---
title: "Cheaper alternatives to Phi 4 · Undominated.ai"
canonical: https://undominated.ai/alternatives/microsoft__phi-4/
description: "12 models score at least as high as Phi 4 and cost no more on a balanced workload. 12 give up nothing measurable; the other 0 name what they drop. Measured prices and independent scores."
---

# Cheaper alternatives to Phi 4 · Undominated.ai

> 12 models score at least as high as Phi 4 and cost no more on a balanced workload. 12 give up nothing measurable; the other 0 name what they drop. Measured prices and independent scores.

# Cheaper alternatives to Phi 4

Microsoft · LMArena 1216.8 · $0.087/M on a balanced workload · prices as of 2026-08-28

## 12 models are both better and cheaper, giving up nothing.

Of the 133 models carrying both an independent score and a published price at the same delivery mode, **12** score at least as high as Phi 4 and cost no more under at least one workload. **12** match or beat it on every capability we hold — context, maximum output, input modes, tool use and extended reasoning.

Dominance is not a property of a model. It is a property of a model, a capability lens and a workload, and all three are named on every row. Scores are LMArena Elo, used under CC BY 4.0 from the official dataset; prices are blended per million tokens. A higher score is not a drop-in replacement.

This page is built from that comparison and nothing else. When no model is both better and cheaper than Phi 4, it is not generated.

## Where it holds

| Workload | Phi 4 $/M | Better and cheaper | With a trade |
| --- | --- | --- | --- |
| Balanced | $0.087 | 8 | 0 |
| Summarise | $0.073 | 11 | 0 |
| Chat | $0.098 | 8 | 0 |
| Code gen | $0.112 | 6 | 0 |
| Agentic | $0.081 | 11 | 0 |

Workload mixes are defined on the [methodology page](/methodology/). Cached input is priced at the cached rate, and reasoning tokens at the worse of the reasoning and output rates.

## What you would be replacing

| Intelligence | 1216.8 |
| --- | --- |
| Context window | 16K |
| Max output | 15K |
| Input modes | text |
| Tool use | no |
| Extended reasoning | no |

A replacement has to clear every line above, not just the score. [Full record for Phi 4](/models/microsoft__phi-4/).

## Better and cheaper, nothing given up 12

Each of these matches or beats Phi 4 on context, maximum output, input modes, tool use and reasoning, scores at least as high, and costs no more.

| # | Model | Intelligence | $/M | Holds under |
| --- | --- | --- | --- | --- |
| 1 | Qwen3 30B A3B Instruct 2507 Qwen | 1384.3 +167.5 | $0.084 | Balanced −4% Summarise −25% Agentic −13% |
| 2 | Solar Pro 4 Upstage | 1376.2 +159.4 | $0.052 | Balanced −40% Summarise −62% Chat −40% Code gen −27% Agentic −64% |
| 3 | gpt-oss-120b OpenAI | 1365.6 +148.8 | $0.070 | Balanced −20% Summarise −41% Chat −8% Agentic −29% |
| 4 | Gemma 3 12B Google | 1334.2 +117.4 | $0.075 | Balanced −14% Summarise −25% Chat −8% Code gen −2% Agentic −19% |
| 5 | Granite 4.1 8B IBM | 1291.6 +74.8 | $0.063 | Balanced −29% Summarise −29% Chat −29% Code gen −29% Agentic −29% |
| 6 | Gemma 3 4B Google | 1290.8 +74 | $0.063 | Balanced −29% Summarise −29% Chat −29% Code gen −29% Agentic −29% |
| 7 | gpt-oss-20b OpenAI | 1287.8 +71 | $0.055 | Balanced −37% Summarise −52% Chat −29% Code gen −20% Agentic −44% |
| 8 | Mistral Small 3 Mistral | 1233.6 +16.8 | $0.058 | Balanced −34% Summarise −30% Chat −37% Code gen −39% Agentic −32% |
| 9 | DeepSeek V4 Flash 0423 DeepSeek | 1431.6 +214.8 | $0.067 summarise | Summarise −9% Chat −4% Agentic −32% |
| 10 | MiMo-V2.5 Xiaomi | 1427.3 +210.5 | $0.079 agentic | Agentic −1% |
| 11 | GLM 4.7 Flash Z.ai | 1352.9 +136.1 | $0.063 summarise | Summarise −15% |
| 12 | GPT-5 Nano OpenAI | 1320.3 +103.5 | $0.055 summarise | Summarise −26% Agentic −6% |

## What this compares, and what it leaves out

 - Quality is LMArena Elo, used under CC BY 4.0 from the official dataset. The 95% confidence interval on a difference between two scores is about ±10.68 points, so a gap smaller than that is marked *tie* rather than an improvement — see [significance bands](/significance/).
 - 188 further models at this delivery mode carry a price but no independent score. They are absent from the comparison in both directions — unrated is not a zero, and an unmeasured model is neither an alternative nor a worse buy.
 - Retired models are never offered as an alternative, and a model only competes against its own delivery mode: batch trades latency for price, so it is not a like-for-like swap.
 - Nothing here measures latency, throughput, rate limits or how a model behaves on your prompts. Two models with the same index score are not interchangeable.
 - The models that nothing beats on both axes are on the [value frontier](/frontier/), and every other model something cheaper beats is [listed here](/alternatives/).