WHAT NEW IN X11.06 FRI

X 上的 AI 合集档案 · 23 个合集 · 70,791 条

gpt-6IMAGE

开源 harness unreal-agent++ 省 72% 调用费

作者发布本地开源 harness unreal-agent++,称 1461 次 gpt-6-astra 付费调用比直连便宜 72%(实付 439.56 美元 vs 直连 1551.61 美元)。

staccoverflow.eth
@STACCoverflow2026.09.24收录原推发布于 2026.09.23
开源 harness unreal-agent++ 省 72% 调用费(第 1 张,共 3 张)
开源 harness unreal-agent++ 省 72% 调用费(第 2 张,共 3 张)
开源 harness unreal-agent++ 省 72% 调用费(第 3 张,共 3 张)
staccoverflow.eth @STACCoverflow原推发布于 2026.09.23
在 X 上打开原推

a wise piggy once told me to give the tech to the people for free. A friend on a Spaces was sad they spent $20 on some AI credits and it bought them a prompt or two. I have an answer. It doesn't matter how rich you are, if you use AI this will make it 5x cheaper for frontier models (or you get 5x as much use out of frontier models, dollar for dollar). Ok I realized that they don't actually have a harness at all, so I shoved this in ChatGPT Desktop like I did with the previous OpenZoo. So, this is leagues above n beyond Unreal Agent's offering for fun n profit, or whatever. Good luck. Have fun. Make babies. It's a no-brainer. And I just cut myself from my rake, entirely, in the interest of the greater good. Introducing unreal-agent++: An open-source harness with state-of-the-art cost efficiency We don't charge you a penny. Everything stays local. We had a cloud version of this before where we only charged you a % what we saved you, but data-aware folks raised some good concerns and now it's not cloud - but we learned an awful lot along the way, and kept receipts. On the receipt, 1,461 paid gpt-6-astra calls were 72% cheaper than direct while maintaining performance: $439.56 actually billed, would have cost $1,551.61 direct. This should actually be even greater savings building on top of @unreallabsai - I invite any of you to run the benchmarks, and let me know! unreal-agent++ is a general-purpose harness with frontier efficiency both at coding and non-coding benchmarks unreal-agent++ is built around asynchronous execution. Tools run independently of model turns, so the model can start long-running tasks, continue other work, and use results as they arrive. Efficiency comes from the async execution loop, a single bash tool, and a system prompt. Our token optmizations are local-only, we never colelct or store any of it. Whatever risk you may have taken here with your harness already, is the same if you use these optimizations. There is no server copy to retain. https://github.com/staccDOTsol/unreal-agent unreal-agent++ is fully open-source & open-license https://github.com/staccDOTsol/unreal-agent Powered by @undacappn https://github.com/AnOversizedMooseWithSocks/leCore If you’re interested in optimizing or deploying cost-efficient agents in your company, here ya go. You’re very welcome. Download the latet release here: https://openzoo.fun Dearest staccfam and leos fam: blow this fucking tweet up so @sequoia and @firstround notice us

这条是引用着另一条写的:x.com/unreallabsai/status/2102435462065385775

x.com/staccoverflow/status/2102633896466690190采集于 2026.09.24 10:19