vs
1:1
overall capability wins (1 ties across 3 scored). Price/speed category wins do not decide overall. Tied overall.
- Pricing0:3
- Speed2:0
- Specs1:1
Verdict
Phi-3 Medium 14B offers ~3.9x larger context (128K vs 33K); Phi-4 Reasoning allows ~4.0x more max output (16K vs 4K); Phi-4 Reasoning is ~3.4x cheaper on a blended token basis than Phi-3 Medium 14B.
- Phi-3 Medium 14B offers ~3.9x larger context (128K vs 33K)
- Phi-4 Reasoning allows ~4.0x more max output (16K vs 4K)
- Phi-4 Reasoning is ~3.4x cheaper on a blended token basis than Phi-3 Medium 14B
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Phi-3 Medium 14B
- $0.34
- Phi-4 Reasoning
- $0.11
Phi-4 Reasoning is about 3.2x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Microsoft
Microsoft
Family
Phi
Phi
License
MIT
MIT
Open weights
Yes
Yes
Release
May 21, 2024
Apr 30, 2025
Knowledge cutoff
-
-
API / provider
Azure AI Foundry
Azure / self-host
Modalities
text → text
text → text
Specs
Context window
128K
33K
Max output
4K
16K
Parameters
14B
14B
Pricing
Input $/1M
$0.17
$0.07
Output $/1M
$0.68
$0.14
Blended $/1M (3∶1)
$0.30
$0.09
Speed
tok/s
90
80
TTFT (s)
0.28
0.5
Reasoning
MMLU-Pro
—
74.3
GPQA Diamond
—
65.8
Humanity's Last Exam
—
—
AIME 2025
—
62.9
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
—
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
53.8
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
37.6
—
SciCode
—
—
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,198
—
| Metric | Phi-3 Medium 14B | Phi-4 Reasoning | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Microsoft | Microsoft | — |
| Family | Phi | Phi | — |
| License | MIT | MIT | — |
| Open weights | Yes | Yes | — |
| Release | May 21, 2024 | Apr 30, 2025 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Azure AI Foundry | Azure / self-host | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 128K | 33K | A +95K |
| Max output | 4K | 16K | B +12K |
| Parameters | 14B | 14B | tie |
| Pricing | |||
| Input $/1M | $0.17 | $0.07 | B +$0.10 |
| Output $/1M | $0.68 | $0.14 | B +$0.54 |
| Blended $/1M (3∶1) | $0.30 | $0.09 | B +$0.21 |
| Speed | |||
| tok/s | 90 | 80 | A +10 |
| TTFT (s) | 0.28 | 0.5 | A +0.22 |
| Reasoning | |||
| MMLU-Pro | — | 74.3 | — |
| GPQA Diamond | — | 65.8 | — |
| Humanity's Last Exam | — | — | — |
| AIME 2025 | — | 62.9 | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | — | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 53.8 | — |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 37.6 | — | — |
| SciCode | — | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,198 | — | — |
Phi-3 Medium 14B
- Context
- 128K / 4K out
- Parameters
- 14B
- Price
- $0.17 / $0.68
- Speed
- 90 tok/s · 0.28s TTFT
- Modalities
- text → text
- License
- MIT
Phi-3 Medium 14B instruct — Microsoft's mid-size SLM with long context before Phi-3.5 / Phi-4.
Phi-4 Reasoning
- Context
- 33K / 16K out
- Parameters
- 14B
- Price
- $0.07 / $0.14
- Speed
- 80 tok/s · 0.5s TTFT
- Modalities
- text → text
- License
- MIT
Reasoning-tuned Phi-4 variant for math and science on constrained hardware.
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.