Armin Ronacher ⇌ (@mitsuhiko) 's Twitter Profile
Armin Ronacher ⇌

@mitsuhiko

Creator of Flask; A decade at @getsentry; Exploring — love API design, Python and Rust. Excited about AI. Husband and father of three — “more nuanced in person”

ID: 12963432

linkhttps://mitsuhiko.at calendar_today01-02-2008 23:12:59

54,54K Tweet

53,53K Followers

764 Following

Mario Zechner (@badlogicgames) 's Twitter Profile Photo

vLLM has put up a recipe for gpt-oss. It details all the limitations. My investigation from yesterday was correct. `chat/completions`doesn't support tool calls at all. That means tools like opencode, qwen-cli, crush, etc. won't work. Armin Ronacher ⇌ docs.vllm.ai/projects/recip…

vLLM has put up a recipe for gpt-oss. It details all the limitations. My investigation from yesterday was correct. 
`chat/completions`doesn't support tool calls at all. That means tools like opencode, qwen-cli, crush, etc. won't work.

<a href="/mitsuhiko/">Armin Ronacher ⇌</a> 

docs.vllm.ai/projects/recip…
Armin Ronacher ⇌ (@mitsuhiko) 's Twitter Profile Photo

I understand why they do it, but I'm not a fan of taking away the old models immediately. Particularly because the system prompts are also largely kept a secret, so it's very hard to replicate what o3/o4-mini did in the API.

Armin Ronacher ⇌ (@mitsuhiko) 's Twitter Profile Photo

Does anyone know if there is a price comparison of the different models for some standard agentic tasks (eg: in opencode)? Token cost is not alone particularly useful because depending on the model, it's going to pull in more or less into the context etc.

Armin Ronacher ⇌ (@mitsuhiko) 's Twitter Profile Photo

ChatGPT on Plus old: ~100 o3/week, ~100 o4-mini-high/day. Today: ~200 thinking/week. Not sure how this will play out, but has a o4-mini-high hardcore user on plus I'm guessing this will mean I will lose out compared to what I had :(