DeepSeek·released 2025-12-01✓ 2 sources
DeepSeek-V3.2 benchmark scores: 15 benchmarks tracked. 33% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 15 benchmarks·best result 87.81 on OTIS Mock AIME 2024-2025 (reproduced)·5 of 15 independently reproduced·$0.28/$0.4 per M tokens
Consensus: LiteLLM · Cross-check: OpenRouter · See every model’s pricing →
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| OTIS Mock AIME 2024-2025 | 87.81accuracy (%) | reproduced· optimizedT1 | 2025-12-01Epoch AI |
| GPQA diamond | 77.9accuracy (%) | reproduced· optimizedT1 | 2025-12-01Epoch AI |
| DTBench | 76accuracy (%) | unverifiedT1 | 2025-12-01https://conceptualreasoning.ai/dtbench |
| Aider polyglot | 74.2accuracy (%) | unverified· optimizedT1 | 2025-12-01https://aider.chat/docs/leaderboards/ |
| ARC-AGI | 57accuracy (%) | unverified· optimizedT2 | 2025-12-01 |
| Terminal Bench | 39.55accuracy (%) | unverifiedT1 | 2025-12-01https://www.tbench.ai/leaderboard/terminal-bench/2.0 |
| FrontierMath-2025-02-28-Private | 38.77accuracy (%) | reproduced· optimizedT1 | 2025-12-01Epoch AI |
| LMCA | 33.91accuracy (%) | unverifiedT1 | 2025-12-01https://conceptualreasoning.ai/lmca |
| APEX-Agents | 21.3accuracy (%) | unverifiedT2 | 2025-12-01 |
| CL-bench | 12.4accuracy (%) | unverifiedT2 | 2025-12-01 |
| Chess Puzzles | 9.51accuracy (%) | reproducedT1 | 2025-12-01Epoch AI |
| ProofBench | 8accuracy (%) | unverifiedT1 | 2025-12-01https://www.vals.ai/benchmarks/proof_bench |
| CL-bench Life | 7.4accuracy (%) | unverifiedT2 | 2025-12-01 |
| ARC-AGI-2 | 4.03accuracy (%) | unverifiedT2 | 2025-12-01 |
| FrontierMath-Tier-4-2025-07-01-Private | 3.5accuracy (%) | reproducedT1 | 2025-12-01Epoch AI |
Claims drawn from cited facts, not live model generation.
This model was developed by DeepSeek and released in 2025 at a large, frontier-scale 685.36 billion parameters. That large scale implies more capability headroom but also greater training and inference expense, trading off the efficiency and lower cost of smaller models.
4 cited facts
This model is scored on 15 tracked benchmarks. It holds no current top score on any of them. Its scores are independently reproduced, meaning outside evaluation backs the numbers up rather than leaving them as vendor claims.
3 cited facts
On average the model trails the best reported score by about 39.56 points across its benchmarks, a wide gap that reads as well behind the leaders rather than at the frontier — a score gap under a disclosed harness, not a capability verdict. It holds the top score on 0 benchmarks, so there is no category where it currently leads the field.
2 cited facts
6 facts cross-checked across data sources: 4 corroborated, 1 single-source, 1 disputed.
Source facts, citations, and refresh stamp for this record.
Sources: epoch_benchmarks, huggingface_models, litellm_prices, openrouter_models
DeepSeek-V3.2 is an AI model developed by DeepSeek, released 2025-12-01. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
DeepSeek-V3.2 has recorded scores on 15 benchmarks, each shown with its evidence status.
DeepSeek-V3.2 has recorded scores on 15 benchmarks — GPQA diamond, OTIS Mock AIME 2024-2025, DTBench, LMCA, Chess Puzzles, FrontierMath-2025-02-28-Private, and 9 more. The full table above shows each score with its evidence status.
5 of 15 recorded scores (33%) are independently reproduced rather than self-reported by the lab.
DeepSeek-V3.2 has 15 tracked claims: 5 independently reproduced, 10 unverified.
Listed API pricing: $0.28 per million input tokens, $0.4 per million output tokens. See the pricing block for the full breakdown.