Gemini 3.6 Flash refuses query that Google Search AI answers correctly
magallanes2010 · reddit · 2026-09-29
A Reddit user found Gemini 3.6 Flash replying "I am a language model who does not have the capacity to understand what you are asking" to a Spanish-language question about public bidding on their country's platform, while Google Search's AI answered the same question correctly.
The author says it's the first time Gemini responded this way, and is trying to switch from ChatGPT (which "randomly lies") but finds the service-level inconsistency frustrating — highlighting possible over-blocking or regional/language inconsistencies in Gemini's serving.
More from Models
- You don't need max: Sonnet-5.5 at max burns so many tokens it costs as much as Opus — adonis_singh · 2026-09-29
- Opus-5.5 at low beats Fable-5.1 at max for ~50x less money — adonis_singh · 2026-09-29
- Opus-5.5 and Sonnet-5.5 jump +52/+50 points on eyebench-v3, but Astra still leads at 1.9x cheaper — adonis_singh · 2026-09-29
- antirez: stop obsessing over t/s — short thinking phases and fast prefill are what matter — antirez · 2026-09-29
- IQuest-Q1: a free open-weight 320B model that diagnoses training issues and fixes its own code — LearnWithBishal · 2026-09-29
- Open-Source IQuest-Q1 Claimed to Self-Diagnose Training Issues and Fix Code Bugs — LearnWithBishal · 2026-09-29