DeepSeek-V3.2 (Thinking; Novita): benchmark scores, price and ranking
DeepSeek-V3.2 (Thinking; Novita) from DeepSeek ranks #636 of 743 LLMs on the LLMs Tiger aggregate benchmark score (44.5), based on 7 benchmark results captured up to 2026-10-13.
- Organisation
- DeepSeek
- Type
- Open-weight / local model
- Open weights
- Yes
- Cheapest API price
- $0.259 in / $0.800 out per million tokens (via openrouter, observed 2026-10-13T20:47Z)
- Aggregate rank
- #636 of 743 (score 44.5)
Relative to other models, DeepSeek-V3.2 (Thinking; Novita) is strongest on DTBench (#37 of 146) and weakest on APEX-Agents (#16 of 16).
Across the 6 benchmarks where at least ten models are scored, DeepSeek-V3.2 (Thinking; Novita) is in the top 10% on 0 and in the top half on 2.
Its cheapest listed input price, $0.259 per million tokens, is lower than that of 63% of the 427 models with a listed price.
DeepSeek-V3.2 (Thinking; Novita) benchmark results
| Benchmark | Score | Rank | Source | Captured |
|---|---|---|---|---|
| DTBench | 85.6% | #37 of 146 | Epoch AI Benchmarking Hub | 2026-10-13 |
| Terminal-Bench 2.0 | 39.6% | #12 of 28 | Epoch AI Benchmarking Hub | 2026-10-13 |
| LMCA | 28.8% | #57 of 108 | Epoch AI Benchmarking Hub | 2026-10-13 |
| WebDev Arena | 1,325 | #52 of 67 | Epoch AI Benchmarking Hub | 2026-10-13 |
| CL-bench Life | 7.4% | #7 of 9 | Epoch AI Benchmarking Hub | 2026-10-13 |
| CL-bench | 12.4% | #14 of 16 | Epoch AI Benchmarking Hub | 2026-10-13 |
| APEX-Agents | 21.3% | #16 of 16 | Epoch AI Benchmarking Hub | 2026-10-13 |
About DeepSeek-V3.2 (Thinking; Novita)
DeepSeek-V3.2 (Thinking; Novita) is an open-weight model that can run on your own hardware from DeepSeek. The cheapest API listing we track charges $0.259 per million input tokens and $0.800 per million output tokens, through openrouter. We hold 7 benchmark results for it across 3 areas, and none is in the top tenth of its board.
Where DeepSeek-V3.2 (Thinking; Novita) stands
On our aggregate score, DeepSeek-V3.2 (Thinking; Novita) is #636 of 743 ranked models, in the lower half. Its score is 44.5; the scale is centred on 50, so it sits 5.5 points below the typical model. That beats 14% of the ranked field. Among the 33 ranked DeepSeek models it is #30. The next one up is DeepSeek-R1-Distill-Qwen-1.5B at 44.6. The next one down is DeepSeek-V3 (Mar 2025) at 44.1.
