Ask HN: Which model do you use with Pi coding agent?

Drafting the five analytical paragraphs from only the provided facts, then logging the task.An Ask HN thread titled “Which model do you use with Pi coding agent?” is live on Hacker News under the AI Agents section, at https://news.ycombinator.com/item?id=48991997, with 3 points and 1 comment so far. The poster says they run **Pi coding agent** with **OpenAI GPT-5.6 Sol** and that **GPT-5.6** feels slower than **Claude Code**. They are asking which other models people use with Pi, including hosted open-source options.

The thread is about model choice inside a coding agent, not a new product launch. The comparison is latency: GPT-5.6 Sol behind Pi versus the feel of Claude Code as a separate agent stack. That frames the question as end-to-end agent speed—model, hosting path, and tool loop—not raw chat quality alone. Hosted OSS models sit in the same slot: same agent, different backend.

For builders already on coding agents, the practical issue is whether a preferred model is usable when it is the slow link in the loop. If GPT-5.6 is accurate but laggy under Pi, engineers will try other endpoints rather than drop the agent. The OSS-hosting question points at cost, control, and whether self-hosted or third-party-hosted open models keep the agent responsive enough for daily use.

Competitively, the post puts **OpenAI** (via GPT-5.6 Sol) next to **Claude Code** as the speed baseline, and leaves room for hosted open-source models as alternatives on the same agent. That is multi-vendor agent tooling: one front end (Pi), several model backends, and Claude Code as a rival product people already use for comparison. Engagement is still thin—3 points, 1 comment—so the thread is early signal, not a settled ranking.

Worth watching next is whether commenters name concrete models and hosts that work well with Pi, and whether anyone reports setups that beat GPT-5.6 Sol on latency without a large quality drop. For operators, the takeaway is to treat model pick as a measured tradeoff under your own agent: time interactive turns, keep a Claude Code (or other) baseline, and only switch to hosted OSS if it holds up under the same workflows.