Phi 4 vs Qwen2.5 72B Instruct
Qwen2.5 72B Instruct leads Phi 4 on the Artificial Analysis Intelligence Index, 7.7 to 5.9, and Phi 4 costs 2.9x less per output token ($0.14 vs $0.40 per million). Updated October 22, 2026; 43 benchmarks have a result for both.
Phi 4 or Qwen2.5 72B Instruct: which to pick
Neither model has a coding score yet, so the two cannot be ranked on coding.
Pick Phi 4 for output tokens 2.9x cheaper ($0.14 vs $0.40 per million) and work measured by LiveCodeBench (Artificial Analysis run) or Terminal-Bench Hard (Artificial Analysis run), where it leads.
- Shared results
- 43
- Phi 4 only
- 3
- Qwen2.5 72B Instruct only
- 10
- Like-for-like categories
- 7 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
No call
Neither has a coding score yet.
One side not measured
Intelligence
AA Intelligence Index
Qwen2.5 72B Instruct
Qwen2.5 72B Instruct leads, 7.7 against 5.9.
Same board, both measured
Long documents
Listed context window
Qwen2.5 72B Instruct
Qwen2.5 72B Instruct takes 33K tokens against 16K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Phi 4
Phi 4 $0.0001, Qwen2.5 72B Instruct $0.0006 per turn.
List prices
At a glance
- Coding index
- Phi 4 not benchmarked, Qwen2.5 72B Instruct not benchmarked.
- Widest category
- Agentic work, 19.6 points apart on average over 2 shared benchmarks.
- Widest single gap
- τ²-Bench Telecom: Phi 4 0.0% against Qwen2.5 72B Instruct 34.5%, 34.5 points apart on a 0 to 100 scale.
- Output price
- Phi 4 $0.14 against Qwen2.5 72B Instruct $0.40 per million output tokens, 2.9x apart.
- Context window
- Phi 4 takes 16K tokens, Qwen2.5 72B Instruct 33K.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | Phi 4 | Bars | Qwen2.5 72B Instruct | Ties |
|---|---|---|---|---|
| Coding | 0 | 4 | ||
| Intelligence | 1 | 1 | ||
| Agentic work | 0 | 2 | ||
| Reasoning | 3 | 3 | ||
| Math | 4 | 2 | ||
| Knowledge | 0 | 6 | ||
| General | 0 | 17 | ||
| All shared boards | 8 | 35 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index04
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| LiveCodeBench (Artificial Analysis run) | 23.1% | 27.6% | Qwen2.5 72B Instruct |
| Terminal-Bench Hard (Artificial Analysis run) | 3.8% | 4.5% | Qwen2.5 72B Instruct |
| WeirdML | Not measured | 16.0% | Only Qwen2.5 72B Instruct measured |
| LMArena Text: coding | 1231 | 1292 | Qwen2.5 72B Instruct |
| LMArena Text: software and IT | 1228 | 1292 | Qwen2.5 72B Instruct |
| Neither model has a result on the other 34 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index11
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| AA Intelligence Index | 5.9 | 7.7 | Qwen2.5 72B Instruct |
| Epoch Capabilities Index | 130.5 | 129.0 | Phi 4 |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| τ²-Bench Telecom | 0.0% | 34.5% | Qwen2.5 72B Instruct |
| METR time horizons | Not measured | 35.8% | Only Qwen2.5 72B Instruct measured |
| BALROG | 11.6% | 16.2% | Qwen2.5 72B Instruct |
| TheAgentCompany | Not measured | 5.7% | Only Qwen2.5 72B Instruct measured |
| Berkeley Function Calling Leaderboard (BFCL) | 28.8% | Not measured | Only Phi 4 measured |
| Neither model has a result on the other 36 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index33
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| GPQA Diamond (Artificial Analysis run) | 57.5% | 49.1% | Phi 4 |
| Humanity's Last Exam (Artificial Analysis run) | 3.8% | 3.6% | Phi 4 |
| GPQA Diamond (Epoch AI run) | 56.1% | 49.1% | Phi 4 |
| BIG-Bench Hard | Not measured | 79.8% | Only Qwen2.5 72B Instruct measured |
| LMArena Text: hard prompts | 1220 | 1271 | Qwen2.5 72B Instruct |
| LMArena Text: expert | 1203 | 1245 | Qwen2.5 72B Instruct |
| LMArena Text: hard prompts in English | 1230 | 1279 | Qwen2.5 72B Instruct |
| Neither model has a result on the other 15 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index42
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| AA Math Index | 18.0 | 14.0 | Phi 4 |
| AIME 2025 (Artificial Analysis run) | 18.0% | 14.0% | Phi 4 |
| OTIS Mock AIME | 13.8% | 8.1% | Phi 4 |
| MATH level 5 | 64.9% | 63.2% | Phi 4 |
| LMArena Text: math | 1246 | 1283 | Qwen2.5 72B Instruct |
| LMArena Text: mathematical work | 1254 | 1290 | Qwen2.5 72B Instruct |
| Neither model has a result on the other 7 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index06
