phi-4: benchmark scores, price and ranking
phi-4 from Microsoft Research ranks #584 of 743 LLMs on the LLMs Tiger aggregate benchmark score (45.4), based on 6 benchmark results captured up to 2026-10-13.
- Organisation
- Microsoft Research
- Type
- Open-weight / local model
- Open weights
- Yes
- Released
- 2024-12-12 (source: epoch)
- Cheapest API price
- $0.070 in / $0.140 out per million tokens (via deepinfra, observed 2026-10-13T20:47Z)
- Aggregate rank
- #584 of 743 (score 45.4)
Relative to other models, phi-4 is strongest on MMLU-Pro (#40 of 136) and weakest on BALROG (#30 of 34).
Across the 6 benchmarks where at least ten models are scored, phi-4 is in the top 10% on 0 and in the top half on 2.
Its cheapest listed input price, $0.070 per million tokens, is lower than that of 87% of the 427 models with a listed price.
phi-4 benchmark results
| Benchmark | Score | Rank | Source | Captured |
|---|---|---|---|---|
| MMLU-Pro | 70.4% | #40 of 136 | HuggingFace MMLU-Pro leaderboard | 2026-10-13 |
| Epoch ECI | 130 | #87 of 204 | Epoch AI Benchmarking Hub | 2026-10-13 |
| GPQA Diamond | 56.1% | #175 of 275 | Epoch AI Benchmarking Hub | 2026-10-13 |
| OTIS Mock AIME | 13.8% | #136 of 189 | Epoch AI Benchmarking Hub | 2026-10-13 |
| Chess Puzzles | 1.0% | #97 of 122 | Epoch AI Benchmarking Hub | 2026-10-13 |
| BALROG | 11.6% | #30 of 34 | Epoch AI Benchmarking Hub | 2026-10-13 |
About phi-4
phi-4 is an open-weight model that can run on your own hardware from Microsoft Research. It was released on 2024-12-12, according to Epoch AI. The cheapest API listing we track charges $0.070 per million input tokens and $0.140 per million output tokens, through deepinfra. We hold 6 benchmark results for it across 3 areas, and none is in the top tenth of its board.
Where phi-4 stands
On our aggregate score, phi-4 is #584 of 743 ranked models, in the lower half. Its score is 45.4; the scale is centred on 50, so it sits 4.6 points below the typical model. That beats 21% of the ranked field.
phi-4, benchmark by benchmark
MMLU-Pro puts phi-4 at 70.4%, #40 of 136. That beats 71% of the models tested, but 17.7 points short of Darwin-180B-RSI (88.1%). The results either side are Qwen2.5-72B (71.6%) and qwen3-4b-instruct-2507 (69.6%). It covers 57-subject knowledge + reasoning. The board median is 55.3%, so phi-4 is 15.1 points above it. Source: HuggingFace MMLU-Pro leaderboard, captured 2026-10-13.
On Epoch ECI, phi-4 reaches 130. It ranks #87 of 204, ahead of 57% of the field, and 26 points behind Qwen 3.8 Max at 156. The results either side are Grok-2 (Dec 2024) (130) and Gemma 3 27B (130). It covers Epoch Capabilities Index: Epoch's own aggregate of ~40 benchmarks. The board median is 127, so phi-4 is 3 points above it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
phi-4 lands below the middle on GPQA Diamond: 56.1%, #175 of 275, with 63% of the field scoring higher. GPT-6 Astra (max) leads at 95.8%; the closest result above is magistral-small-2506 at 56.1%. It covers expert-level science questions. The board median is 71.7%, so phi-4 is 15.7 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
phi-4 lands below the middle on OTIS Mock AIME: 13.8%, #136 of 189, with 71% of the field scoring higher. Claude Fable 5.1 (max) leads at 100.0%; the closest result above is DeepSeek-V3 at 15.8%. It covers competition maths (AIME 2024-25 style). The board median is 57.8%, so phi-4 is 44.0 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
phi-4 lands below the middle on Chess Puzzles: 1.0%, #97 of 122, with 79% of the field scoring higher. GPT-6 Astra (max) leads at 72.0%; the closest result above is mistral-small-3.2-2506 at 1.0%. It covers chess tactics puzzles. The board median is 12.5%, so phi-4 is 11.5 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
BALROG is a weak spot. phi-4 scores 11.6%, #30 of 34, behind 85% of the models measured. The leader, GPT-6 Astra (max), reaches 68.3%. It covers long-horizon game-playing agents. The board median is 27.6%, so phi-4 is 16.0 points below it. Source: Epoch AI Benchmarking Hub, captured 2026-10-13.
Strengths and weaknesses
Reasoning, maths and science: uneven, ahead of 45% of the field on average across 3 benchmarks. Its best board here is MMLU-Pro (#40 of 136) and its weakest is OTIS Mock AIME (#136 of 189). General knowledge and chat: uneven, ahead of 39% of the field on average across 2 benchmarks. Its best board here is Epoch ECI (#87 of 204) and its weakest is Chess Puzzles (#97 of 122).
Price and value
The median input price across the 427 priced models is $0.500 per million tokens, so phi-4 costs less than the typical listing. Models that score within 1.5 points or better and cost less: Qwen2.5-Coder-7B-Instruct ($0.010, 49.5), DeepSeek V4.1 Flash (high) ($0.010, 48.9), Qwen2.5-Coder-3B-Instruct ($0.010, 46.5). Similar scores cost more on gemini-2.0-flash-lite ($0.075) and Qwen3-32B ($0.080).
Reading these numbers
The 6 scores on this page come from 2 sources: Epoch AI Benchmarking Hub (5), HuggingFace MMLU-Pro leaderboard (1). No official leaderboard supplied any of them, so treat them as third-party measurements. With 6 results the aggregate rank can move a lot when one new score arrives.
Models ranked nearby
- #581 Qwen3-Coder-480B-A35B-Instruct (45.5)
- #582 ollama/codegeex4 (45.5)
- #583 Claude 3.5 Sonnet (Oct 2024) (45.4)
- #585 gpt-4-turbo (45.4)
- #586 GPT-3.5 Turbo (Nov 2023) (45.4)
- #587 internlm2-math-plus-20b (45.4)
Compare it on the live leaderboard → · How the aggregate score works