Phi-3 Pricing: Plans & API Token Cost Calculator 2026
Free to self-host under MIT, or roughly half a cent per thousand output tokens on Azure. Every cost here is hardware or a per-token rate, never a subscription.
Phi-3 plans and pricing
Phi-3-mini-4k-instruct
Phi-3-mini-128k-instruct
Phi-3.5-mini-instruct
Phi-3-small-8k-instruct
Phi-3-small-128k-instruct
Phi-3-medium-4k-instruct
Phi-3-medium-128k-instruct
Phi-3 pricing: the quick answer
Phi-3 Medium is a free open-source model, and hosted API access through Azure costs $0.14 per 1M input tokens and $0.56 per 1M output tokens as of July 8, 2026. The weights carry an MIT license, so downloading and self-hosting costs nothing beyond your own hardware. There is no subscription; the Azure serverless rate is pure pay-as-you-go across a context window up to 128k tokens. The output rate runs four times input, and both are low enough that a small on-device model like this is usually cheaper to run than any flat-rate chat plan, whether you host it yourself or route through Azure.
- Phi-3-mini-4k-instructCustom
- Phi-3-mini-128k-instructCustom
- Phi-3.5-mini-instructCustom
- Phi-3-small-8k-instructCustom
- Phi-3-small-128k-instructCustom
- Phi-3-medium-4k-instructCustom
- Phi-3-medium-128k-instructCustom
Phi-3 is free to start, against a $7.50/mo median across 12 large language models tools we track.
Phi-3 cost calculator
Phi-3 Hidden Costs & Pitfalls
The model is free to download, so every cost here is infrastructure or a managed-hosting rate. What you pay depends on whether you run it locally or hand the hosting to Azure.
Phi-3 pricing, read against its live plans and category
Positioning
Phi-3 Medium is free to download under an MIT license, and the only recurring cost is inference. On Azure's serverless API it runs $0.14 per 1M input tokens and $0.56 per 1M output, one of the cheapest hosted rates in the category. The family is built to run efficiently on small hardware, so self-hosting is genuinely viable, and the 128k context window comes without a flat-rate subscription. For developers who want a capable small model without monthly fees, the economics are hard to beat.
Cost drivers
- 1The token rate is only the inference bill. Fine-tuning or hosting a custom deployment on Azure adds separate cloud compute charges. Self-hosting locally is free of per-token fees but may require a hardware upgrade if your machine cannot hold the model. And because the Azure rate has no monthly floor, high-volume usage can accumulate faster than expected if nobody is watching the token count.
Watch-outs
Complaints center on tooling rather than billing, with some open-source runtimes slow to support the architecture.
Strengths
What makes it worth running:
- Runs efficiently on-device for offline AI on phones and modest hardware
- A 128k context window on a small, cheap-to-serve model
- MIT license with no per-token fee when you self-host
Editor’s take
Phi-3 Medium suits developers who want an efficient small model they can self-host for free or run on Azure for well under a cent per thousand tokens. If you need stronger reasoning on complex tasks than a small-parameter model delivers, step up to a larger open model or a managed frontier API instead.
Oleh KemFounder & Lead AnalystPhi-3 price history
Cheaper Large Language Models tools
Frequently asked questions
Sources & verification
| Source | What was checked | Last checked |
|---|---|---|
| Official Pricing Page | Source of verified tiers | July 8, 2026 |
| Official Website | Official vendor website | — |
| G2 | G2 verified user reviews · 4/5 | — |
| Capterra | Capterra verified user reviews · 4/5 | — |
Every fact on this Phi-3 pricing page is tied to a named source and a verification date. Freshness-sensitive figures trace to the sources above; verify against the vendor before relying on them.
Explore Phi-3
Every page on Phi-3 in one place, you are on pricing.
Snapshot, score and verdict
How to get API access, limits, SDKs and what it costs
You are here
Hidden fees, discounts and how to negotiate
Compared and ranked vs peers
Price and feature change history
Browse the full Large Language Models category