GPT-4-turbo outputs shorter completions when told it's December, N=477 test
gleech · x · 2026-11-11
An experiment by Rob Lynch shows the date in the system prompt affects gpt-4-turbo output length via API.
- Same code-completion prompt: "May" system prompt averaged 4298 characters, "December" averaged 4086
- N = 477 completions per group, t-test p < 2.28e-07 — statistically significant
- Discussion of possible causes: weekend/weekday load differences, compute load-balancing shifting to newer weights making the old model 'lazier', or funny accidental anthropomorphic correlations
- Author suggests a testable follow-up: compare weekend vs weekday outputs
More from Models
- SenseNova 6.8 Flash debuts with long-term memory and multi-format output — SarahAnnabels · 2026-11-11
- Dev claims ~700 tokens/s running Qwen 3 27B on a single RTX 5090 — IgorCarron · 2026-11-11
- Qwen4 Max Preview appears sans benchmarks, rumors call it disappointing — kimmonismus · 2026-11-11
- Inkling 2 Small goes free on OpenCode: 594B MoE, 1M context, multimodal — ZhongRuiqi · 2026-11-11
- Gemini free tier users report usage limit quietly nerfed from ~2% to 6% per prompt — ItsMichaelRay · 2026-11-11
- Researcher lists three explanations for "model nerfing," one intentional — gleech · 2026-11-11