DeepSeek-V3.2 (Thinking; Novita), benchmark by benchmark
DTBench puts DeepSeek-V3.2 (Thinking; Novita) at 85.6%, #37 of 146. That beats 75% of the models tested, but 13.3 points short of Claude Opus 5.5 (max) (98.9%). The results either side are DeepSeek v4 Flash (max) (86.4%) and o3 (high) (84.8%). It covers dynamic tool-use benchmark. The board median is 68.8%, so DeepSeek-V3.2 (Thinking; Novita) is 16.8 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
On Terminal-Bench 2.0, DeepSeek-V3.2 (Thinking; Novita) reaches 39.6%. It ranks #12 of 28, ahead of 57% of the field, and 29.9 points behind Gemini 3 Pro Preview at 69.4%. The results either side are MiniMax-M2.5 (42.7%) and claude-opus-4-1-20250805 (38.0%). It covers real terminal tasks (agent + model). The board median is 37.3%, so DeepSeek-V3.2 (Thinking; Novita) is 2.2 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
DeepSeek-V3.2 (Thinking; Novita) lands below the middle on LMCA: 28.8%, #57 of 108, with 52% of the field scoring higher. Claude Opus 5.5 (max) leads at 68.2%; the closest result above is Claude Sonnet 4 (unknown thinking) at 29.0%. It covers language-model capability assessment. The board median is 29.2%, so DeepSeek-V3.2 (Thinking; Novita) is 0.4 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
DeepSeek-V3.2 (Thinking; Novita) lands below the middle on WebDev Arena: 1,325, #52 of 67, with 76% of the field scoring higher. Claude Opus 5.5 (max) leads at 1,820; the closest result above is MiMo-V2-Flash at 1,331. It covers building web apps, crowd-voted (Elo). The board median is 1,433, so DeepSeek-V3.2 (Thinking; Novita) is 109 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
CL-bench is a weak spot. DeepSeek-V3.2 (Thinking; Novita) scores 12.4%, #14 of 16, behind 81% of the models measured. The leader, GPT-5.4 (xhigh), reaches 27.9%. It covers learning new rules from context. The board median is 17.7%, so DeepSeek-V3.2 (Thinking; Novita) is 5.3 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
APEX-Agents is a weak spot. DeepSeek-V3.2 (Thinking; Novita) scores 21.3%, #16 of 16, behind 94% of the models measured. The leader, Claude Sonnet 5.5 (max), reaches 75.5%. It covers long professional agent tasks. The board median is 47.8%, so DeepSeek-V3.2 (Thinking; Novita) is 26.4 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
The full table above also lists the other 1 result, including boards with fewer than ten models, where a rank says little.
Strengths and weaknesses
Software engineering: uneven, ahead of 40% of the field on average across 2 benchmarks. Its best board here is Terminal-Bench 2.0 (#12 of 28) and its weakest is WebDev Arena (#52 of 67). Agents and tool use: uneven, ahead of 37% of the field on average across 2 benchmarks. Its best board here is DTBench (#37 of 146) and its weakest is APEX-Agents (#16 of 16). General knowledge and chat: weak, ahead of 30% of the field on average across 2 benchmarks. Its best board here is LMCA (#57 of 108) and its weakest is CL-bench (#14 of 16).
Price and value
The median input price across the 427 priced models is $0.500 per million tokens, so DeepSeek-V3.2 (Thinking; Novita) costs less than the typical listing. Models that score within 1.5 points or better and cost less: Qwen2.5-Coder-7B-Instruct ($0.010, 49.5), DeepSeek V4.1 Flash (high) ($0.010, 48.9), Qwen2.5-Coder-3B-Instruct ($0.010, 46.5). Similar scores cost more on gemini-2.5-flash ($0.300) and Qwen3-Coder-480B-A35B-Instruct ($0.380).
Reading these numbers
The 7 scores on this page come from 1 source: Epoch AI Benchmarking Hub (7). No official leaderboard supplied any of them, so treat them as third-party measurements. With 7 results the aggregate rank can move a lot when one new score arrives.
Models ranked nearby
- #633 qwen-turbo-2024-11-01 (44.5)
- #634 Mistral-7B-v0.2 (44.5)
- #635 c4ai-command-a-03-2025 (44.5)
- #637 mistral-small-3.1-2503 (44.4)
- #638 Yi-1.5-34B-Chat (44.4)
- #639 granite-4.1-3b (44.4)
More from DeepSeek
- DeepSeek V4 Pro 0813 (max) (#23, 54.3)
- DeepSeek V4 Flash 0731 (max) (#42, 53.1)
- DeepSeek-V3.2-Exp (high) (#67, 52.4)
- deepseek-v4.1-flash-max (#75, 52.2)
- DeepSeek-V2.5-1210 (#101, 51.7)
- DeepSeek v4 Pro (max) (#108, 51.6)
Compare it on the live leaderboard → · How the aggregate score works