Microsoft ships seven in-house AI models that undercut OpenAI on price
Microsoft's Ignite batch spans code, security, image, speech and routing models, with MAI Code 2 hitting 98% of GPT-6.1 Sol's performance at 40% of the cost, plus licensed fine-tuning for enterprises.

Why it matters
- MAI Code 2 averages 98 percent of GPT-6.1 Sol's performance at 40 percent of the cost on four benchmarks, Microsoft says.
- MAI Cyber 2 Flash scores 92.3 percent on CyberGym-E2E and costs $0.45 per task versus $34.38 for GPT-6 Astra.
- Microsoft built its new decision model on Google's Gemma 4 26B A4B; it scores 68 on Artificial Analysis' Decision Index v2.1.
- MAI Licensed Weights will let enterprises fine-tune MAI models in their own Azure environments, launching 'shortly' with no pricing yet.
- Promotional MAI Code 2 pricing runs through May 17, 2027: $0.50 per million input tokens, then $1.
Microsoft's AI division released seven specialized models at its Ignite conference, all positioned as cheaper in-house replacements for the OpenAI and Anthropic systems that power its own products. The move targets the economics of always-on AI agents, where token costs accumulate continuously rather than per query.
Microsoft AI (MAI) released the full batch at once, mirroring its approach at Build in June, and framed the lineup around price-to-performance. The company cites rapidly rising token spending among customers running AI agents around the clock as the driving factor. The code models target everyday enterprise coding tasks, and Microsoft says they are not yet optimized for harder work such as foundational research or long-horizon agentic workflows. The backdrop matters: Microsoft and OpenAI restructured their partnership in April, leaving OpenAI free to sell through any cloud provider while Microsoft retains a non-exclusive license to OpenAI's models through 2032.
How close is MAI Code 2 to GPT-6.1 Sol?
By Microsoft's numbers, very close — at a fraction of the price. MAI Code 2 averages 98 percent of GPT-6.1 Sol's performance across four benchmarks at 40 percent of the cost, Microsoft says. Against Claude Opus 5.5, it reaches 96 percent of the performance at 5 percent of the cost.
Microsoft tested on DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. Three of the tests ran on Microsoft's internal MSBench platform using the GitHub Copilot CLI without sub-agents. The cost figures include a 50 percent promotional discount Microsoft applies to its own model.
In blind comparisons of frontend work — UIs, games, data visualizations, and SVG graphics — human evaluators preferred MAI Code 2's output 64 percent of the time over GPT-6 Sol, according to Microsoft.
The lighter MAI Code 2 Flash, successor to MAI-Code-1.1-Flash, handles code explanations, targeted edits, and routine fixes. Microsoft says it matches GPT-6 Luna and Claude Haiku 5.5 at 65 and 20 percent of their respective costs.
Pricing in GitHub Copilot: MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens at the promotional price through May 17, 2027, then rises to $1 and $5. Code 2 Flash costs $0.05 and $0.25 until that date, then $0.10 and $0.50.
What does the security model do — and why was the benchmark changed?
MAI Cyber 2 Flash solves 92.3 percent of tasks in the CyberGym-E2E vulnerability benchmark, Microsoft says. GPT-6 Astra reaches 91.5 percent but needs twice the reasoning budget. Per task, the MAI model costs less than 2 percent of GPT-6 Astra; per-token costs are one hundredth. On WebExploitBench, it scores 40.4 percent, ahead of GPT-5.6 Cyber at 38.4 percent.
Microsoft disclosed a data-contamination issue during evaluation. Ten percent of the roughly 900 CyberGym tasks draw on the same open-source projects and vulnerability reports as parts of Microsoft's own training data. The company says the reference solutions never entered training, but it removed those tasks for all models regardless and scored only the remainder.
A comparison with GPT-6 Cyber is missing, because that model is available only to selected customers. Microsoft says agentic benchmarks with tools require extra safety measures, with results planned for the coming weeks.
The context window spans one million tokens, four times that of MAI Cyber 1.1 Flash. Output tokens cost 86 percent less than the predecessor's. MAI Cyber models already handle nearly 40 percent of scans on Microsoft's security platform MDASH, according to the press release. Pricing is $0.10 per million input tokens and $0.50 per million output tokens, versus $10 and $50 for GPT-6 Astra — which works out to $0.45 per CyberGym-E2E task instead of $34.38. In MDASH, the model automatically replaces the existing one in the MAI profile for eligible customers starting mid-December, and it remains exclusive to Microsoft's security products.
Why does Microsoft's decision model run on Google's Gemma?
Microsoft-Decision-1.1 handles routing, classification, prioritization, and workflow control. It picks from developer-approved actions and returns calibrated confidence scores; when uncertainty is high, applications can escalate rather than act on a wrong decision, and the model abstains when information is incomplete.
