M
Phi 4OPEN
Microsoft Research 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.
Prices update automatically — checked daily against provider list prices.
- $0.09 per 1M tokens, assuming 3 input for 1 output the convention used by Artificial Analysis
- cheaper than 80% of the 191 open-weight models tracked here
Benchmarks & performance
Independent benchmark scores for Phi 4, measured by Artificial Analysis. Higher is better.
For the same budget, Phi 4 delivers 3.3 times more capability. Open weights also mean you can self-host it and pay nothing per token.
Models at this level cost $0.4 per 1M tokens (median of 14) — this one costs $0.09.
Cheaper and better on this index: DeepSeek V4 Flash 0731 · Solar Pro 4 · Ling-3.0-flash and 8 more
About this model
Phi 4 is an open-weight AI model by Microsoft. You can download and self-host it for free; the prices below are hosted-API list prices, tracked daily, for when you prefer convenience over self-hosting.
Frequently asked questions
What is Phi 4?
Phi 4 is an AI language model from Microsoft. It is open-weight: you can download it and run it on your own hardware, for free. It scores 4.6 on the Artificial Analysis intelligence index.
Is Phi 4 free?
The weights are free and open — you can self-host Phi 4 and pay nothing per token. If you prefer a hosted API, list prices are $0.07 per million input tokens and $0.14 per million output tokens.
What is Phi 4 good at?
Independent benchmarks from Artificial Analysis give it GPQA 57.5%, MMLU-Pro 71.4%, Humanity's Last Exam 3.8%, Long Context Reasoning 0%, LiveCodeBench 23.1%, SciCode 26%, MATH-500 81%, AIME 14.3%, AIME 2025 18%, IFBench 23.5%, τ²-Bench 0%, Terminal-Bench Hard 3.8%. It is particularly used for mathematical reasoning.
How fast is Phi 4?
It generates about 43.4 tokens per second, with a median 0.93s delay before the first token. Measured independently by Artificial Analysis.
Can I self-host Phi 4?
Yes. Phi 4 has open weights, so you can download it and run it on your own GPU or server with tools like Ollama, vLLM or llama.cpp — with no per-token cost.