Microsoft Launches Seven MAI Models to Undercut GPT and Claude Costs
Microsoft AI drops seven models at Ignite that undercut GPT and Claude on price while matching performance on everyday coding, image editing, voice and security tasks.
- Microsoft launched seven new MAI models at Ignite focused on undercutting GPT and Claude on cost per task
- MAI Code 2 hits 98% of GPT 6.1 Sol quality at 40% of the price, live in GitHub Copilot
- MAI Cyber 2 Flash matches GPT-6 Astra on CyberGymE2E at under 2% of the per-task cost
- MAI Image 2.7 ranks #1 on Artificial Analysis image editing at one fifth of GPT Image 2.6's price
- MAI Realtime 1 speech-to-speech starts at $0.01 per audio minute, beating GPT-Realtime 2.1
- MAI Licensed Weights lets enterprises run and fine-tune frontier MAI models inside their own Azure tenant
Microsoft introduced seven in-house MAI models at Ignite 2026, spanning coding, image generation and editing, real-time voice, transcription, and cyber defense. The company also announced a model router and an Azure licensing program. Its Ignite overview centers on a shared commercial claim: specialized models can approach the quality of larger GPT and Claude systems while costing substantially less to run.
Microsoft reports some comparisons as cost per task and others as usage prices, so the percentages are not directly interchangeable. Developers evaluating the lineup need to measure total cost per accepted result, including retries, tool calls, latency, human corrections, and routing overhead.
MAI spreads across the stack
| Product | Primary workload |
|---|---|
| MAI Code 2 | Repository-level coding, debugging, and front-end development |
| MAI Code 2 Flash | Lower-cost, lower-latency coding tasks |
| MAI Image 2.7 and Flash | Image generation and editing |
| MAI Realtime 1 | Live speech-to-speech applications |
| MAI Cyber 2 Flash | Continuous security scanning and remediation |
| MAI Transcribe 2 | Clinical transcription through Dragon Copilot |
| Decision 1.1 | Routing requests among models |
| MAI Licensed Weights | Private model adaptation inside an Azure tenant |
Code 2 targets the pull-request queue
Microsoft averaged MAI Code 2 results across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1, benchmarks intended to approximate repository work, terminal use, and software issue resolution. On that composite, the company says Code 2 delivered 98% of GPT-6.1 Sol’s performance at 40% of the cost and 96% of Claude Opus 5.5’s performance at 5% of the cost.
MAI Code 2 Flash targets faster routine work. Microsoft prices it at 65% of GPT-6 Luna’s cost and 20% of Claude Haiku 5.5’s cost on a similar task mix. Those aggregate figures can conceal substantial differences by language, repository size, context length, tool use, and output length.
Microsoft describes both models as optimized for targeted edits, repository-wide reasoning, debugging, and merge-ready patches. Their current scope excludes fundamental science research and very long-running agent tasks, such as autonomous refactors that span a large codebase for several days.
In blind side-by-side evaluations of front-end application work, reviewers preferred Code 2 over GPT-6 Sol in 67% of cases. Microsoft says it tuned the model for dashboards, visualizations, and other visually oriented coding tasks.
Both models are available in GitHub Copilot, according to the Code 2 announcement. Code 2 Flash is also in limited-access preview through Microsoft Foundry. Microsoft lists OpenRouter, Vercel, Hermes, OpenCode, and Kilo as upcoming integrations.
Media models push down per-minute costs
MAI Image 2.7 reached first place on Artificial Analysis for image editing, according to Microsoft. The company says it beat GPT Image 2.6 Sunburst at its maximum setting while costing one-fifth as much. Microsoft also announced a Flash variant, although the supplied launch details do not provide separate benchmark or pricing figures for it.
MAI Realtime 1 is Microsoft’s first live speech-to-speech model. The company reports higher scores than GPT-Realtime 2.1 on external tests and Dynamics 365 benchmarks, with pricing starting at $0.01 per audio minute. A trial is available through MAI Labs.
