Phi 4
[Microsoft Research](/microsoft) Phi-4 is designed to perform well in complex reasoning tasks and can operate efficiently in situations with limited memory or where quick responses are needed.
Phi-4Released 10 Jan 2025 (Jan 2025)Pricing
List prices at Azure AI Foundry, in US dollars per million tokens.
| Tier | Input / 1M | Output / 1M | |
|---|---|---|---|
| Standard | $0.125 | $0.50 |
What it costs a month
At Azure AI Foundry list prices, 30 days.
- Chat assistant5,000 requests a day, 2K in / 0.5K out$75.00
- RAG over documents2,000 requests a day, 8K in / 0.6K out$78.00
- Coding agent1,000 requests a day, 40K in / 2K out$180
- Overnight batch20,000 requests a day, 4K in / 0.4K out$420
Where to run it
2 hosts serve Phi 4. The comparison is on blended price, 3 parts input to 1 part output.
| Host | In / out per 1M | Model id | |
|---|---|---|---|
Azure AI FoundryCloud Phi-4 | $0.125 / $0.50 Retires 17 Jan 2027 | Phi-4 | Retires 17 Jan 2027 |
DeepInfraInference phi-4 | $0.07 / $0.14 −60% vs headline | phi-4 |
Capabilities
Core
- Streaming
- Function calling
- JSON mode
- Structured outputs
Multimodal
- Vision
- Image input
- Audio
- Image generation
- Video
Advanced
- Reasoning
- Tool calling
- MCP
- Prompt caching
- Batch API
- Embeddings
- Takes in
- Text
- Produces
- Text
- Open weights
- License: MITRuns in about 35.2 GB of GPU memoryRuntimes: Ollama, llama.cpp, vLLM
Cheaper alternatives
Current models that cost less and keep most of what this one can do: tools, vision, reasoning, structured output, caching and enough context.
Same model, cheaper host
- DeepInfra60% lessphi-4$0.07 / $0.14
Smaller Microsoft models
No cheaper Microsoft model keeps its key capabilities.
Other frontier labs
- 9% less
- 9% less
Open-weight picks
- 9% less
- 18% less
- 18% less
- 27% less
Benchmarks
Scores published on the model card.
- BBH52.4%
- GPQA20.8%
- IFEval5.9%
- MATH Lvl 531.6%
- MMLU-Pro47.6%
- MUSR23.8%
- Open LLM Average30.4%
Call it
Using Azure AI Foundry's API.
from litellm import completion
response = completion(
model="azure_ai/Phi-4",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)Quick answers
How much does Phi 4 cost?
On Azure AI Foundry, Phi 4 costs $0.125 per million input tokens and $0.50 per million output tokens. The cheapest host is DeepInfra, at $0.07 per million input tokens and $0.14 per million output tokens.
Which APIs offer Phi 4?
Phi 4 runs on 2 APIs: Azure AI Foundry and DeepInfra.
What is Phi 4's context window?
16K tokens, with up to 16K tokens in one reply.
Is Phi 4 being retired?
Azure AI Foundry retires it on 17 Jan 2027. 1 host keeps serving it.
What are cheaper alternatives to Phi 4?
Models that cost less and keep most of what it can do: Claude Haiku 5.5 on Anthropic (9% less), GPT-6 Luna on OpenAI (9% less) and Ministral 14B 2512 on Mistral AI (9% less). Phi 4 itself costs 60% less on DeepInfra. Savings are on a mix of 3 input tokens to 1 output token.
When was Phi 4 released?
Microsoft released it on 10 Jan 2025.
What is Phi 4's API model id?
Use Phi-4 with Azure AI Foundry. Other hosts use their own ids: phi-4 on DeepInfra.
Sources
Raw catalog entry for Phi-4 on Azure AI Foundry
{
"deprecation_date": "2027-01-17",
"deprecation_source": "https://learn.microsoft.com/en-us/azure/foundry/concepts/model-lifecycle-retirement",
"successor_model": "microsoft--phi-4",
"input_cost_per_token": 1.25e-7,
"litellm_provider": "azure_ai",
"max_input_tokens": 16384,
"max_output_tokens": 16384,
"mode": "chat",
"output_cost_per_token": 5e-7,
"source": "https://prices.azure.com/api/retail/prices?$filter=serviceName%20eq%20'Foundry%20Models'%20and%20armRegionName%20eq%20'eastus'%20and%20priceType%20eq%20'Consumption'",
"supports_function_calling": true,
"supports_pdf_input": false,
"supports_tool_choice": true,
"supports_vision": false
}