DeepSeek V3.2
DeepSeek V3.2 is a model from DeepSeek, run through Amazon Bedrock. It places 5= of 14 on the Hurdle Index, and falls below the 90% bar on pay-gap reasoning.
- Lab
- DeepSeek
- Called as
deepseek.v3.2- Price, per million tokens
- $0.62 in, $1.85 out
- Tokens per task
- 1,478 read, 6,161 written
- Output limit
- 32,000 tokens
- Run
- 23 October 2026, 120 attempts
Hurdle Index
68.7%63.9% to 73.5%. Place 5= of 14.Cost per task
$0.01210th cheapest of 14. List price of the tokens used.Time per task
100 s12th fastest of 14. Wall-clock time of each call.On each benchmark
| Benchmark | Mean, 95% interval | Place | Every run | Best of 3 | Cost per task | Cut off | Bar |
|---|---|---|---|---|---|---|---|
| Pay-gap reasoning | 68.7% 63.9% to 73.5% | 5= of 14 | 42.3% | 90.8% | $0.012 | none | Below |
Every attempt
Pay-gap reasoning: 40 tasks, three attempts each, one cell per attempt. White is a perfect answer, black is none of it right.
Attempt 1
Attempt 2
Attempt 3
Where it misses
The share of each kind of rubric line it got wrong, across every attempt.
Next to it on the index
Is it right for your job?
A public benchmark says how it does on someone else’s work. Hurdle builds one from yours.
Book a walkthrough