Dragon Copilot Transcribe, built on MAI Transcribe 2, now handles all doctor-patient conversation transcription within Dragon Copilot. Microsoft reports 10 million encounters each month, a 50% reduction in multilingual transcription errors, and adoption among one-third of US physicians.
Cyber 2 Flash is priced to stay on
Continuous security agents create a recurring inference bill because they monitor systems, inspect changes, and propose patches around the clock. Microsoft positions MAI Cyber 2 Flash for that workload, where a modest per-task saving can compound across large scan volumes.
On CyberGymE2E, an end-to-end cybersecurity benchmark, Microsoft says Cyber 2 Flash matched GPT-6 Astra’s quality at less than 2% of the cost per task. MAI Cyber now accounts for almost 40% of scans in MDASH, and 90% of scans that begin with MAI continue using it. Output-token pricing has fallen 86% from the previous Flash release.
Licensed weights move into Azure tenants
Model weights are the numerical parameters learned during training and used to generate a model’s responses. Through MAI Licensed Weights, Microsoft plans to deploy those parameters inside a customer’s Azure environment, allowing the customer to adapt the model with proprietary data while keeping the resulting checkpoint private.
Under Microsoft’s stated policy, customer data and adaptations remain under customer control and enter Microsoft’s training process only with express agreement. The arrangement provides direct access to the model inside Azure without releasing the weights publicly.
Microsoft cites three customer projects to support its specialization strategy:
- AT&T used MAI models to analyze 5G networks. Microsoft says MAI outperformed GPT-6.1 Sol on standard tasks at one-fiftieth of the cost.
- Woven by Toyota used MAI models to repair software. The system produced merge-ready fixes in 44% of cases where its existing frontier models required human intervention, at less than one-tenth of the cost.
- Alaska Airlines increased aircraft maintenance report classification accuracy from 81% to 91%, exceeding the results from leading frontier models at a lower cost.
Key contract variables for licensed weights include update rights, fine-tuning permissions, Azure-region availability, portability, telemetry, safety controls, indemnity, and restrictions on redistribution. Those terms will determine how much operational control the program provides in practice.
Decision 1.1 controls model escalation
Decision 1.1 examines incoming requests and routes each one to an appropriate model. Routine work can remain on a cheaper system, while requests requiring deeper reasoning can reach a larger model. This approach lowers blended inference cost when the router classifies requests accurately.
Microsoft says Decision 1.1 leads Jev, Google, and OpenAI on the Artificial Analysis Decision Index. The company also reports that it runs 50 times faster and 79 times cheaper than GPT-6.1 Sol. Production evaluations should track routing errors because an incorrect downgrade can create retries, poor outputs, or additional human review.
Specialization changes the model stack
Microsoft’s portfolio supports a tiered architecture in which Flash models handle high-volume requests, specialized checkpoints serve narrow domains, and a router escalates selected tasks. Licensed weights add private adaptation for organizations that need proprietary knowledge embedded within an Azure deployment.
MAI Work Flash illustrates how Microsoft is applying that structure internally. Derived from Code 2 Flash, it has been further optimized to serve multiple products across Microsoft 365 and Fabric IQ. The model family therefore functions as a set of workload-specific components instead of a single general-purpose endpoint.
Test cost per accepted result
A production pilot should replay representative workloads and compare the complete path from request to accepted output. Useful evaluation criteria include:
- Task success: accepted patches, resolved incidents, accurate classifications, or completed conversations.
- Human effort: review time, corrections, rejected outputs, and escalations.
- Blended cost: input and output tokens, audio minutes, tool calls, retries, and router overhead.
- Operational performance: latency, throughput, rate limits, regional availability, and failure recovery.
- Deployment controls: data retention, private networking, audit logs, model updates, and licensing rights.
Teams already using GitHub Copilot or Microsoft Foundry can compare the Code 2 family against their current models on the same repositories and acceptance criteria. Voice developers can evaluate whether Realtime 1’s per-minute price supports longer sessions, while organizations considering private model adaptation can assess Licensed Weights against their governance and portability requirements.