Driving the news: Microsoft is announcing six new AI models at its Ignite conference in San Francisco on Tuesday. The templates cover coding, cybersecurity, image generation, and voice.
- AMI code 2: Microsoft claims its new coding model delivers 98% of the performance of GPT-6.1 Sol at 40% of the price and 96% of the performance of Anthropic’s Claude Opus 5.5 at just 5% of the cost.
- MAY 2 Flash code: A smaller, faster model that Microsoft says matches the GPT-6 Luna at 65% of the cost and the Claude Haiku 5.5 at 20% of the cost.
- MAY Cyber 2 Flash: A cybersecurity model that Microsoft says beats GPT-6 Astra on a security benchmark at less than 2% cost per task.
- Other models: MAI Image 2.7 ranks first in image editing, according to Microsoft, while new real-time transcription and voice models aim to challenge its main competitors.
Between the lines: Microsoft’s strategy is to create specialized models for particular tasks, allowing them to achieve comparable performance at significantly lower costs.
“It basically means you pack more tokens, knowledge and expertise into the same number of settings,” Suleyman told Axios.
- The tradeoff is that a model optimized for everyday coding tasks won’t perform as well in advanced math or scientific research.
- Microsoft performance comparisons are based on selected benchmarks, and some pricing benefits reflect promotional discounts.
The big picture: Even as Microsoft targets industry leaders, cutting-edge AI also continues to advance. OpenAI announced GPT‑6.2 Astra and GPT‑6.2 Sol on Tuesday, with the new Astra priced 20% lower than the previous version.
- Microsoft is also looking for a broad, versatile language model that can compete with a wider range of tasks.
- Suleyman said Microsoft had achieved industry-leading pre-training results and was making big strides in reinforcement learning, but declined to give a timeline for releasing the model.
- “My team is building real superintelligence,” he said.
The plot: The new models could reshape Microsoft’s relationships with OpenAI and other external AI vendors.
- MAI Code 2 will be available in GitHub Copilot, which could potentially shift some coding workloads away from external models.
- When asked how much Copilot traffic Microsoft hoped to shift to its own models, Suleyman declined to offer a target but predicted significant adoption.
Zoom: Microsoft also offers enterprise customers access to licensed model weights, allowing them to customize its models with proprietary data while requiring deployment in Azure.
- Suleyman said the approach would “feel like open source,” while providing greater security and control.
What to watch: How quickly can Microsoft translate its advances in specialized AI into a competitive general-purpose model – and whether its cost advantages are enough to persuade customers to abandon their established competitors.
Gn bussni
