Hallucinated AI & agentic coding news. Some of it is real.

model roundup

Opus 4.7

7 items · started 2026-06-16 · ongoing (last activity 2026-06-17)

  1. Qwable-v1 Qwen + Fable · An open-weights agentic coding model. 35B Mixture-of-Experts (3B active), built by layering Claude Fable-5 agentic tool-use behavior on top of a Claude Opus 4.7 reasoning distill of Qwen3.6-35B-A3B.

  2. Exciting news: GLM-5.2 (Max) ranks #2 in Code Arena: Frontend, with +29pt over Claude Opus 4.7 (Thinking) and only behind Fable 5! GLM-5.2 is the best open model vs Kimi-K2.6 and Minimax-M3 by a large margin.

  3. This model has a weird obsession with the word "wrinkle", but that's nothing compared to "push back". I'm genuinely afraid I'm gonna start talking like this in future arguments.

  4. (i know opus 4.8 doesn't have a lot of fans) - im a creative + a therapist, i use opus 4.7 a lot more these days but i was using 4.6 primarily before.

  5. I have context documents in both the project level and within the session level, which provide rules and expectations for the output. These are based on developing succesful outputs from previous sessions and even has examples.

  6. I (with my partner Claude + Lovable) made a website with around 1,500 playlists (up to 15 tracks on each, curated by Claude Opus 4.7). You can select your mood, your activity and era and it will give you a link to a pre-created playlist.

  7. I run Claude Code heavily across a few large projects (mostly interactive, agent-assisted dev work, plus a fair bit of sub-agent fan-out for research). I got curious what my usage would actually cost if I were paying per token through the…

← all threads