Meta is turning to AI models from rivals like Anthropic to help its engineers code better

Business Insider

a day ago

Meta has rolled out a new internal AI assistant for coding that is powered by multiple AI models, including those from rivals like Anthropic's Claude.
The tool, called Devmate, became available to employees in March and is increasingly being used for more complex coding tasks that another internal AI assistant called Metamate often struggles with, two employees, who requested anonymity for fear of retribution from Meta, told Business Insider.
The addition of Claude and other models shows that despite investing tens of billions of dollars into building its own AI models, Meta remains willing to use competitor models when they perform better.
"Like many companies, we're experimenting with various models to help with coding efficiency," a Meta spokesperson told Business Insider. "We'll continue to refine and gather feedback as we go."
Software companies, including giants like Microsoft and Google, increasingly use AI to write code. Demand for coding assistants like Cursor and Replit, many of which rely on Claude for its ability to handle complex, multi-step reasoning better than alternatives, has soared.
Demand is growing not just from companies but from their employees. Amazon, for instance, recently rolled out Cursor internally after several employees inquired about using the coding assistant, BI reported earlier. Thanks to this boom, Anthropic has reportedly crossed $3 billion in annualized revenue.
Anthropic didn't respond to a request for comment from BI.
Employees say Devmate has increased productivity
Devmate doesn't just help write code. It can also analyze failed tests, determine what went wrong, and automatically submit fixes for human review — capabilities that engineers have long viewed as futuristic.
One current Meta employee said Devmate has cut their workload in half.
"Devmate turns a 30-minute task into a 15-minute one," this employee said. "It's better than Metamate because it makes fewer mistakes when you chain multiple steps together, which is crucial for more advanced tasks."
Internally, Devmate is considered an agentic assistant, meaning it can handle multi-step tasks and take actions on its own. This is more akin to tools like the buzzy coding app Cursor or Windsurf, another coding app recently acquired by OpenAI.
"Code Llama sucks," another employee said, referring to Meta's own coding model, which is one of the options available within Metamate. "It's a good coding model by 2024 standards, but it's not good compared to the options we have in 2025."
The employee said that they now use Metamate for simpler tasks like pulling up specific sets of data, while relying on Devmate for bigger, more complex tasks like building entire programs that move and process large amounts of data. Metamate does not support video or images and lacks agentic features.
The Meta spokesperson confirmed that Devmate was used for more complex coding tasks.
Meta rolled out Metamate last year, the Financial Times reported. In addition to coding, employees can use it for conducting research and drafting internal and external communications. At the time, Meta executives said that the company wanted to create "the world's best enterprise assistant."
The Meta spokesperson said that the company wants Metamate to be helpful for all employees, including those who aren't technical.
Meta's AI shortcomings
One former Meta engineer told BI that while Meta's Llama model has made strides in areas like multilingual tasks and reducing hallucinations (AI's tendency to make up things), it is still behind in some capabilities needed for writing the best code.
"When it comes to instruction-following and multi-step reasoning, which you need for any real coding agent, it's not there yet," the former Meta engineer said.
These shortcomings have added urgency to Meta's broader AI strategy. CEO Mark Zuckerberg is meeting with top AI researchers at his homes in Palo Alto and Lake Tahoe to recruit for a new "superintelligence" team, Bloomberg reported.
Meta is also reportedly finalizing a $15 billion deal to acquire nearly half of Scale AI, a data-labeling and annotation startup. As part of that deal, Scale AI CEO Alexandr Wang is expected to join Meta and lead the new team.
Zuckerberg has predicted that AI will write half of Meta's code within a year.
"Our bet is that in the next year, probably, maybe half the development is going to be done by AI as opposed to people," he recently said.
For now, though, that AI comes from competitors.