Yes, two ways failing to combine tool calls is a real deficiency for the LLM's user: it costs money and can get you rate limited.
Every turn forces you to pay $ for cached inputs and possibly gets you rate limited on TPM.
At 500k context and $0.26/M-tok cache read price, every time it combines two tool calls instead of taking separate turns would save you $0.13. That could add up.
Maybe the provider sets a rate limit of 2 million TPM. That's just 4 turns in 1 minute before you start getting 429s. Cutting them in half by combining calls can save that.
All else equal, we definitely want models combining tool calls!
It seems like thinking should be the place where it can iteratively draft before creating an output.
I've wondered if text diffusion models could in theory write better, thanks to iteratively rewriting/refining the text.
It seems to me at least that avenues like that are worth pursuing, so I'm happy to see efforts like Mercury 3 and Diffusion Gemma.
For some reason the providers for MiMo 2.6 Pro are all very slow. I'd much rather use something quick. DeepSeek was great for that.
I haven't tried MiMo ultraspeed.
Model | Reasoning | Intelligence Index | Artificial Analysis million output tokens for the intelligence index
-|-|-|-
Step 5 | ? | 44 | 160
GLM 5.3 | Max | 45 | 210
MiMo V2.6 Pro | ? | 46 | 140
Kimi K3 | Max | 44 | 160
Qwen Max 0902 | ? | 45 | 190
DeepSeek 4.1 Flash | Max | 39 | 250
GLM 5.3-flash | Max | 42 | 180
GPT-6 Astra | Low | 46 | 10
It looks reasonable by open model standards. This doesn't capture the fact that DeepSeek and Step 5 have much higher token/s than the rest, other than MiMo Ultraspeed. MiMo V2.6 is either slow but cheap, or fast but expensive.
Just based on these numbers, it looks good.
Astra-low is one of the fastest and cheapest because it doesn't use many tokens, but I've never tried it.
I liked DeepSeek and GLM when I used them.
I tried it a few times and liked the speed, but often found it ended up looping, i.e. repeating the same token sequence (e.g. the same sequence of 5 paragraphs) over and over again until it hit the max output limit. This doesn't end up happening every session, but does every now and then.
My impression of DSv4.1-flash was very positive aside from this. But that was enough for me to stick with GLM-5.3(-flash), which both gave me consistently great results
I was using a vibe coded bare bones harness.
I was wondering if this was normal from DSv4.1-flash, or if its my harnesses fault.
I've had that looping issue with open models too. But never 4.1. I wonder if it's a model + harness combo? But yeah, one loop issue and I'm done with a model forever.
Yeah, that's what I'd been leaning towards.
No mcp, but I'll see if I can reproduce and debug it, since other people don't seem to have that problem as badly as I've experienced it (and the idea of having a nasty bug like that bothers me).
No mcp support.
I'll try copying deepseek harness's basic tool call formats as a starting point.
I was just trying deepseek v4.1 flash on OpenRouter.
After running reasonably well for a while, I return to the window and see that my scrollback is nothing but this repeated over and over again:
```
OK.
Let me write.
Let me go.
OK.
Let me write the script.
Let me go.
```
I'm not sure to what extant this is a model problem, vs some providers being fairly broken.
If I chose a single provider, I could know how to blame and to avoid them. With OpenRouter, I don't know which provider I was on when this happened.
If Q4 takes less than 1.3x as many tokens as bf16 or q8, it could still end up being faster, given how decode tends to be bandwidth bound.
The kv cache was still bf16, so a few ops are the same between quants.
Mercor, Tacit Labs, Handshake AI...
I suspect companies like these play a big part in model improvements, generating high quality benchmark/task-focused data for training.
However, these do require educated, white collar, workers.
I'm curious about this because I've had Fable decompile games and help me understand what's going on inside the game itself and it never complained. I'm not sure what it takes to trip the "safety" guards but digging into game code and data files doesn't seem to be a barrier at all. I've used CC to build some personal game mods a few times now. Once for a game with no modding capability explicitly exposed.
Oddly not much to trip it. I once ran “touch AKAMAI.md” and then had Claude run git status and it fell back to Opus as if I had just broken into the pentagon itself.
> A narrow set of frontier LLM development tasks, such as distributed training infrastructure, ML accelerator design, and kernel development for certain non-standard chips.
My work is mostly on the Nvidia b200, which apparently gets flagged as non-standard.
Opus 5 works, but sometimes I do wonder if it's surreptitiously trying to sabotage the efforts -- possibly deliberately, but more likely by something like Fable's initial launch, which did come with secretly degraded performance when detecting kernel work. Anthropic was open at the time that such a mechanism existed, but disabled it due to backlash. More likely than not, this is just paranoia on my end...
I think I'd rather have the model stop once the obvious solutions failed and ask me. It can suggest more creative ideas, but I don't necessarily want it to try implementing them.
Every turn forces you to pay $ for cached inputs and possibly gets you rate limited on TPM. At 500k context and $0.26/M-tok cache read price, every time it combines two tool calls instead of taking separate turns would save you $0.13. That could add up.
Maybe the provider sets a rate limit of 2 million TPM. That's just 4 turns in 1 minute before you start getting 429s. Cutting them in half by combining calls can save that.
All else equal, we definitely want models combining tool calls!
reply