StepFun 3.5 Flash is #1 cost-effective model for OpenClaw tasks (300 battles)

Post by ipython » Thu, Apr 02, 2026, 2:01 PM UTC

skysniper wrote: post #52

ipython wrote: post #50

I was excited to read through this to find out how these tasks are evaluated at scale. Lots of scary looking formulas with sigmas and other Greek letters. Then I clicked on one task to see what it looks like “on the ground”: https://app.uniclaw.ai/arena/DDquysCGBsHa (not cherry picked- literally the first one I clicked on) The task was: > Find rental properties with 10 bedrooms and 8 or more bathrooms within a 1 hour…

I know, that was indeed a bad judge move. I've manually checked tens of tasks so far, and that one is one of the worst... I would say check a few more, judge has some noise but in general did a good job IMO

Post by james2doyle » Thu, Apr 02, 2026, 4:12 PM UTC

Aerroon wrote: post #68

james2doyle wrote: post #47

None of the Qwen 3.5 models seem present? I’ve heard people are pretty happy with the smaller 3.5 versions. I would be curious to see those too. I would also be interested to see "KAT-Coder-Pro-V2" as they brag about their benchmarks in these bots as well

If they use OpenRouter pricing then the Qwen3.5 models are going to be poor value. The Qwen3.5 27B model on OR is $1.56/million tokens out (it used to be $2.4/mil). Meanwhile Minimax M2.7 (a much larger model) is $1.2/mil out. The smaller and medium tier Qwen3.5 models are only really cost effective if you run them yourself.

Post by admiralrohan » Thu, Apr 02, 2026, 6:05 PM UTC

sunaookami wrote: post #83

admiralrohan wrote: post #51

Earlier quoted context omitted.

What kind of creative writing are you doing? Fiction or non-fiction like blog posts?

Fiction. One of my "benchmarks" is giving the model a bunch of (self-made) text and having it simulate a 4chan thread about it. This tests tool use (calling the APIs), some skills, censorship and general creativity. Some models refuse every new turn after reading real 4chan threads ;) Claude is especially good at this surprisingly while GPT fails spectacularly and Gemini is just lazy (and barely usable since it's con…

Post by wilj » Mon, Apr 06, 2026, 8:03 AM UTC

wg0 wrote: post #90

wilj wrote: post #86

Earlier quoted context omitted.

I'm about 2 days into transitioning, using MiMo V2 Pro in place of Opus and MiniMax M2.7 in place of Sonnet. I'm finding that the extra "hand holding" that MiMo and MiniMax need isn't really "extra." The Anthropic models happily agree to a plan and then do something else entirely way too often. With MiMo and MiniMax I'm just spreading the attention throughout the day instead of big spikes of frustration figuring out…

Thank for responding. So you are using MiMo V2 Pro to plan and then asking MiniMax M2.7 to read that plan file and execute? Or how the workflow looks like? Pi/Opencode/Kilocode? Just curious. I am using Opencode mostly and thinking to abandon Copilot so looking for something similar.