Microsoft Releases Seven MAI Models: Coding Model Delivers 98% of GPT-6.1 Sol's Quality at 40% of the Price
Microsoft AI, the company's internal MAI lab, has released seven models designed to let agents run around the clock without huge token bills. According to the company, its coding model delivers 98% of GPT-6.1 Sol's quality on development tasks at roughly 40% of the cost; against Opus 5.5, it reaches 96% of the quality for 5% of the price. It's worth testing these claims on your own workloads: routine agent steps can go to the cheaper model, while the flagship is reserved for complex steps.
AI-processed from Hacker News AI; edited by Hamidun News
An agent that works not for five minutes but around the clock quickly turns the token bill into the biggest line item in the budget. Microsoft AI, the company's internal lab known as MAI, opened its announcement with this problem: it released seven models at once and promises leader-level quality for a small fraction of the price. The most notable figure concerns code. Microsoft says its coding model reaches 98% of GPT-6.1 Sol's quality (meaning it solves nearly as many engineering tasks as GPT-6.1 Sol does) and costs about 40% of its price. All numbers in this article are Microsoft's own data; there is no independent verification yet.
Code, Cybersecurity, and Mid-Tier Models
Comparing the coding model with Anthropic's Opus 5.5, it reaches 96% of that model's results but costs only 5% of the price. For a team where an agent fixes bugs and writes tests around the clock, the difference between 100% and 96% of quality is barely noticeable, while a 20x difference in price is very noticeable.
A separate cybersecurity model was tested on the CyberGym-E2E benchmark, where an agent completes a task end to end, from finding a vulnerability to delivering the result. According to Microsoft, it surpasses GPT-6 Astra in quality, and a single task costs less than 2% of what it costs with the competitor. If you automate vulnerability search or incident analysis, a price like that turns an expensive experiment into an ordinary line item.
The general-purpose model is positioned more modestly. It shows quality comparable to GPT-6 Luna at 65% of its price, and quality comparable to Haiku 5.5 at 20% of the price. It is a candidate for the workhorse role on high-volume requests where a flagship is overkill.
Another model in the family is for decision-making. Microsoft says it is 50 times faster and 79 times cheaper than GPT-6.1 Sol. Models like this are needed when an agent must quickly choose its next step: which tool to call, where to route a request, whether human review is needed. The company does not specify which tasks the comparison was measured on, so it is more reliable to rely on your own data.
Images, Voice, and Speech Transcription
The image editing model ranked first on the Artificial Analysis leaderboard and outperformed GPT Image 2.6 Sunburst in its maximum mode, while costing five times less. Editing is the most common scenario in products: removing a background, fixing a product photo, replacing a detail.
According to Microsoft, its real-time voice model is better than GPT-Realtime 2.1 and starts at $0.01 per minute of audio. For call centers and voice assistants, the price per minute determines whether the service pays off with thousands of parallel calls.
The most down-to-earth example is MAI Transcribe 2. Microsoft calls it the fastest and most accurate multilingual speech transcription model. It is already running across 10 million patient visits per month, and its error rate across different languages has been cut in half. This is not a lab test but production use in medicine, where an error is costly.
Its Own Chip and What Developers Should Do
The seventh item in the announcement concerns hardware. Microsoft's second-generation accelerator is already running in production and serving the company's largest models. It delivers 1.7x more tokens per watt. Energy efficiency largely determines what price per token Microsoft can afford to set.
What should practitioners do? First, don't take the percentages from the press release at face value: quality is measured on tests chosen by the vendor, and "price per task" and "price per token" are different metrics. Second, assemble your own set of 50 to 100 real tasks and run the new models alongside your current ones. Third, if your agent has many repetitive steps, revisit the architecture: a cheap, fast model handles routine work and selects actions, while the expensive flagship is brought in only for the hard parts.
There is also a strategic storyline. Microsoft, which long relied on OpenAI, is now selling its own models and openly comparing them with GPT and Claude on price. If independent tests confirm the claims, pressure on competitors' pricing will grow, and choosing a model on the principle of "the strongest" will become less economical.
- Nov 21, 2026Islo привлекла $15 млн от Insight Partners на «фабрику ПО» с ИИ-агентами
- Nov 21, 2026Applied Digital Asks Oracle and Meta for Billions in Power Guarantees
- Nov 21, 2026GitHub Copilot now lets teams share plugins by default
- Nov 21, 2026Microsoft Releases TokenEconomizer: Coding Agents Spend 18.4–30.9% Less Money
Key Questions Answered
What are the key details on Code, Cybersecurity, and Mid-Tier Models?
Comparing the coding model with Anthropic's Opus 5.5, it reaches 96% of that model's results but costs only 5% of the price. For a team where an agent fixes bugs and writes tests around the clock, the difference between 100% and 96% of quality is barely noticeable, while a 20x difference in price is very noticeable. A separate cybersecurity model was tested on the CyberGym-E2E benchmark, where an agent completes a task end to end, from finding a vulnerability to delivering the result.
What are the key details on Images, Voice, and Speech Transcription?
The image editing model ranked first on the Artificial Analysis leaderboard and outperformed GPT Image 2.6 Sunburst in its maximum mode, while costing five times less. Editing is the most common scenario in products: removing a background, fixing a product photo, replacing a detail. According to Microsoft, its real-time voice model is better than GPT-Realtime 2.1 and starts at $0.01 per minute of audio.
What are the key details on Its Own Chip and What Developers Should Do?
The seventh item in the announcement concerns hardware. Microsoft's second-generation accelerator is already running in production and serving the company's largest models. It delivers 1.7x more tokens per watt. Energy efficiency largely determines what price per token Microsoft can afford to set. What should practitioners do? First, don't take the percentages from the press release at face value: quality is measured on tests chosen by the vendor, and "price per task" and "price per token" are different metrics.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.