Signum
Feed
Useful signal18 Nov 2026medium confidence

Microsoft AI unveils seven in-house MAI models at Ignite for code, security, image, speech and routing, priced below OpenAI and Anthropic equivalents

Edited by · written with AI from the linked sources (how)

At Ignite, Microsoft's MAI division released seven specialized models: MAI Code 2, MAI Code 2 Flash, MAI Work Flash, MAI Cyber 2 Flash (selected customers only), MAI Image 2.7 (plus Flash), MAI Realtime 1 and Decision 1.1 (built on Google's Gemma 4 26B A4B), and Retriever-1 (preview). Pricing is set: MAI Code 2 costs $0.50/$2.50 per million input/output tokens in GitHub Copilot at a 50% promotional price through May 17, 2027, then $1/$5. MAI Code 2 Flash is $0.05/$0.25, then $0.10/$0.50. MAI Cyber 2 Flash is $0.10/$0.50. Decision 1.1 is $0.042 per million input tokens, with free output. Enterprise customers will also be able to fine-tune selected model weights in Azure with their own data.

EconomicsCapabilityAccessAdoption

Entities: Microsoft, Microsoft AI (MAI), OpenAI, Anthropic, Google, Artificial Analysis

72Useful signal
1 source
0 primary
Was this useful?
01

What happened

At Ignite, Microsoft's MAI division released seven in-house models covering code, work tasks, security, images, speech and routing. Prices are set: MAI Code 2 is $0.50/$2.50 per million input/output tokens in GitHub Copilot at a 50% promotional rate until 17 May 2027, then $1/$5. MAI Code 2 Flash is $0.05/$0.25, rising to $0.10/$0.50. MAI Cyber 2 Flash ($0.10/$0.50) is for selected customers only, Retriever-1 is a preview, and Decision 1.1 is built on Google's Gemma 4 26B A4B rather than a wholly Microsoft base. Enterprises are also promised the ability to fine-tune selected model weights in Azure with their own data.

02

Why it matters

Microsoft is building a credible second source of models for its own products, which weakens its dependence on OpenAI and gives it leverage on price in negotiations with both OpenAI and Anthropic. Teams running always-on agents or heavy Copilot coding workloads now have a cheaper option to test against their current spend. Weight-level fine-tuning in Azure could matter to regulated buyers, but details and availability are not yet clear. Real impact depends on whether the models hold up on your own tasks, not on Microsoft's headline benchmarks.

03

What is noise

The claim that MAI Code 2 reaches 96-98% of GPT-6.1 Sol and Claude Opus 5.5 is Microsoft's own testing, and no primary evidence links are provided, so treat it as unverified. The "fraction of the cost" comparison leans on a 50% promotional price that lasts about six months, so the lasting gap is smaller. "Replacing outside models" in the headline overstates things: the coverage describes a lineup and some Copilot placement, not a confirmed wholesale switch, and one model rests on Google's Gemma.

04

Watch next

  1. 01Independent benchmark results, for example from Artificial Analysis, and real-world coding evaluations of MAI Code 2 against Opus 5.5 and GPT-6.1 Sol, especially on long agentic tasks rather than single prompts.
  2. 02Whether GitHub Copilot and Microsoft 365 products actually route meaningful traffic to MAI models, and whether OpenAI or Anthropic respond with price cuts or changed terms before the May 2027 promotion ends.
  3. 03Availability and terms of Azure weight fine-tuning, wider access to MAI Cyber 2 Flash and Retriever-1, and the list prices once the 50% discount lapses.

Coverage

1 story

Newsletter

Get the no-hype AI news straight to your inbox

Weekly. No spam, unsubscribe any time. Privacy

More economics signals

Full feed →