# DeepSeek V4 Flash vs o3

> DeepSeek V4 Flash is the stronger model overall, scoring 53.6 to 47.4 on the Noometry Index.

- Canonical page: https://noometry.com/compare/deepseek-v4-flash-vs-o3
- Last updated: 2026-10-10
- Shared benchmarks: 34

## Summary

- They share 34 benchmarks with published results for both. DeepSeek V4 Flash scores higher in 7 categories and o3 in 1 category; 5 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where DeepSeek V4 Flash leads 53.7 to 32.1.
- The biggest single-benchmark swing is ARC-AGI-2: 61.4% for DeepSeek V4 Flash and 6.5% for o3.
- DeepSeek V4 Flash is cheaper at $0.15 / $0.60 per million input/output tokens, against $2 / $8 for o3.
- DeepSeek V4 Flash accepts more context: 1M tokens versus 200K.
- DeepSeek V4 Flash has downloadable open weights; the other is API-only.

## Snapshot

| | DeepSeek V4 Flash | o3 |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 53.6 | 47.4 |
| Rank | 32 | 53 |
| Context | 1M | 200K |
| Input $/M | $0.15 | $2 |
| Output $/M | $0.60 | $8 |
| Weights | Open | Proprietary |

## Coding

- DeepSeek V4 Flash: 47.6 (#56)
- o3: 47.3 (#57)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| WeirdML | 63% | 52.4% |
| LMArena Coding | 1457 | 1408 |
| ALE-Bench | 1,306 | 933.55 |
| SWE-bench Verified | — | 62.3% |
| FrontierCode | 18.8% | — |
| SWE-bench Verified (bash only) | — | 58.4% |
| Aider Polyglot | — | 81.3% |
| LMArena WebDev | 1581 | — |
| SciCode | 49.9% | — |
| GSO | — | 8.8% |
| CadEval | — | 74% |

## Agentic & Tool Use

- DeepSeek V4 Flash: —
- o3: 34.3 (#42)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| Berkeley Function Calling Leaderboard | — | 63% |
| GDPval | — | 30.8% |
| DeepResearch Bench | — | 45.2% |
| OSWorld | — | 23% |
| LMArena Search | — | 1144 |
| METR Time Horizons | — | 65.4% |

## Reasoning

- DeepSeek V4 Flash: 53.7 (#25)
- o3: 32.1 (#66)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| ARC-AGI-2 | 61.4% | 6.5% |
| SimpleBench | 61.1% | 53.1% |
| Kagi LLM Benchmark | 52.2% | 67.6% |
| ARC-AGI-1 | 89% | 60.8% |
| CritPt | 16.6% | 1.4% |
| Chess Puzzles | 33% | 38% |
| LMArena Hard Prompts | 1443 | 1402 |
| Mystery Game Puzzles | 34% | 29% |
| DTBench | 90.9% | 84.8% |
| LMCA | 41.7% | 39.7% |
| Epoch Capabilities Index | 154.51 | 146.88 |
| NYT Connections (extended) | 89.6% | — |
| EnigmaEval | — | 13.1% |
| ForecastBench | — | 62.5 |

## Math

- DeepSeek V4 Flash: 60.0 (#31)
- o3: 48.9 (#49)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| FrontierMath (Tiers 1-3) | 57.5% | 33.3% |
| OTIS Mock AIME 2024-2025 | 94.4% | 84.4% |
| LMArena Math | 1427 | 1426 |
| FrontierMath Tier 4 | 24.4% | — |
| MathArena Final-Answer Competitions | 76.5% | — |
| ProofBench | 56% | — |
| Omni-MATH | — | 71.4% |
| MATH Level 5 | — | 97.8% |
| FrontierMath (Feb 2025 set) | — | 18.7% |
| FrontierMath Tier 4 (v1) | — | 2.1% |

## Knowledge

- DeepSeek V4 Flash: 55.5 (#43)
- o3: 54.5 (#47)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| GPQA Diamond | 91% | 81.8% |
| SimpleQA Verified | 33.6% | 49.4% |
| LMArena Expert | 1441 | 1403 |
| Humanity's Last Exam | — | 20.3% |
| MMLU-Pro | — | 85.9% |
| Confabulations | — | 14.4% |
| GPQA (HELM) | — | 75.3% |

## Multimodal

- DeepSeek V4 Flash: —
- o3: 40.7 (#35)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| LMArena Vision | — | 1214 |
| GeoBench | — | 74% |
| VPCT | — | 52% |

## Multilingual

- DeepSeek V4 Flash: 53.2 (#71)
- o3: 51.9 (#99)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| LMArena Non-English | 1420 | 1402 |
| LMArena Chinese | 1467 | 1438 |
| LMArena French | 1438 | 1430 |
| LMArena German | 1418 | 1420 |
| LMArena Japanese | 1403 | 1404 |
| LMArena Korean | 1383 | 1371 |
| LMArena Russian | 1428 | 1407 |
| LMArena Spanish | 1437 | 1395 |

## Instruction Following

- DeepSeek V4 Flash: 75.1 (#76)
- o3: 73.0 (#122)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| LMArena Instruction Following | 1421 | 1368 |
| IFEval | — | 86.9% |

## Long Context

- DeepSeek V4 Flash: 43.8 (#81)
- o3: 53.2 (#6)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| LMArena Longer Query | 1435 | 1372 |
| Fiction.LiveBench | — | 88.9% |
| CL-bench | — | 17.8% |

## Writing & Preference

- DeepSeek V4 Flash: 64.0 (#59)
- o3: 63.7 (#62)

| Benchmark | DeepSeek V4 Flash | o3 |
|---|---|---|
| LMArena Text | 1432 | 1410 |
| LMArena Creative Writing | 1403 | 1359 |
| EQ-Bench Creative Writing | 1559 | 1676 |
| LMArena Multi-Turn | 1449 | 1405 |
| Short-Story Creative Writing | — | 83.9% |
| WildBench | — | 86.1% |

## FAQ

### Is DeepSeek V4 Flash better than o3?

DeepSeek V4 Flash is the stronger model overall, scoring 53.6 to 47.4 on the Noometry Index.

### Which is cheaper, DeepSeek V4 Flash or o3?

DeepSeek V4 Flash is cheaper. It lists at $0.15 per million input tokens and $0.60 per million output tokens; o3 lists at $2 and $8.

### Is DeepSeek V4 Flash or o3 better for coding?

They score almost the same on coding (47.6 vs 47.3); test both on your own repository before choosing.

### Which has the bigger context window?

DeepSeek V4 Flash does, with 1M tokens against 200K.

### How many benchmarks do DeepSeek V4 Flash and o3 share?

34 benchmarks have published results for both models. DeepSeek V4 Flash has 41 scored results on Noometry and o3 has 63.