Microsoft · language model
Phi 4: pricing, context and benchmarks
Phi 4 is a language model from Microsoft released on January 10, 2025. It costs $0.07 per million input tokens and $0.14 per million output tokens, with a 16.4K-token context window. Epoch AI ranks it #115 of 125 rated models.
What you need to know about Phi 4
Phi 4 was released by Microsoft (United States) on January 10, 2025. It accepts text as input and produces text. Its context window reaches 16,384 tokens, larger than 4 % of the models in the registry, and it can write up to 14,745 tokens in one go.
Released more than 1 year ago, it belongs to a generation most vendors have since superseded.
Phi 4 pricing
On OpenRouter, Phi 4 costs $0.07 per million input tokens and $0.14 per million output tokens. Counting three tokens read for every token written, its average price is $0.088 per million, more expensive than 7 % of paid models (8.7 times cheaper than the median of $0.76).
A concrete example: a task that sends 2,000 tokens (about 1,500 words) and receives a 500-token answer costs $0.0002. At 1,000 tasks a day, budget $6 a month. The price calculator redoes this with your own volumes.
Performance and benchmarks
The Epoch Capabilities Index (ECI), which aggregates published results across many benchmarks, gives Phi 4 a score of 130.5 (range 124.7 to 132). That places it #115 of 125 rated models.
Scores published by Epoch AI on the benchmarks the site tracks:
| Benchmark | What it measures | Score | Date |
|---|---|---|---|
| GPQA Diamond | PhD-level science questions | 56.1 % | Dec 12, 2024 |
| OTIS Mock AIME | maths competition problems | 13.8 % | Dec 12, 2024 |
For at least equal performance, Qwen3.7 Flash and gpt-oss-20b cost at least 30% less.
Features, privacy and local use
- It does not support tool calling: it is not suited to driving an AI agent that has to act.
- It can answer in an enforced JSON format (structured outputs).
- Its knowledge stops at June 2024.
Its weights are open and downloadable from Hugging Face (mit licence): you can run it on your own machines without sending your data to a third party. With 14,659,507,200 billion parameters, it needs about 8,191,637,993 GB of memory at 4-bit quantisation. Check with the VRAM calculator. It was downloaded 364,764 times on Hugging Face last month.
Verdict
Phi 4 scores 40/100 on our grid, which only uses published figures.
Strengths
- Low price: cheaper than 93 % of paid models.
- Open weights: self-hostable, your data stays with you.
Limitations
- Short context (16.4K tokens): long documents must be split.
- No tool calling: unusable as an agent's brain.
- Modest performance: behind 91 % of rated models.
- Older generation, often replaced by a newer model from the same vendor.
- Text only: reads neither images nor files.
Avoid if…
- you are building an agent that must call tools;
- you analyse long documents in one go;
- you want to host it on a single consumer computer (≈ 8,191,637,993 GB needed at 4-bit).
Alternatives
- Gemma 3 12B$0.08
- Qwen3 14B$0.15
- Gemma 3 27B$0.15
- Qwen3 30B A3B$0.22
- gpt-oss-20b$0.04
- Gemma 3 4B$0.04
Comparisons and tools
Learn more
Frequently asked questions
How much does Phi 4 cost?
$0.07 per million input tokens and $0.14 per million output tokens on OpenRouter (checked October 25, 2026). A task with 2,000 tokens read and 500 written costs $0.0002.
How big is Phi 4's context window?
16,384 tokens, roughly 12,000 words, with a maximum of 14,745 output tokens.
Can Phi 4 run locally?
Yes, its weights are published under the mit licence. Expect about 8,191,637,993 GB of memory at 4-bit.
Is Phi 4 suitable for an AI agent?
No: without tool calling it cannot act, only answer.
How good is Phi 4?
Epoch AI ranks it #115 of 125 rated models (ECI 130.5). On GPQA Diamond, it scores 56.1%.
What are the alternatives to Phi 4?
The closest models on price and performance are Gemma 3 12B, Qwen3 14B, and Gemma 3 27B.
Related articles
- Mistral AI: models, pricing and agentsOctober 25, 2026
- LLM benchmarks: how to read model rankingsOctober 24, 2026
- Running a local LLM: hardware, software, and limitationsOctober 24, 2026
- Open-source or closed models for your AI agents?October 5, 2026
Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score