Model comparison
DeepSeek V4 Pro vs Kimi K2.5
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 48.1 on the Noometry Index.
Last verified . 37 shared benchmarks.
DeepSeekSummary
- They share 37 benchmarks with published results for both. DeepSeek V4 Pro scores higher in 7 categories and Kimi K2.5 in 2 categories; 6 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where DeepSeek V4 Pro leads 56.5 to 31.2.
- The biggest single-benchmark swing is ARC-AGI-2: 61.3% for DeepSeek V4 Pro and 11.8% for Kimi K2.5.
- Kimi K2.5 is cheaper at $0.45 / $2.25 per million input/output tokens, against $0.66 / $1.98 for DeepSeek V4 Pro.
- DeepSeek V4 Pro accepts more context: 1M tokens versus 262K.
Side by side
| DeepSeek V4 Pro | Kimi K2.5 | |
|---|---|---|
| Provider | DeepSeek | Moonshot AI |
| Noometry Index | 54.3 | 48.1 |
| Released | 2026-04-24 | 2026-01-27 |
| Weights | Open | Open |
| Context window | 1M | 262K |
| Max output | 393K | 262K |
| Input $ / M tokens | $0.66 | $0.45 |
| Output $ / M tokens | $1.98 | $2.25 |
| Results tracked | 48 | 51 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek V4 Pro leads
DeepSeek V4 Pro: 52.4 (#34), Kimi K2.5: 48.8 (#53)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| SWE-bench Verified | 77.6% | 73.8% |
| LMArena WebDev | 1582 | 1437 |
| SciCode | 51% | 49% |
| WeirdML | 66.2% | 45.6% |
| LMArena Coding | 1470 | 1474 |
| ALE-Bench | 1,403 | 821.65 |
| FrontierCode | 28.6% | — |
| SWE-bench Verified (bash only) | — | 70.8% |
| SWE-bench Multilingual | — | 67.3% |
Agentic & Tool Use Kimi K2.5 leads
DeepSeek V4 Pro: 32.8 (#58), Kimi K2.5: 34.2 (#48)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| Vending-Bench 2 | 3,285 | 1,198 |
| Terminal-Bench | — | 43.2% |
| APEX-Agents | 47.3% | — |
| OSWorld | — | 63.3% |
Reasoning DeepSeek V4 Pro leads
DeepSeek V4 Pro: 56.5 (#24), Kimi K2.5: 31.2 (#80)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| ARC-AGI-2 | 61.3% | 11.8% |
| Kagi LLM Benchmark | 53.5% | 78.5% |
| NYT Connections (extended) | 91.3% | 69.9% |
| ARC-AGI-1 | 90.5% | 65.3% |
| CritPt | 18% | 3.1% |
| Chess Puzzles | 47% | 12% |
| LMArena Hard Prompts | 1461 | 1453 |
| Epoch Capabilities Index | 155.31 | 148.03 |
| SimpleBench | — | 46.8% |
| EnigmaEval | — | 3.4% |
| Thematic Generalization | — | 69.4% |
| Mystery Game Puzzles | 43% | — |
| DTBench | 93.9% | — |
| LMCA | 45.5% | — |
| Surface Evolver Bench | 40% | — |
| ForecastBench | 56.1 | — |
Math DeepSeek V4 Pro leads
DeepSeek V4 Pro: 64.8 (#30), Kimi K2.5: 51.8 (#53)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| MathArena Final-Answer Competitions | 76.6% | 62.3% |
| OTIS Mock AIME 2024-2025 | 98.6% | 92.2% |
| LMArena Math | 1455 | 1470 |
| FrontierMath (Tiers 1-3) | 64.6% | — |
| FrontierMath Tier 4 | 26.8% | — |
| ProofBench | 50% | — |
| FrontierMath (Feb 2025 set) | — | 27.9% |
| FrontierMath Tier 4 (v1) | — | 4.2% |
Knowledge DeepSeek V4 Pro leads
DeepSeek V4 Pro: 59.5 (#31), Kimi K2.5: 53.6 (#56)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| GPQA Diamond | 91.7% | 87.6% |
| SimpleQA Verified | 52.9% | 34.3% |
| Vectara Hallucination Rate | 8.6% | 14.2% |
| LMArena Expert | 1464 | 1466 |
| Humanity's Last Exam | — | 24.4% |
Multimodal Not comparable
DeepSeek V4 Pro: —, Kimi K2.5: 41.1 (#39)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| LMArena Vision | — | 1269 |
| LMArena Document | — | 1430 |
Multilingual Too close to call
DeepSeek V4 Pro: 54.4 (#45), Kimi K2.5: 53.9 (#53)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| LMArena Non-English | 1439 | 1433 |
| LMArena Chinese | 1486 | 1495 |
| LMArena French | 1472 | 1454 |
| LMArena German | 1458 | 1441 |
| LMArena Japanese | 1445 | 1421 |
| LMArena Korean | 1447 | 1410 |
| LMArena Russian | 1453 | 1435 |
| LMArena Spanish | 1458 | 1450 |
Instruction Following Too close to call
DeepSeek V4 Pro: 76.1 (#47), Kimi K2.5: 75.3 (#64)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| LMArena Instruction Following | 1448 | 1431 |
Long Context Kimi K2.5 leads
DeepSeek V4 Pro: 45.0 (#51), Kimi K2.5: 52.1 (#7)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| CL-bench Life | 13.5% | 13.2% |
| LMArena Longer Query | 1458 | 1445 |
| Fiction.LiveBench | — | 86.1% |
| CL-bench | — | 19.3% |
Writing & Preference Too close to call
DeepSeek V4 Pro: 65.5 (#46), Kimi K2.5: 65.1 (#53)
| Benchmark | DeepSeek V4 Pro | Kimi K2.5 |
|---|---|---|
| LMArena Text | 1451 | 1445 |
| LMArena Creative Writing | 1446 | 1423 |
| EQ-Bench Creative Writing | 1553 | 1579 |
| LMArena Multi-Turn | 1467 | 1444 |
| EQ-Bench 4 | 1166 | — |
Frequently asked questions
Is DeepSeek V4 Pro better than Kimi K2.5?
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 48.1 on the Noometry Index.
Which is cheaper, DeepSeek V4 Pro or Kimi K2.5?
Kimi K2.5 is cheaper. It lists at $0.45 per million input tokens and $2.25 per million output tokens; DeepSeek V4 Pro lists at $0.66 and $1.98.
Is DeepSeek V4 Pro or Kimi K2.5 better for coding?
DeepSeek V4 Pro scores higher on coding benchmarks: 52.4 versus 48.8 in the Noometry coding category.
Which has the bigger context window?
DeepSeek V4 Pro does, with 1M tokens against 262K.
How many benchmarks do DeepSeek V4 Pro and Kimi K2.5 share?
37 benchmarks have published results for both models. DeepSeek V4 Pro has 48 scored results on Noometry and Kimi K2.5 has 51.