Microsoft Strikes Deals With OpenAI’s Top Rivals For AI Coding Assistant
Microsoft is turning up the heat on OpenAI. Microsoft’s developer platform GitHub announced on Tuesday that its wildly popular GitHub Copilot tool—which previously relied entirely on OpenAI models—will now also incorporate AI models from Anthropic and Google.
GitHub Copilot will be able to toggle between Anthropic’s Claude 3.5 Sonnet, Google’s Gemini 1.5 Pro, and OpenAI’s GPT-4o, o1-preview, and o1-mini, the company said. The new development is striking given Microsoft’s long record of relying primarily on OpenAI models for its AI products.
Microsoft’s investments in OpenAI grant it the rights to reuse the startup’s models in its own software, and GitHub Copilot became the earliest Microsoft product built on OpenAI models when it first launched widely in June 2022, months before OpenAI shipped ChatGPT.
That partnership helped OpenAI grab early mindshare with software developers, and was also a boon to Microsoft, which consistently pointed to GitHub Copilot as an early example of revenue coming from a generative AI product (the company implied in July that GitHub Copilot is on track to generate $300 million annually from subscriptions). And because Microsoft already had the rights to use OpenAI’s models, it didn’t have to pay sticker price for the AI underlying the tool, which theoretically helped GitHub’s margins on selling the $10-per-seat software.
Today’s announcement shows just how much the AI landscape has changed since GitHub Copilot’s launch. When I spoke to GitHub CEO Thomas Dohmke about GitHub’s decision to start paying Anthropic and Google to include their models in its software, he said the move stemmed from developer demand for those models.
“Developers may have personal preference, or they may have seen a benchmark or an article on The Information about a new model that they want to use instead,” Dohmke said. “We’re not saying one model is better than the other. We believe that’s for developers to decide.”
Software developers and founders that I’ve spoken to see the move as a no-brainer given Anthropic’s recent advances in the code-writing space. As my colleagues Erin, Steph and Amir reported earlier this month, Anthropic has taken the lead over OpenAI in the race to develop artificial intelligence that can automatically write code or help developers improve their software.
Models made by Anthropic are beating OpenAI at some code-writing tasks, including in internal tests that OpenAI has been conducting.
“Claude 3.5 has been out for two quarters now and it’s just measurably, meaningfully better at reasoning and coding than [OpenAI’s] GPT-4o,” said Siqi Chen, founder and CEO of the enterprise software startup Runway. Chen said his company’s developers use GitHub Copilot but have simultaneously kept subscriptions to Anthropic and to other code-writing AI tools like Cursor and Augment to test which one performs best.
GitHub Copilot’s pricing will remain the same for now, at $10 per seat for individuals and enterprise tiers at $19 and $39, depending on how many features companies want to pay for. When I asked Dohmke whether paying for Google or Anthropic models will affect GitHub’s margins, he wouldn’t comment directly, but said GitHub has a “cost model for each individual model” and aims to “set the right price to achieve the best possible outcome for our customers.”
Meanwhile, GitHub also announced today that it’s rolling out another product that will compete more directly with OpenAI: a new application called Spark, which will use generative AI to create entire applications based on short prompts that users write in natural language.
For instance, a user could write, “make a budgeting tool to track my travel expenses” or “build an online store for my pottery business” or something similar, and Spark would automatically code the applications and host them online (GitHub is initially releasing Spark in a limited preview and will announce pricing details later).
My colleagues reported this month that OpenAI is working on a similar product that could use AI to automate such coding tasks. GitHub’s Spark, like Copilot, won’t rely entirely on OpenAI’s models: users will be able to decide between OpenAI or Anthropic models when they use the tool.