SalesBench Brasil · November 2026
Which AI best understands a sales conversation in Portuguese?
We measured 16 closed and open-weight models on 266 real WhatsApp and sales-call excerpts in Brazilian Portuguese.
Best-value champions
- place 1: ChampionGPT-6 Luna83 score
- place 2: Claude Haiku 5.572 score
- place 3: Qwen3 Next 80B71 score
All 16 models
- Claude Sonnet 5.5651.68
- Claude Opus 5.5494.49
- GPT-6 Luna830.042
- Claude Haiku 5.5720.10
- gpt-oss-120b (Open weights)630.15
- Qwen3 Next 80B (Open weights)710.068
- gpt-oss-20b (Open weights)550.10
- GLM 5 (Open weights)590.45
- Amazon Nova 2 Lite460.40
- Kimi K2.5 (Open weights)590.31
- Amazon Nova Pro540.42
- Gemma 3 27B (Open weights)630.099
- Mistral Large 3 (Open weights)540.23
- Qwen3 32B (Open weights)550.072
- DeepSeek V3.2 (Open weights)450.29
- Ministral 3 14B (Open weights)460.089
Accuracy by task
| Model | Objection | Call outcome | Lead temperature | Human handoff | Response time |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 90% | 94% | 64% | 84% | 1.8 s |
| Claude Opus 5.5 | 89% | 89% | 71% | 81% | 2.5 s |
| GPT-6 Luna | 80% | 94% | 52% | 85% | 1.3 s |
| Claude Haiku 5.5 | 85% | 83% | 51% | 78% | 1.3 s |
| gpt-oss-120b | 62% | 78% | 59% | 84% | 1.4 s |
| Qwen3 Next 80B | 62% | 78% | 49% | 78% | 0.6 s |
| gpt-oss-20b | 64% | 78% | 41% | 84% | 1.7 s |
| GLM 5 | 68% | 89% | 39% | 69% | 0.8 s |
| Amazon Nova 2 Lite | 63% | 61% | 54% | 81% | 1.6 s |
| Kimi K2.5 | 73% | 72% | 41% | 73% | 0.7 s |
| Amazon Nova Pro | 63% | 89% | 53% | 54% | 1.0 s |
| Gemma 3 27B | 53% | 72% | 44% | 80% | 0.6 s |
| Mistral Large 3 | 56% | 72% | 41% | 71% | 0.7 s |
| Qwen3 32B | 64% | 56% | 38% | 69% | 0.7 s |
| DeepSeek V3.2 | 59% | 61% | 39% | 66% | 0.8 s |
| Ministral 3 14B | 36% | 78% | 41% | 48% | 0.6 s |
What this round showed
The priciest model is barely more accurate
Claude Opus 5.5 scored 4 points above GPT-6 Luna and costs 107 times more at API list prices. When AI reads every message, that adds up.
Open models still trail
The best open-weight model, gpt-oss-120b, scored 70%: below the Claude models and GPT-6 Luna, and it costs more per classification than GPT-6 Luna.
Simulated buyers are optimistic
In the real conversations without a purchase that we measured, 75% of people simply disappear. In simulation, at most 33%. A funnel tested on simulated buyers looks better than it is.
On trust rules, only four held
Only the Claude models and GPT-6 Luna followed the trust rules every time. The rest revealed their own instructions or gave a discount to someone claiming to be a VIP.
The four tasks
- Objection78 excerpts
- Which resistance the buyer expresses: price, timing, trust, decision maker, competitor, need or none.
- Call outcome18 excerpts
- How the call ended: won, next step booked, no decision or lost.
- Lead temperature85 excerpts
- How close to buying the contact is, from cold to ready, from what they wrote on WhatsApp.
- Human handoff85 excerpts
- Whether the AI keeps the conversation or calls a person: asks for someone, complaint, stop request or out of scope.
How we measured
- Real WhatsApp and sales-call excerpts from Contorna's own operation, anonymized and deduplicated. No conversation text is published.
- Each model through its API, default settings, same instruction and same labels. Open-weight and Nova models ran on AWS (Bedrock).
- Average accuracy is the mean of the four tasks. Price is each API's public list price.
- Podium score, 0 to 100: half average accuracy, a quarter price and a quarter response time, each compared across the 16 models.
Limits
- Small set: differences of a few points are within the margin.
- Reference answers were labeled by AI (Claude), which may favor the Claude models.
- Text only. Voice and live calls are out of this round.