---
title: "Microsoft's coding model costs 40% of a rival, but the price tag has an expiry date"
date: 2026-11-20
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-model-mai-code-pricing-1120
language: en
---

# Microsoft's coding model costs 40% of a rival, but the price tag has an expiry date

> Microsoft shipped seven in-house models. MAI Code 2 matches GPT-6.1 Sol at 98% quality for 40% of the price, but that price is a six-month promotion that doubles next May.

On November 17, at its Ignite conference in San Francisco, Microsoft released seven in-house models at once.

The headliner was MAI Code 2, a coding model that went live inside GitHub Copilot the same day. Microsoft's claim is blunt: averaged across four coding benchmarks, it matches GPT-6.1 Sol at 98% of the quality and 40% of the price.

The line aimed at Anthropic is sharper still: about 96% of Claude Opus 5.5, at about 5% of the cost.

## Four price tags, one order of magnitude apart

According to Microsoft's published test data, MAI Code 2 costs $0.309 per task on average. GPT-6.1 Sol comes in at $0.761, and Claude Opus 5.5 at $6.306.

Set three numbers side by side and the gap is roughly an order of magnitude. It is also the easiest thing to remember from the keynote.

The lighter MAI Code 2 Flash is cheaper still, at $0.052 per task under promotional pricing. Across DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas, it averaged a 61.0% pass rate, against 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5.

Access is tiered. Code 2 is limited to Copilot Pro, Business and Enterprise subscribers, while Flash reaches Free and Student users as well.

## The 40% is a discount, not a list price

Beijing News reported that the pricing uses a 50% promotional rate, and that Code 2's unit price doubles from May 17, 2027.

That single sentence rewrites every comparison above.

After the doubling, 40% becomes 80%, and 5% becomes 10%. The same task that costs $0.309 today would cost about $0.618, against a rival's $0.761. Still cheaper, but no longer dramatically so.

A practitioner told Beijing Business Today something worth keeping: the unit price is 25% of the competitor's, yet the measured cost per run is 40%, which means Code 2 burns more tokens to solve a task. Cheap per token is not the same as cheap per job. The bill is price multiplied by consumption.

## Security and images shipped the same day

MAI Cyber 2 Flash is the security entry. Microsoft says it scores 92.3% on the CyberGym-E2E benchmark, against 91.5% for GPT-6 Astra, at less than 2% of the task cost. It will not be offered through a general API; from mid-December it rolls out to eligible customers inside the MDASH system.

MAI Image 2.7 tops the Artificial Analysis image-editing leaderboard with an Elo of 1,215, four points above GPT Image 2.6 Sunburst at 1,211.

Four points sits inside the statistical error band, and Microsoft acknowledges as much in its own materials.

One more line is easy to skip: Microsoft says Code 2 targets everyday enterprise coding and has not been tuned for fundamental research or very long-running agentic tasks.

The discount has boundaries, and the boundaries are in the footnotes.

## What is really being sold is habit

Mustafa Suleyman, who leads Microsoft AI, told Reuters that this is not a single model but a "hill-climbing machine" that co-optimizes performance and price.

Put the seven models together and the strategy reads clearly. Microsoft is not chasing first place on every leaderboard. It is trying to replace externally purchased models inside surfaces it already owns: Copilot, Microsoft 365, Azure.

The six months of promotional pricing buys exactly that replacement window. Developers who build habits at half price face migration costs when the price doubles.

So the 40% is not a measure of technical advantage. It is a measure of customer acquisition cost. The real test arrives next May, when the tag changes.

## The margin question nobody asked

Cheap tokens do not automatically produce cheap software. Microsoft is also one of the largest resellers of OpenAI models, and every workload it moves from a partner model to an in-house one improves its own gross margin, because it stops paying a partner for tokens it can generate itself.

That is the part of the announcement that matters to the income statement, and it does not appear in any benchmark table. MAI Code 2 was trained on enterprise coding tasks of the kind that fill GitHub Copilot logs: fixing defects, editing across files, reading a repository. Those tasks are high-volume, low-glamour and highly repeatable, which is exactly where a price advantage compounds.

The risk sits on the other side. If enterprises route work to the cheaper model and it needs more tokens to finish, the savings reappear as rework and review time. A small quality gap on a benchmark can become a much larger gap in a code review queue.

*Disclaimer: This article draws on Microsoft's public disclosures and reporting by Beijing News, Phoenix and Reuters. It is for reference only and does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-mai-code-pricing-1120
Free to quote with attribution and a link to the original.
