$0.39 · 2.9 min|$0.05 · 18.6 min|$0.00 · 43.1 min (attempt)| |Bowling|$0.14 · 1.7 min|$0.06 · 13.7 min|$0.00 · 34.7 min| |Foosball|$0.22 · 2.0 min|$0.06 · 11.1 min|$0.00 · 36.4 min| |**Total**|**$0.75 · 6.6 min**|**$0.17 · 43.4 min**|**$0.00 · 114.2 min**| So 77% less than Sol alone. Local side was a single RTX 3090 (24 GB). Why it works is pretty boring. Most of the tokens in a coding run are the code itself. The planning and the review are a small slice. So if the expensive model only splits the task and reads what comes back, the bulk of the output lands on the model you don't pay per token for. It's the same idea as aider's architect mode, just with the editor side on your own GPU and several of them running in parallel. The catches: * It's slow. 43 minutes for the three games against under 7 with Sol alone. The local GPU sets the pace. * The saving isn't flat. 87% on pool, 57% on bowling. One run per game, so I wouldn't quote 77% as a constant. * This is API spend only. The GPU and the power bill are still mine. * The planner checks that the promised files exist, not that they do the right thing. Review has to catch the rest. The tool I ran it in is Atomic Agent, open source, it has this as a mode called Fusion (disclaimer: I work on it). But the split itself works in anything that lets you set a planner model and an editor model separately. Curious if anyone's found the floor here. What's the smallest local model you've had hold up as the "hands" under a strong planner before the back-and-forth kills the savings?
document fetched 2026-10-01 00:23 · canonical https://reddit.com/r/ChatGPT/comments/1wtqp7l/using_gpt61_sol_only_as_the_planner_and_letting_a