// Microsoft
Phi 4
A Microsoft model. 1 prices across 1 channels side by side, each with its source and check date – plus specs, licence, speed and benchmarks.
- Official in/out per 1M
- —
- Lowest (DeepInfra)
- $0.07 / $0.14
- Context
- 16K
- AA index
- 5.9
Prices checked 2026-10-22Official sources linked on every rowNo estimates
01Specs at a glance
Official values first; gaps filled from OpenRouter’s API where the vendor page is silent.
| Vendor | Microsoft |
|---|---|
| Context window | 16K (16,384 tokens) |
| Max output | 15K |
| Input | Text |
| Output | Text |
| Tool calling / JSON | No / Yes |
| Weights & licence | Open weights · MIT · Hugging Face |
| Parameters | 14.7B (from weight files) |
| Released | Jan 10, 2025 |
| Model IDs | microsoft/phi-4 |
02Prices on 1 channels
Per 1M tokens, in the currency each source publishes (1 USD = 6.7 CNY for the blended column). Sorted: official first, then cheapest.
| Channel | Source | Input | Output | Cache hit | Blended $ | Context | Notes |
|---|---|---|---|---|---|---|---|
| DeepInfra · bf16 Via OpenRouter / provider listed there | OpenRouter 10-22 | $0.07 | $0.14 | — | 0.087 | 16K | DeepInfra price as listed on OpenRouter (pass-through; quantization=bf16; discount field=0); not checked against DeepInfra's own page |
“Official” = vendor pricing page; “Platform” = the platform’s own pricing page; “OpenRouter” = price listed on OpenRouter, not checked on the platform’s site. Click a source tag to open the original page.
03Where you can use it
Derived only from the price sources above. Country restrictions are set by each vendor – check their terms before you sign up.
- Via OpenRouter / provider listed there
DeepInfra - Self-host the open weights
GPU / VRAM sizing and build plans→ - Free option
phi-4→
04Performance
All results →Speed and AA index from Artificial Analysis public leaderboards (not our own tests); benchmark scores from our benchmark library, best result per benchmark.
- AA indexPhi-4
- 5.9
- Output speed (median)
- 40 tok/s
- Time to first token
- 2.67s
- AA index per $ (blended)
- 67.69
| Benchmark | Best score | Source | Date |
|---|---|---|---|
| LMArena Text | 1256 | official | 2026-10-08 |
| Open LLM Leaderboard v2 | 30.4% | official | 2025-03-13 |
| GPQA Diamond | 56.1% | independent | 2025-01-31 |
| MMLU | 84.8% | paper | 2024-12-12 |
| MMLU-Pro | 70.4% | vendor | 2026-03-11 |
| AIME (OTIS Mock AIME 2024–2025) | 13.8% | independent | 2025-02-25 |
| MATH Level 5 | 64.9% | independent | 2025-01-31 |
| BFCL v4 (function calling) | 28.8% | official | 2026-10-17 |
05Price changes we recorded
All price changes →From our daily re-checks of the sources above.
No price change recorded since we started tracking (2026-10-10).
06More from Microsoft
FAQ
How much does the Phi 4 API cost?
There is no official pay-as-you-go price in our data. The lowest listed price is $0.07 / $0.14 on DeepInfra.
What is the context window of Phi 4?
16K tokens of context, up to 15K output tokens. Platforms may cap it lower – see the table.
Is Phi 4 open source?
Yes, the weights are published on Hugging Face (microsoft/phi-4). Licence: MIT.