Two frontier labs quietly killed the temperature knob, breaking production determinism assumptions
SmartEnoughToFail · reddit · 2026-10-23
A Reddit user noticed that two frontier labs have removed sampling control in recent weeks. Claude Sonnet 5.5 now returns a 400 error for any non-default temperature or topp, while Gemini 4 Argon silently accepts but ignores temperature, topp and topk — so code that appears pinned at temp 0 isn't.
The official rationale is that thinking models handle variability internally. But two years of standard advice — lower the temperature for consistent extraction, classification and structured output — means a huge amount of production glue now sits on a no-op or hard error, discovered only when outputs drift.
The author contrasts this with local Llama/Qwen setups where sampler, seed and topk stay fully under user control and never change semantics without notice. He acknowledges the counterargument: if reasoning models are more consistent at default than older models at temp 0, losing the knob could be an upgrade — but nobody has measured it. He's now running 200 extraction inputs 20 times each, measuring byte-identical output rates, with local Qwen at temp 0 near-perfect, and routing closed models through a gateway to diff outputs on the same eval set.
More from coding & agent
- Playwright's auto-retry hid a UI flash bug — MutationObserver test caught it — SeanOliver · 2026-10-23
- Next.js 16.2 quirk: server action promise resolves before refreshed page data renders — SeanOliver · 2026-10-23
- The worst proactive agent is an email inbox filter that competes for your attention — manosaie · 2026-10-23
- Four Practical Tricks for Taming Exploding Coding Agent Sessions — srchvrs · 2026-10-23
- LarkAgents launches: first customer saves 50 hours a week automating ERP and reordering — ycombinator · 2026-10-23
- One Hidden Line in a Webpage Footer Can Hijack Your AI Agent — anacondainc · 2026-10-23