Model comparison
Claude Haiku 4.5 vs DeepSeek V4 Pro
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 39.5 on the Noometry Index.
Last verified . 36 shared benchmarks.
DeepSeekSummary
- They share 36 benchmarks with published results for both. Claude Haiku 4.5 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 15.1.
- The biggest single-benchmark swing is NYT Connections (extended): 14.3% for Claude Haiku 4.5 and 91.3% for DeepSeek V4 Pro.
- DeepSeek V4 Pro is cheaper at $0.66 / $1.98 per million input/output tokens, against $1 / $5 for Claude Haiku 4.5.
- DeepSeek V4 Pro accepts more context: 1M tokens versus 200K.
- DeepSeek V4 Pro has downloadable open weights; the other is API-only.
Side by side
| Claude Haiku 4.5 | DeepSeek V4 Pro | |
|---|---|---|
| Provider | Anthropic | DeepSeek |
| Noometry Index | 39.5 | 54.3 |
| Released | 2025-10-15 | 2026-04-24 |
| Weights | Proprietary | Open |
| Context window | 200K | 1M |
| Max output | 64K | 393K |
| Input $ / M tokens | $1 | $0.66 |
| Output $ / M tokens | $5 | $1.98 |
| Results tracked | 53 | 48 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek V4 Pro leads
Claude Haiku 4.5: 44.0 (#78), DeepSeek V4 Pro: 52.4 (#34)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| LMArena WebDev | 1330 | 1582 |
| SciCode | 43.3% | 51% |
| WeirdML | 45.4% | 66.2% |
| LMArena Coding | 1453 | 1470 |
| ALE-Bench | 653.48 | 1,403 |
| SWE-bench Verified | — | 77.6% |
| FrontierCode | — | 28.6% |
| SWE-bench Verified (bash only) | 66.6% | — |
| SWE-bench Multilingual | 64.7% | — |
Agentic & Tool Use Too close to call
Claude Haiku 4.5: 33.6 (#52), DeepSeek V4 Pro: 32.8 (#58)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| Vending-Bench 2 | 458.89 | 3,285 |
| Terminal-Bench | 35.5% | — |
| APEX-Agents | — | 47.3% |
| Berkeley Function Calling Leaderboard | 68.7% | — |
| DeepResearch Bench | 45.5% | — |
| BALROG | 31.2% | — |
| ExploitBench | 13.7% | — |
Reasoning DeepSeek V4 Pro leads
Claude Haiku 4.5: 15.1 (#320), DeepSeek V4 Pro: 56.5 (#24)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| ARC-AGI-2 | 4% | 61.3% |
| NYT Connections (extended) | 14.3% | 91.3% |
| ARC-AGI-1 | 47.7% | 90.5% |
| CritPt | 0% | 18% |
| Chess Puzzles | 8% | 47% |
| LMArena Hard Prompts | 1420 | 1461 |
| DTBench | 73.6% | 93.9% |
| LMCA | 30.9% | 45.5% |
| Epoch Capabilities Index | 142.41 | 155.31 |
| ForecastBench | 61.4 | 56.1 |
| Kagi LLM Benchmark | — | 53.5% |
| Mystery Game Puzzles | — | 43% |
| Surface Evolver Bench | — | 40% |
Math DeepSeek V4 Pro leads
Claude Haiku 4.5: 44.9 (#78), DeepSeek V4 Pro: 64.8 (#30)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 66.7% | 98.6% |
| LMArena Math | 1396 | 1455 |
| FrontierMath (Tiers 1-3) | — | 64.6% |
| FrontierMath Tier 4 | — | 26.8% |
| MathArena Final-Answer Competitions | — | 76.6% |
| ProofBench | — | 50% |
| Omni-MATH | 56.1% | — |
| MATH Level 5 | 96.4% | — |
| FrontierMath (Feb 2025 set) | 5.9% | — |
| FrontierMath Tier 4 (v1) | 2.1% | — |
Knowledge DeepSeek V4 Pro leads
Claude Haiku 4.5: 37.7 (#153), DeepSeek V4 Pro: 59.5 (#31)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| GPQA Diamond | 71.2% | 91.7% |
| SimpleQA Verified | 13.2% | 52.9% |
| Vectara Hallucination Rate | 9.8% | 8.6% |
| LMArena Expert | 1442 | 1464 |
| MMLU-Pro | 77.7% | — |
| GPQA (HELM) | 60.5% | — |
Multimodal Not comparable
Claude Haiku 4.5: 26.8 (#118), DeepSeek V4 Pro: —
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| Blueprint-Bench 2 | 0% | — |
| LMArena Document | 1420 | — |
Multilingual DeepSeek V4 Pro leads
Claude Haiku 4.5: 49.9 (#129), DeepSeek V4 Pro: 54.4 (#45)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Non-English | 1377 | 1439 |
| LMArena Chinese | 1417 | 1486 |
| LMArena French | 1408 | 1472 |
| LMArena German | 1375 | 1458 |
| LMArena Japanese | 1339 | 1445 |
| LMArena Korean | 1347 | 1447 |
| LMArena Russian | 1381 | 1453 |
| LMArena Spanish | 1420 | 1458 |
Instruction Following DeepSeek V4 Pro leads
Claude Haiku 4.5: 71.4 (#149), DeepSeek V4 Pro: 76.1 (#47)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Instruction Following | 1414 | 1448 |
| IFEval | 80.1% | — |
Long Context DeepSeek V4 Pro leads
Claude Haiku 4.5: 43.6 (#92), DeepSeek V4 Pro: 45.0 (#51)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Longer Query | 1427 | 1458 |
| CL-bench Life | — | 13.5% |
Writing & Preference DeepSeek V4 Pro leads
Claude Haiku 4.5: 57.9 (#123), DeepSeek V4 Pro: 65.5 (#46)
| Benchmark | Claude Haiku 4.5 | DeepSeek V4 Pro |
|---|---|---|
| LMArena Text | 1396 | 1451 |
| LMArena Creative Writing | 1372 | 1446 |
| EQ-Bench 4 | 1064 | 1166 |
| LMArena Multi-Turn | 1409 | 1467 |
| EQ-Bench Creative Writing | — | 1553 |
| WildBench | 83.9% | — |
Frequently asked questions
Is Claude Haiku 4.5 better than DeepSeek V4 Pro?
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 39.5 on the Noometry Index.
Which is cheaper, Claude Haiku 4.5 or DeepSeek V4 Pro?
DeepSeek V4 Pro is cheaper. It lists at $0.66 per million input tokens and $1.98 per million output tokens; Claude Haiku 4.5 lists at $1 and $5.
Is Claude Haiku 4.5 or DeepSeek V4 Pro better for coding?
DeepSeek V4 Pro scores higher on coding benchmarks: 52.4 versus 44.0 in the Noometry coding category.
Which has the bigger context window?
DeepSeek V4 Pro does, with 1M tokens against 200K.
How many benchmarks do Claude Haiku 4.5 and DeepSeek V4 Pro share?
36 benchmarks have published results for both models. Claude Haiku 4.5 has 53 scored results on Noometry and DeepSeek V4 Pro has 48.