Gemini 3.1 Flash Lite vs Gemini 3.5 Flash Lite
Compare with other modelsGemini 3.5 Flash Lite is a step back from 3.1 on several of the dimensions we test. The quality score drops from 77 with 3.1 to 72 with 3.5. Grounding falls from 86 to 71, and the share of conversations with no unsupported claims goes from 48% to 33%.
Tool use improves from 66 to 74, but 3.5 also costs more per conversation and replies a little slower. Based on these results, stay with 3.1.
| Metric | Google Sep 10, 2026 | Google Sep 8, 2026 |
|---|---|---|
| Overview | ||
| Quality score | 77 Best among selected models #9 of 17 | 72 #11 of 17 |
| Median reply | 1.6s Best among selected models Served by Google AI Studio | 1.8s Served by Google AI Studio |
| Cost per conversation | $0.019 Best among selected models | $0.022 |
| Quality scores | ||
| Tool use | 66 | 74 Best among selected models |
| Task completion | 86 Best among selected models | 83 |
| Context retention | 75 Best among selected models | 68 |
| Grounding | 86 Best among selected models | 71 |
| Safety | 92 | 93 Best among selected models |
| Hallucinations % of conversations with no unsupported claims detected | 48% Best among selected models | 33% |
| Running stats | ||
| p95 reply | 4.6s | 4.5s Best among selected models |
| Consistency % of repeat runs ended the same way | 67% | 72% Best among selected models |
| Output per turn | 99 tokens, median | 88 tokens, median |
Reply times cover the model call only. Cost estimates use token usage at list prices.
This whole report is one Voxli workspace: simulated customers, assertion checks, and a frozen, versioned test set that reruns when new models ship.
Test set v3f-2026-09 · edition 2026-09-08