DeepSeek-V3.2: benchmarks, price and rank
DeepSeek-V3.2 from DeepSeek is 103rd of 123 rated AI models on Spezia's leaderboard. It costs $0.26 in, $0.38 out per million tokens.
- 103rd of 123 rated models
- Rating 716
- Rated on 10 benchmarks
- Open weights
- Released Sep 29, 2025
- 164,000 tokens of context
Where it is strongest
Not in the top tenth on any benchmark in use.
Where it falls behind
- 18th of 18 on IMO-AnswerBench
- 55th of 56 on APEX-Agents
- 51st of 56 on BrowseComp
Price: $0.26 in, $0.38 out per million tokens.
Its headline results
The benchmarks that count most in the rating, with its place among the models measured the same way, and where each number came from.
- APEX-Agents21.3%
Long tasks written by investment banking analysts, management consultants and corporate lawyers from their own work, done across files and apps in a realistic work setup.
55th of 56. Independent result from Epoch AI.
- BrowseComp40.1%
Questions whose answers are hidden deep on the web and take many searches and pages to track down.
51st of 56. The maker's own number, from DeepSeek, via LLM Stats.
- Humanity's Last Exam19.8%
About 2,500 questions written by experts in more than a hundred fields to be as hard as they could make them, many needing a diagram or image.
54th of 74. The maker's own number, from DeepSeek, via LLM Stats.
- Vending-Bench 2$1,034
The model runs a simulated vending machine business for a year on its own: finding suppliers, ordering stock, setting prices and paying the daily fees.
53rd of 60. Independent result from Epoch AI.
- Toolathlon35.2%
Long tasks across many real apps and their tools, such as email, calendars, spreadsheets, code repositories and online stores.
33rd of 37. The maker's own number, from DeepSeek, via LLM Stats.
- LMArena text1,425
People put a question to two anonymous models and vote for the better answer.
66th of 107. Independent result from LMArena.
- LMArena WebDev1,362
People ask two anonymous models to build the same web page or app and vote for the better one.
79th of 98. Independent result from LMArena.
- IMO-AnswerBench78.3%
Problems from the International Mathematical Olympiad, the hardest contest for school students, where the model has to reach the right final answer.
18th of 18. The maker's own number, from DeepSeek, via LLM Stats.
- SWE-bench Verified67.8%
Real bug reports and feature requests from popular open-source Python projects. The model has to change the code so the project's hidden tests pass.
39th of 56. The maker's own number, from DeepSeek, via LLM Stats.
- GPQA Diamond79.9%
Graduate-level questions in biology, physics and chemistry, written by PhDs so that searching the web does not give the answer away.
The maker's own number, from DeepSeek, via LLM Stats.
4 more results in its full profile.
Every result with its source, what it costs per task, what jumped with its release, side by side with any other model, and answers to your questions about when to use it.
Models ranked near it
How the rating works, and every benchmark it uses: the model leaderboard.