Microsoft launches cheap AI models the same day OpenAI cuts prices

Reviewed byNidhi Govil

10 Sources

Share

Microsoft launched a family of in-house AI models at Ignite, claiming near-parity with Claude and GPT at a fraction of the price. The claims rest on Microsoft-chosen benchmarks and a 50% promotional discount, and the company still sells its rivals' models.

Microsoft put its own AI models up against OpenAI and Anthropic on Tuesday, announcing a family of in-house systems for coding, cybersecurity, image generation and speech at the Ignite conference in San Francisco. The move puts Microsoft in direct competition on price and performance with the same firms it invests in and resells

7

2

5

4

.

Source: ZDNet

Source: ZDNet

Mustafa Suleyman, chief executive of Microsoft AI, says specialization is the trick: "It basically means that you cram more tokens and knowledge and expertise into the same amount of parameters"

5

. "This isn't just one model. It's not a flash in the pan. This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price," he told Reuters

7

. The company says MAI Code 2 delivers 96% of the coding performance of Claude Opus 5.5 for 5% of its cost

5

3

1

6

4

. It puts the model at 98% of GPT-6.1 Sol's performance for 40% of the price

5

. A smaller Flash version, Microsoft says, matches GPT-6 Luna for 65% of the price and Claude Haiku 5.5 for 20% of the cost

5

6

4

.

Source: Axios

Source: Axios

The fine print in Microsoft's own numbers

The caveats come from Microsoft itself, per its own disclosures. The cost figures use a 50% introductory discount, and the benchmarks were Microsoft's own choices. The company's documentation says the models are not built for the hardest agentic, research or mathematical work

1

6

5

. ZDNet's columnist flags that Microsoft picked its own benchmarks, though he suspects the numbers are not padded, since Microsoft would benefit most if they were honest

1

.

On security, Microsoft's own evaluation has MAI Cyber 2 Flash scoring 92.3% on the CyberGym end-to-end vulnerability-discovery benchmark, at under 2% of the cost per task

5

6

. OpenAI's GPT-6 Astra scored 91.5% on the same test

6

. That is a 0.8-point margin on a test Microsoft scored, after it excluded 92 of 905 tasks for overlapping with its training data

5

3

6

4

. There is no comparison yet against OpenAI's restricted GPT-6 Cyber model

6

.

One independent ranking backs the image model, and it is the launch's only externally checked claim

6

. MAI Image 2.7 tops Artificial Analysis's human-preference editing leaderboard, 1,215 points to 1,211 for GPT Image 2.6 Sunburst, a four-point gap whose uncertainty ranges overlap

6

. It ranks fourth on the same site's text-to-image list

6

. Microsoft prices it at about $35 per 1,000 images against roughly $211 for the same GPT Image 2.6 Sunburst at maximum quality

6

4

5

. Decrypt's quick hands-on tests found its outputs beat GPT Image and Google's Nano Banana on realism and overall quality. The reviewer tried a few generations, and noted the model is heavily censored

6

.

Attacks on stage, and a rival's same-day answer

The keynote went after the rivals by name. "Simply stated, anything you can do in Claude or ChatGPT, you can do better with Copilot and for less," said Judson Althoff, chief executive of Microsoft's commercial business

3

. Copilot chief Jacob Andreou said rivals' document editors force compromises: "That means that Copilot is the only way to draft a PowerPoint presentation without spinning a meter"

3

. A GigaOm study Microsoft commissioned put the monthly enterprise cost per user at $30 for Copilot versus $81 for ChatGPT and $205 for Claude

3

. The study was paid for by Microsoft, based on modeled workloads, and uses enterprise pricing that is not comparable to the consumer subscriptions rivals advertise

3

.

Source: GeekWire

Source: GeekWire

Going by the internal figures, the distribution gap shows up inside Microsoft's own products. Internal data reported by The Information puts Claude for Microsoft 365 at under a tenth of the weekly users of paid Copilot in Office

3

. Anthropic, meanwhile, launched Claude-based spreadsheet and whiteboard tools last week, moving into Microsoft's core business, and its IPO is expected this week

3

.

OpenAI answered the same day, releasing GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, with the new Astra priced 20% below its predecessor. Satya Nadella announced from the Ignite stage that GPT-6.2 is available in Microsoft Foundry, which means Microsoft's Sol comparison targets the previous generation

5

3

6

.

A portfolio, and a chip under it

The lineup runs to seven models on Interesting Engineering's list, including MAI Realtime 1, which Microsoft calls its first speech-to-speech model, priced from $0.01 per audio minute

4

5

. This is the second wave of in-house models this year. Microsoft released seven homegrown models at its Build conference in June, part of what Suleyman has called a push for long-term self-sufficiency

3

. Microsoft also said its Maia 200 AI accelerator is in production and running its biggest models. By the company's account alone, it delivers as much as 1.7 times more tokens per watt than leading GPU hardware

4

.

Microsoft says that predecessor handles almost 40% of scans in MDASH and keeps 90% of scans that start with it

7

6

4

. For enterprises that want the models on their own data, Microsoft launched MAI Licensed Weights, letting customers customize the models and run them inside Azure. Telecom carrier AT&T built a 5G network analytics model through the predecessor program

5

3

4

. Suleyman says the setup gives customers open-source-style flexibility with more security and control

5

. Suleyman also said Microsoft is building a frontier general-purpose model with state-of-the-art pre-training results, but declined to give a release timeline; "My team is building true superintelligence," he said

5

.

Microsoft remains an investor in OpenAI and Anthropic, and still sells their models through Copilot and Foundry

3

1

. MAI Code 2 ships as an optional choice in GitHub Copilot from launch day

5

. Suleyman declined to give a target for how much Copilot traffic would shift, predicting significant adoption

3

1

5

.

© 2026 TheOutpost.AI All rights reserved