Microsoft API: models, pricing and limits
A practical guide to microsoft, using data from the AI & LLM Index. The goal is a shortlist and a decision, not a tour of every option.
The Microsoft lineup
Microsoft has 14 models in the index. Software company behind the Phi models. The lineup spans 3 categories, from multimodal to small open-weight.
The strongest model by average score is phi-4. The cheapest by input price is Phi-3.5 1.5B.
Pricing
Across the lineup, the average input price is $2.04 per million tokens and the average output price is $9.46. That average hides a wide range: individual models run from $0.25 to $3.96 on input.
Use the average as a starting point, then check the specific model you plan to deploy. Output price and cached input pricing can change the ranking for long or repetitive workloads.
Strengths and limits
14 of the 14 models are open weights. That gives you a self-hosting path for part of the lineup, while the proprietary models remain API-only.
The best benchmark result in the lineup is 75.4 on the index average. Compare that with the best model from other providers before treating Microsoft as the default.
How to choose
- Start with the task, not the brand.
- Pick the cheapest model that clears your evaluation set.
- Keep a second provider as a fallback for outages and price changes.
- Check licence and data-handling terms for every model you shortlist.
Microsoft models at a glance
| Model | Input / 1M | Context | Avg score |
|---|---|---|---|
| Phi-4 | $3.96 | 200,000 | 75.4 |
| Phi-4 Instruct | $3.79 | 128,000 | 75.3 |
| Phi-4 Chat | $3.66 | 65,536 | 74.7 |
| Phi-4 Base | $3.15 | 32,768 | 71.2 |
| Phi-3.5 | $2.81 | 16,384 | 70.9 |
| Phi-3.5 Instruct | $2.67 | 16,384 | 70.1 |
| Phi-3.5 Chat | $2.64 | 16,384 | 69.0 |
| Phi-3.5 Base | $2.24 | 65,536 | 66.6 |
| Phi-4 7B | $0.91 | 8,192 | 56.8 |
| Phi-4 8B | $0.96 | 32,768 | 54.9 |
Related
FAQ
Which Microsoft model should I start with?
Start with phi-4 for quality-sensitive work and Phi-3.5 1.5B for high-volume tasks.
What is the most important factor in microsoft?
Fit to the task beats raw benchmark rank. Define the failure modes, pick the benchmarks that resemble them, then test a small shortlist on your own examples.
How many models should I evaluate?
Three is usually enough: a quality candidate, a value candidate and a fallback. More candidates add work without changing the decision.
How do I control cost?
Compare blended 3:1 price, watch output tokens, use caching where available and route simple requests to a cheaper model.