Microsoft to Release New Coding Model Next Week in Comeback Attempt
Microsoft next week will unveil a suite of new homegrown AI models at its annual Build conference for app developers in San Francisco. The question is whether it can win over these people.
The company plans to unveil a coding model to boost the competitiveness of Microsoft-owned GitHub Copilot, a coding assistant whose early lead in the AI coding market was eroded by Cursor and Claude Code, according to someone with direct knowledge of the plans.
Microsoft also is planning to announce new models that come in various sizes and specialize in tasks like transcription, thinking (i.e. reasoning), speech and images, the person said. The family of models will build on the earlier homegrown models that Microsoft previewed earlier this year. (Speech and transcription are becoming a big deal among developers, as Stephanie described in vivid detail in Wednesday’s edition.)
A Microsoft spokesperson declined to comment. Microsoft AI chief Mustafa Suleyman, in his excitement, previewed one of the forthcoming image models focused on image generation on Tuesday.
It’s an important test for Microsoft’s in-house AI capabilities. Models launched by the team under Suleyman haven’t topped AI leaderboards since he joined the company two years ago; Suleyman has said he was restricted in training cutting-edge models by Microsoft’s deal with OpenAI, which prevented Microsoft from pursuing human-level AI before April, when the companies renegotiated the matter.
So far, Microsoft has primarily relied on models from OpenAI (which it gets for free through 2032) and Anthropic (which it pays for) to power its GitHub Copilot as well as its 365 Copilot software, which automates tasks within Office apps.
When Microsoft loses access to OpenAI’s technology in six years or so, it wants to have legit AI of its own so it isn’t beholden to pricey suppliers.
Microsoft will market its new models as cheaper than the slightly more capable ones from OpenAI and Anthropic, the person said. That mirrors how Microsoft has positioned some of its earlier homegrown models, and is similar to Google’s strategy. Of course, we won’t be able to take Microsoft’s word about the new AI capabilities until developers confirm them.
Microsoft has said it is using some homegrown models to power AI features in Copilot software for Office apps, driving down the cost of running the features. The new models are expected to also help in that regard, said the person with knowledge of the launch.
The company also wants developers to buy the new models through its Azure cloud the same way they buy access to OpenAI and Anthropic models.
Relying on models from Anthropic has prompted Microsoft to raise the prices of some products, such as GitHub Copilot, and cap how much developers can use Anthropic models through that product. If Microsoft’s in-house coding model can serve as a cheaper alternative to Claude, it could help Microsoft win back market share among cost-conscious developer customers.