vs
5:2
overall capability wins (1 ties across 8 scored). Phi-4 leads overall.
- Reasoning3:1
- Coding2:0
- Pricing3:0
- Speed2:0
- Specs0:1
Verdict
Phi-4 leads coding (widest gap: +15.0 on SciCode); Phi-4-multimodal offers ~8.0x larger context (131K vs 16K); Phi-4 is ~1.6x cheaper on a blended token basis than Phi-4-multimodal.
- Phi-4 leads coding (widest gap: +15.0 on SciCode)
- Phi-4-multimodal offers ~8.0x larger context (131K vs 16K)
- Phi-4 is ~1.6x cheaper on a blended token basis than Phi-4-multimodal
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Phi-4
- $0.11
- Phi-4-multimodal
- $0.16
Phi-4 is about 1.5x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Microsoft
Microsoft
Family
Phi
Phi
License
MIT
MIT
Open weights
Yes
Yes
Release
Dec 12, 2024
Feb 1, 2025
Knowledge cutoff
-
-
API / provider
Azure / self-host
Azure AI Foundry
Modalities
text → text
text, image, audio → text
Specs
Context window
16K
131K
Max output
4K
4K
Parameters
14B
—
Pricing
Input $/1M
$0.07
$0.08
Output $/1M
$0.14
$0.32
Blended $/1M (3∶1)
$0.09
$0.14
Speed
tok/s
150
90
TTFT (s)
0.15
0.35
Reasoning
MMLU-Pro
70.4
48.5
GPQA Diamond
57.5
31.5
Humanity's Last Exam
3.8
5.0
AIME 2025
18.0
—
MATH-500
81.0
69.3
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
14.1
—
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
14.5
IFBench
23.5
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
23.1
13.1
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
45.5
—
SciCode
26.0
11.0
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,256
—
| Metric | Phi-4 | Phi-4-multimodal | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Microsoft | Microsoft | — |
| Family | Phi | Phi | — |
| License | MIT | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Dec 12, 2024 | Feb 1, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Azure / self-host | Azure AI Foundry | — |
| Modalities | text → text | text, image, audio → text | — |
| Specs | |||
| Context window | 16K | 131K | B +115K |
| Max output | 4K | 4K | tie |
| Parameters | 14B | — | — |
| Pricing | |||
| Input $/1M | $0.07 | $0.08 | A +$0.010 |
| Output $/1M | $0.14 | $0.32 | A +$0.18 |
| Blended $/1M (3∶1) | $0.09 | $0.14 | A +$0.05 |
| Speed | |||
| tok/s | 150 | 90 | A +60 |
| TTFT (s) | 0.15 | 0.35 | A +0.20 |
| Reasoning | |||
| MMLU-Pro | 70.4 | 48.5 | A +21.9 pts |
| GPQA Diamond | 57.5 | 31.5 | A +26.0 pts |
| Humanity's Last Exam | 3.8 | 5.0 | B +1.2 pts |
| AIME 2025 | 18.0 | — | — |
| MATH-500 | 81.0 | 69.3 | A +11.7 pts |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 14.1 | — | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | 14.5 | — |
| IFBench | 23.5 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | 23.1 | 13.1 | A +10.0 pts |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 45.5 | — | — |
| SciCode | 26.0 | 11.0 | A +15.0 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,256 | — | — |
Phi-4
- Context
- 16K / 4K out
- Parameters
- 14B
- Price
- $0.07 / $0.14
- Speed
- 150 tok/s · 0.15s TTFT
- Modalities
- text → text
- License
- MIT
Small high-quality dense model from Microsoft - punches above its size on STEM.
Phi-4-multimodal
- Context
- 131K / 4K out
- Parameters
- —
- Price
- $0.08 / $0.32
- Speed
- 90 tok/s · 0.35s TTFT
- Modalities
- text, image, audio → text
- License
- MIT
First multimodal Phi — unified text, vision, and speech inputs for on-device assistants and accessibility apps.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.