vs
1:1
overall capability wins (1 ties across 3 scored). Phi-4 leads overall.
- Pricing0:3
- Speed0:2
- Specs1:1
Verdict
Phi-3.5-mini offers ~8.0x larger context (131K vs 16K); Phi-4 is ~2.6x cheaper on a blended token basis than Phi-3.5-mini.
- Phi-3.5-mini offers ~8.0x larger context (131K vs 16K)
- Phi-4 is ~2.6x cheaper on a blended token basis than Phi-3.5-mini
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- Phi-3.5-mini
- $0.26
- Phi-4
- $0.11
Phi-4 is about 2.5x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
Microsoft
Microsoft
Family
Phi
Phi
License
MIT
MIT
Open weights
Yes
Yes
Release
Aug 21, 2024
Dec 12, 2024
Knowledge cutoff
-
-
API / provider
Azure AI Foundry
Azure / self-host
Modalities
text → text
text → text
Specs
Context window
131K
16K
Max output
4K
4K
Parameters
3.8B
14B
Pricing
Input $/1M
$0.13
$0.07
Output $/1M
$0.52
$0.14
Blended $/1M (3∶1)
$0.23
$0.09
Speed
tok/s
130
150
TTFT (s)
0.2
0.15
Reasoning
MMLU-Pro
—
70.4
GPQA Diamond
—
57.5
Humanity's Last Exam
—
3.8
AIME 2025
—
18.0
MATH-500
—
81.0
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
—
14.1
AA-LCR v1.1
—
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
—
23.5
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
23.1
Terminal-Bench 2.1
—
—
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
45.5
SciCode
—
26.0
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-3.5-mini | Phi-4 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | Microsoft | Microsoft | — |
| Family | Phi | Phi | — |
| License | MIT | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Aug 21, 2024 | Dec 12, 2024 | — |
| Knowledge cutoff | - | - | — |
| API / provider | Azure AI Foundry | Azure / self-host | — |
| Modalities | text → text | text → text | — |
| Specs | |||
| Context window | 131K | 16K | A +115K |
| Max output | 4K | 4K | tie |
| Parameters | 3.8B | 14B | B +10 |
| Pricing | |||
| Input $/1M | $0.13 | $0.07 | B +$0.06 |
| Output $/1M | $0.52 | $0.14 | B +$0.38 |
| Blended $/1M (3∶1) | $0.23 | $0.09 | B +$0.14 |
| Speed | |||
| tok/s | 130 | 150 | B +20 |
| TTFT (s) | 0.2 | 0.15 | B +0.05 |
| Reasoning | |||
| MMLU-Pro | — | 70.4 | — |
| GPQA Diamond | — | 57.5 | — |
| Humanity's Last Exam | — | 3.8 | — |
| AIME 2025 | — | 18.0 | — |
| MATH-500 | — | 81.0 | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | — | 14.1 | — |
| AA-LCR v1.1 | — | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | — | 23.5 | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | 23.1 | — |
| Terminal-Bench 2.1 | — | — | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | 45.5 | — |
| SciCode | — | 26.0 | — |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | 1,256 | — |
Phi-3.5-mini
- Context
- 131K / 4K out
- Parameters
- 3.8B
- Price
- $0.13 / $0.52
- Speed
- 130 tok/s · 0.2s TTFT
- Modalities
- text → text
- License
- MIT
Prior Phi SLM generation — still common on Azure for low-cost classification before the Phi-4 refresh.
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.
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.