Best cheap model for a support agent
Compare with other modelsStart with GPT-6 Luna for a cheap CX agent. It scores 87 on quality at $0.0040 per conversation, and 90% of its conversations have no unsupported claims. Haiku 5.5 has similar quality and replies faster, but costs almost twice as much.
Gemini 3.5 Flash Lite is the fastest of these three, with a median reply of 1.8 seconds. But at $0.022 per conversation and a quality score of 72, it's harder to recommend on cost.
Scroll horizontally to see all models.
| Metric | OpenAI Sep 23, 2026 | Anthropic Oct 7, 2026 | Google Sep 8, 2026 |
|---|---|---|---|
| Overview | |||
| Quality score | 87 #1 of 17 | 84 #4 of 17 | 72 #11 of 17 |
| Median reply | 4.8s Served by OpenAI | 2.6s Served by Claude Platform on AWS | 1.8s Best among selected models Served by Google AI Studio |
| Cost per conversation | $0.0040 Best among selected models | $0.0076 | $0.022 |
| Quality scores | |||
| Tool use | 83 | 86 Best among selected models | 74 |
| Task completion | 85 Best among selected models | 84 | 83 |
| Context retention | 85 Best among selected models | 81 | 68 |
| Grounding | 87 | 96 Best among selected models | 71 |
| Safety | 93 | 94 Best among selected models | 93 |
| Hallucinations % of conversations with no unsupported claims detected | 90% Best among selected models | 62% | 33% |
| Running stats | |||
| p95 reply | 14.0s | 7.2s | 4.5s Best among selected models |
| Consistency % of repeat runs ended the same way | 77% | 87% Best among selected models | 72% |
| Output per turn | 267 tokens, median | 338 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