Gemini models keep blowing past length limits while Opus 5.5 over-refuses

jessi_cata · x · 2026-11-09

A user shared real-world eval logs of instruction-following failures:

Gemini 3.8 Flash and Gemini 4 Argon repeatedly ignore output length limits given in the prompt, producing essays far past the cutoff — the attached table shows dozens of essaysizelimit failures, some with 10k+ output tokens against a 16k limit.
Claude Opus 5.5 frequently hit refusal at the prisoner stage; the author says its cyber safeguards are over-active.
The table also shows Qwen3.8 Max Prime failing exactly at its 6000 outputtokenlimit.

Net takeaway: models differ sharply at constraint-following — Gemini can't stay brief, Opus blocks too eagerly.

More from Models