Model comparison
DeepSeek-V3.2-Exp vs DeepSeek V4 Pro
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 44.3 on the Noometry Index. DeepSeek-V3.2-Exp costs 3.4× less per token, which makes it the better buy when DeepSeek V4 Pro's lead doesn't matter for your workload.
Last verified . 38 shared benchmarks.
DeepSeekSummary
- They share 38 benchmarks with published results for both. DeepSeek-V3.2-Exp scores higher in 1 category and DeepSeek V4 Pro in 8 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where DeepSeek V4 Pro leads 56.5 to 22.1.
- The biggest single-benchmark swing is ARC-AGI-2: 4% for DeepSeek-V3.2-Exp and 61.3% for DeepSeek V4 Pro.
- DeepSeek-V3.2-Exp is cheaper at $0.26 / $0.38 per million input/output tokens, against $0.66 / $1.98 for DeepSeek V4 Pro.
- DeepSeek V4 Pro accepts more context: 1M tokens versus 164K.
Side by side
| DeepSeek-V3.2-Exp | DeepSeek V4 Pro | |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Noometry Index | 44.3 | 54.3 |
| Released | 2025-09-29 | 2026-04-24 |
| Weights | Open | Open |
| Context window | 164K | 1M |
| Max output | 66K | 393K |
| Input $ / M tokens | $0.26 | $0.66 |
| Output $ / M tokens | $0.38 | $1.98 |
| Results tracked | 49 | 48 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 46.5 (#65), DeepSeek V4 Pro: 52.4 (#34)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| LMArena WebDev | 1362 | 1582 |
| SciCode | 38.9% | 51% |
| WeirdML | 39.5% | 66.2% |
| LMArena Coding | 1454 | 1470 |
| SWE-bench Verified | — | 77.6% |
| FrontierCode | — | 28.6% |
| SWE-bench Verified (bash only) | 70% | — |
| Aider Polyglot | 74.2% | — |
| SWE-bench Multilingual | 59% | — |
| ALE-Bench | — | 1,403 |
Agentic & Tool Use Too close to call
DeepSeek-V3.2-Exp: 32.7 (#59), DeepSeek V4 Pro: 32.8 (#58)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| APEX-Agents | 21.3% | 47.3% |
| Vending-Bench 2 | 1,034 | 3,285 |
| Terminal-Bench | 39.6% | — |
| Berkeley Function Calling Leaderboard | 56.7% | — |
| TheAgentCompany | 42.9% | — |
Reasoning DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 22.1 (#208), DeepSeek V4 Pro: 56.5 (#24)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| ARC-AGI-2 | 4% | 61.3% |
| Kagi LLM Benchmark | 52.2% | 53.5% |
| NYT Connections (extended) | 36.7% | 91.3% |
| ARC-AGI-1 | 57% | 90.5% |
| CritPt | 2.9% | 18% |
| Chess Puzzles | 14% | 47% |
| LMArena Hard Prompts | 1434 | 1461 |
| DTBench | 87.7% | 93.9% |
| LMCA | 29.1% | 45.5% |
| Epoch Capabilities Index | 146.27 | 155.31 |
| Thematic Generalization | 65% | — |
| Mystery Game Puzzles | — | 43% |
| Surface Evolver Bench | — | 40% |
| ForecastBench | — | 56.1 |
Math DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 41.7 (#87), DeepSeek V4 Pro: 64.8 (#30)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| MathArena Final-Answer Competitions | 57.7% | 76.6% |
| OTIS Mock AIME 2024-2025 | 87.8% | 98.6% |
| ProofBench | 8% | 50% |
| LMArena Math | 1435 | 1455 |
| FrontierMath (Tiers 1-3) | — | 64.6% |
| FrontierMath Tier 4 | — | 26.8% |
| FrontierMath (Feb 2025 set) | 22.1% | — |
| FrontierMath Tier 4 (v1) | 2.1% | — |
Knowledge DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 51.7 (#66), DeepSeek V4 Pro: 59.5 (#31)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| GPQA Diamond | 83.4% | 91.7% |
| Vectara Hallucination Rate | 5.3% | 8.6% |
| LMArena Expert | 1436 | 1464 |
| SimpleQA Verified | — | 52.9% |
Multilingual DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 52.2 (#90), DeepSeek V4 Pro: 54.4 (#45)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| LMArena Non-English | 1409 | 1439 |
| LMArena Chinese | 1461 | 1486 |
| LMArena French | 1433 | 1472 |
| LMArena German | 1440 | 1458 |
| LMArena Japanese | 1374 | 1445 |
| LMArena Korean | 1371 | 1447 |
| LMArena Russian | 1424 | 1453 |
| LMArena Spanish | 1440 | 1458 |
Instruction Following DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 74.5 (#93), DeepSeek V4 Pro: 76.1 (#47)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| LMArena Instruction Following | 1413 | 1448 |
Long Context DeepSeek-V3.2-Exp leads
DeepSeek-V3.2-Exp: 47.6 (#16), DeepSeek V4 Pro: 45.0 (#51)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| CL-bench Life | 9.5% | 13.5% |
| LMArena Longer Query | 1428 | 1458 |
| Fiction.LiveBench | 83.3% | — |
| CL-bench | 13.2% | — |
Writing & Preference DeepSeek V4 Pro leads
DeepSeek-V3.2-Exp: 62.4 (#77), DeepSeek V4 Pro: 65.5 (#46)
| Benchmark | DeepSeek-V3.2-Exp | DeepSeek V4 Pro |
|---|---|---|
| LMArena Text | 1425 | 1451 |
| LMArena Creative Writing | 1403 | 1446 |
| EQ-Bench Creative Writing | 1515 | 1553 |
| LMArena Multi-Turn | 1427 | 1467 |
| EQ-Bench 4 | — | 1166 |
Frequently asked questions
Is DeepSeek-V3.2-Exp better than DeepSeek V4 Pro?
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 44.3 on the Noometry Index. DeepSeek-V3.2-Exp costs 3.4× less per token, which makes it the better buy when DeepSeek V4 Pro's lead doesn't matter for your workload.
Which is cheaper, DeepSeek-V3.2-Exp or DeepSeek V4 Pro?
DeepSeek-V3.2-Exp is cheaper. It lists at $0.26 per million input tokens and $0.38 per million output tokens; DeepSeek V4 Pro lists at $0.66 and $1.98.
Is DeepSeek-V3.2-Exp or DeepSeek V4 Pro better for coding?
DeepSeek V4 Pro scores higher on coding benchmarks: 52.4 versus 46.5 in the Noometry coding category.
Which has the bigger context window?
DeepSeek V4 Pro does, with 1M tokens against 164K.
How many benchmarks do DeepSeek-V3.2-Exp and DeepSeek V4 Pro share?
38 benchmarks have published results for both models. DeepSeek-V3.2-Exp has 49 scored results on Noometry and DeepSeek V4 Pro has 48.