Fable 5 benchmarked a bunch of harnesses for me to test my rust harness I’m building and @theo I might owe you an apology, oh-my-pi is in fact bloatware (at least in my current configs) —- but I had gpt-5.6 blind grade the results. the initial results are impressive for Ocean and terrible for omp, codex, and opencode — I’ll need to test gpt-5.6 and fable 5 next and I’ll need to put the same exact model into Claude code to get true even results
I built Ocean as a harness written in rust that is giving some interesting results benchmarked against Claude code, Codex, Pi, & oh-my-pi
Ocean is punching above its weight so far.
I will be adding a new marker — memory usage
Ocean is a lightweight daemon that runs on under 50mbs of memory
@kyle_mccleary Oh it did that because after installing it printed the ENTIRE CHANGELOG FROM DAY 1 and hosed my terminal.
I'm sorry but this is clearly slop and I will be removing it from my machine immediately.
S-Tier: Fable 5
A- Tier: GPT 5.6sol, GLM 5.2
B- Tier :Opus 4.8, Grok 4.5, gpt-5.5
C- Tier : Deepseek-v4-pro, Sonnet-5, composer 2.5
I haven’t tried Terra or Luna enough to have a take
being in the loop is important
but i keep seeing people misunderstand this - your job isn't to micromanage the agent and get it to execute its steps in a way you think is better than what it'd do on its own
it's to come up with cool things for it to work on
@theo I just like how I can tell fable in plain English which models to use or let fable choose based on cost benefit.. if that’s slop i guess i love slop