Its foundation is Google's mixture-of-experts model Gemma 4 26B A4B — a notable choice given the competitive dynamics. The predecessor was built on Qwen3.5-9B. Microsoft trained the new model with supervised fine-tuning and reinforcement learning on hundreds of synthetic and licensed datasets, and runs it on a proprietary inference engine.
On the Decision Index v2.1 from Artificial Analysis, Decision 1.1 scores 68 points, six more than its predecessor and above Jev 1.14, Gemini Decision 1, and GPT-6 Luna (Decisions). Microsoft also tested it on 102 internal benchmarks with more than 385,000 questions, where it achieves the highest average accuracy among eight decision models.
The model now accepts images, including screenshots. In Copilot, it decides whether a request goes to Copilot or Cowork, performing as accurately as a large language model but 2.2 times faster and 4.4 times cheaper, Microsoft says. In Cowork, its Auto function controls how much reasoning effort a task receives. Pricing is $0.042 per million input tokens, or $0.05 in the EU data zone, with output tokens free.
A seventh model appears in the Ignite Book of News: Microsoft-Retriever-1, which gathers relevant information from enterprise documents for agents in Foundry IQ. Instead of a single search, it plans queries, evaluates results, and issues follow-ups until it has enough material for a grounded answer. It launches as a preview, and the Model Router in Foundry now selects automatically from an expanded model pool.
Where do image and speech models land?
MAI Image 2.7 leads the Artificial Analysis image-editing leaderboard with 1,215 Elo points from 5,478 votes — four points ahead of GPT Image 2.6 Sunburst (max) at one-fifth the cost. For text-to-image generation it ranks fourth. The model costs about $35 per 1,000 images; the Flash variant costs about $10, less than a third, and powers Canvas, a new image editor in MAI Labs. The model can change the color, material, or position of individual objects without altering the rest of an image, and it generates transparent backgrounds. MAI Image has been the default in Bing Image Creator since June, and Adobe, Figma Weave, Minute (Sports Illustrated), Wirestock, and Manus work with it.
MAI Realtime 1 is Microsoft's first speech-to-speech model. It handles interruptions, waits until speakers finish, and recognizes requests amid background noise. For harder tasks it falls back to a more powerful reasoning engine billed per token; the base rate is charged per audio minute. On Big Bench Audio, τ-Voice, and Dynamics 365 tests, it outperforms GPT Realtime 2.1, Microsoft says. Compatibility with the OpenAI Realtime API lets existing voice apps switch easily. Agoda and Reservations.ai are testing it, and Microsoft plans to bring it to Dynamics 365 Contact Center and Dragon Copilot, where nurses would practice patient conversations. It costs one US cent per audio minute, plus $1 per million input tokens and $5 per million output tokens for reasoning.
Dragon Copilot Transcribe, based on MAI Transcribe 2, now handles all doctor-patient transcriptions in Dragon Copilot — 10 million conversations per month, with the multilingual error rate cut in half. Microsoft also now runs its own large models on the Maia 200 chip, its second-generation AI accelerator, said to deliver 1.7 times more tokens per watt.
Can enterprises fine-tune these models themselves?
Yes, through MAI Licensed Weights. Companies can fine-tune MAI models with their own data and run them in their own Azure environment, with data and customizations staying with the customer. Microsoft says it will use them for training its own models only with explicit consent. The company positions the offering as an alternative to open weights, whose provenance and behavior are hard for enterprises to verify. One cited example: US carrier AT&T built a customized model for 5G network operations analysis that reportedly outperformed GPT-6.1 Sol on standard tasks at roughly one-fiftieth of the cost. No launch date or pricing exists yet; interested companies contact their Microsoft account team, with availability "shortly" through the Frontier Company program.
Availability varies by model: MAI Code 2 and Code 2 Flash are in GitHub Copilot, with Code 2 Flash also in limited preview in Foundry; OpenRouter, Vercel, Hermes, Open Code, and Kilo are expected to follow. Image 2.7 and Flash are in public preview in Foundry and on fal, Genspark, OpenRouter, and Vercel. Decision 1.1 is generally available in Foundry, on OpenRouter, and through Vercel's AI Gateway. MAI Realtime 1 can be tried in MAI Labs and as a free Foundry preview until public preview launches. With token spending climbing and its OpenAI license running to 2032, Microsoft now has both the incentive and the model lineup to shift its own products onto in-house systems.
Original: microsoft.ai
More from Marcus Bennett
Show full bio
Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
745 articles
Related articles
- OpenAI launches GPT-6.2 Astra and Sol with 20 percent lower API prices
- GPT-6.1 Terra halves API prices and closes the gap on OpenAI's biggest models
- Microsoft's MAI Code 2 hits 96% of Claude Opus 5.5 at 5% of the cost
- Microsoft AI Launches MAI Code 2 and Flash on Cost-per-Task Math
- OpenAI Launches GPT-6.1 Sol at One-Fifth of Astra's Price