Phi-4: pricing, context window and benchmarks
Phi-4 is a multimodal model from Microsoft. It offers a 200,000-token context window, costs $3.96 per million input tokens and $20.63 per million output tokens, and averages 75.4 across the index's benchmark suite.
Pricing
Phi-4 costs $3.96 per million input tokens and $20.63 per million output tokens. A blended 3:1 input-to-output mix works out to about $8.13 per million tokens. Cached input is listed at $0.53 per million tokens.
| Item | Price | Notes |
|---|---|---|
| Input | $3.96 | per 1M tokens |
| Output | $20.63 | per 1M tokens |
| Cached input | $0.53 | per 1M tokens |
| Blended 3:1 | $8.13 | estimate per 1M tokens |
| 1M in + 1M out | $24.59 | estimate |
Benchmarks
Scores below are modelled estimates on a 0–100 scale. Use them to compare relative position, not as official vendor results.
| Benchmark | Score | Rank | Percentile | Relative |
|---|---|---|---|---|
| MMLU-Pro | 79.5 | 198 | 67th | |
| GPQA Diamond | 69.4 | 284 | 53th | |
| SWE-bench Verified | 66.5 | 180 | 70th | |
| HumanEval | 79.4 | 253 | 58th | |
| MATH-500 | 79.0 | 236 | 61th | |
| AIME 2025 | 74.1 | 180 | 70th | |
| LiveCodeBench | 64.7 | 303 | 50th | |
| IFEval | 83.0 | 246 | 59th | |
| TruthfulQA | 76.5 | 295 | 51th | |
| HellaSwag | 86.3 | 206 | 66th | |
| MMMU | 71.2 | 267 | 56th | |
| Arena Elo | 1392.0 | 240 | 60th |
What Phi-4 is good for
Phi-4 is an open-weight model, so it can be self-hosted or run through a third-party inference provider. That matters for data residency, fine-tuning and long-run cost control. At $3.96 per million input tokens (hosted list price) and a 200,000-token context window, it is a practical option for teams that want more control than a closed API allows.
Strengths
- Large 200,000-token context window for long documents and agent traces.
- Input price of $3.96 per million tokens is at or below the index median.
- Open weights make self-hosting, fine-tuning and data-residency review possible.
- Accepts more than text, including image input where the provider supports it.
Trade-offs
- No standout weakness in the indexed specifications, but real-world behaviour still needs testing.
How it compares
| Model | Provider | Context | Input / 1M | Avg score |
|---|---|---|---|---|
| Llama 3 DeepInfra Instruct | DeepInfra | 65536 | $3.84 | 75.5 |
| Yi-Lightning Chat | 01.AI | 32768 | $3.79 | 75.5 |
| Phi-4 | Microsoft | 200000 | $3.96 | 75.4 |
| Mistral Small 3 | Mistral AI | 32768 | $3.90 | 75.3 |
| Phi-4 Instruct | Microsoft | 128000 | $3.79 | 75.3 |
| Seed-OSS | ByteDance | 32768 | $3.74 | 75.3 |
FAQ
How much does Phi-4 cost?
Phi-4 is listed at $3.96 per million input tokens and $20.63 per million output tokens. Cached input is $0.53 per million tokens. A blended 3:1 input-to-output mix is about $8.13 per million tokens.
What is the context window of Phi-4?
Phi-4 has a context window of 200,000 tokens and can emit up to 128,000 output tokens per request in this index.
Is Phi-4 open weights?
Yes. Phi-4 is listed as an open-weight model under a open weights licence.
How does Phi-4 score on benchmarks?
Phi-4 averages 75.4 across the scored benchmarks in this index, ranking #240 of 600 models. Its strongest relative result is Arena Elo at 1392.0.
What is Phi-4 best for?
Phi-4 is an open-weight model, so it can be self-hosted or run through a third-party inference provider. That matters for data residency, fine-tuning and long-run cost control. At $3.96 per million input tokens (hosted list price) and a 200,000-token context window, it is a practical option for teams that want more control than a closed API allows.