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| MMLU Pro (Artificial Analysis run) | 71.4% | 72.0% | Qwen2.5 72B Instruct |
| MMLU | 84.8% | 85.3% | Qwen2.5 72B Instruct |
| AI2 Reasoning Challenge | Not measured | 94.5% | Only Qwen2.5 72B Instruct measured |
| LMArena Text: science | 1206 | 1275 | Qwen2.5 72B Instruct |
| LMArena Text: medicine and healthcare | 1202 | 1258 | Qwen2.5 72B Instruct |
| LMArena Text: legal and government | 1241 | 1288 | Qwen2.5 72B Instruct |
| LMArena Text: business and finance | 1199 | 1253 | Qwen2.5 72B Instruct |
| Neither model has a result on the other 11 boards here | |||
Web designNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| Neither model has a result on 9 boards here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index017
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| IFBench | 23.5% | 36.9% | Qwen2.5 72B Instruct |
| Long context reasoning (AA-LCR) | 0.0% | Not measured | Only Phi 4 measured |
| DTBench | Not measured | 62.9% | Only Qwen2.5 72B Instruct measured |
| LMCA | Not measured | 13.4% | Only Qwen2.5 72B Instruct measured |
| Creative writing (Lech Mazur) | 62.6% | Not measured | Only Phi 4 measured |
| HellaSwag | Not measured | 84.8% | Only Qwen2.5 72B Instruct measured |
| WinoGrande | Not measured | 82.3% | Only Qwen2.5 72B Instruct measured |
| PIQA | Not measured | 82.6% | Only Qwen2.5 72B Instruct measured |
| LMArena Text | 1217 | 1269 | Qwen2.5 72B Instruct |
| LMArena Text: instruction following | 1201 | 1254 | Qwen2.5 72B Instruct |
| LMArena Text: creative writing | 1181 | 1221 | Qwen2.5 72B Instruct |
| LMArena Text: English | 1231 | 1283 | Qwen2.5 72B Instruct |
| LMArena Text: non-English | 1197 | 1252 | Qwen2.5 72B Instruct |
| LMArena Text: Chinese | 1212 | 1273 | Qwen2.5 72B Instruct |
| LMArena Text: German | 1222 | 1234 | Qwen2.5 72B Instruct |
| LMArena Text: Japanese | 1158 | 1180 | Qwen2.5 72B Instruct |
| LMArena Text: Korean | 1151 | 1188 | Qwen2.5 72B Instruct |
| LMArena Text: French | 1224 | 1280 | Qwen2.5 72B Instruct |
| LMArena Text: Spanish | 1234 | 1255 | Qwen2.5 72B Instruct |
| LMArena Text: Russian | 1209 | 1264 | Qwen2.5 72B Instruct |
| LMArena Text: multi-turn | 1206 | 1272 | Qwen2.5 72B Instruct |
| LMArena Text: longer queries | 1217 | 1281 | Qwen2.5 72B Instruct |
| LMArena Text: writing and language | 1198 | 1247 | Qwen2.5 72B Instruct |
| LMArena Text: entertainment and media | 1162 | 1211 | Qwen2.5 72B Instruct |
| Neither model has a result on the other 22 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Phi 4 | Qwen2.5 72B Instruct | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
Phi 4 costs less
Repository review
50K input + 3K output tokens
Phi 4 costs less
Agent loop
200K input + 10K output tokens
Phi 4 costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | Microsoft Phi 4 | Qwen2.5 72B Instruct |
|---|---|---|
| Maker | Microsoft | Qwen |
| Weights | Open weights | Open weights |
| Released | January 10, 2025 | September 19, 2024 |
| Context window | 16K | 33K |
| Input $/Mtok | $0.07 | $0.36 |
| Output $/Mtok | $0.14 | $0.40 |
| Coding index | Not benchmarked | Not benchmarked |
| Independent measurements behind it | 0 | 0 |
| Cost per solved task | Not computable | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Phi 4 better than Qwen2.5 72B Instruct for coding?
Neither model has a coding score yet, so the two cannot be ranked on coding. Pick Phi 4 for output tokens 2.9x cheaper ($0.14 vs $0.40 per million) and work measured by LiveCodeBench (Artificial Analysis run) or Terminal-Bench Hard (Artificial Analysis run), where it leads.
Which is cheaper, Phi 4 or Qwen2.5 72B Instruct?
Phi 4 costs $0.07 per million input tokens and $0.14 per million output tokens; Qwen2.5 72B Instruct costs $0.36 and $0.40. Cost per solved task is known for neither of them.
Which has the bigger context window, Phi 4 or Qwen2.5 72B Instruct?
Qwen2.5 72B Instruct, with 33K tokens against 16K.
How many benchmarks is the Phi 4 vs Qwen2.5 72B Instruct comparison based on?
43 benchmarks have a result for both models as of October 22, 2026. Phi 4's coding index rests on 0 independent benchmarks and Qwen2.5 72B Instruct's on 0. Where only one model has a result, the page says so instead of naming a winner.