Meta taps rival GPT-4 to power its internal coding tool

Mark Zuckerberg has spent the last year convincing the world that Meta's open Llama models can go toe-to-toe with anything coming out of Silicon Valley's closed labs.

Inside Meta's own offices, however, their engineers have quietly been relying on the competition to get their daily coding done.

According to a report from Fortune, Meta's internal AI coding assistant, known as Metamate, has been pulling answers from OpenAI's GPT-4 alongside Meta's own Llama models since at least early this year.

A two-engine setup under the hood

Metamate, which was originally developed under the name Code Compose, functions much like GitHub Copilot or Anthropic's Claude inside Meta's developer stack.

Two current and former Meta employees confirmed that the tool routes queries differently depending on the prompt. If Llama is not the ideal fit for a certain coding problem, the system taps OpenAI's GPT-4 instead.

Employees genuinely like the tool, calling it one of the better generative AI implementations Meta has built. But they also keep expectations measured, describing its current capability as roughly on par with a competent intern who handles routine coding tasks rather than high-level architecture.

It is not just Meta's engineering wing opening its wallet to OpenAI, either. The Chan Zuckerberg Initiative, the philanthropic outfit led by Zuckerberg and Priscilla Chan, is also a paying OpenAI customer building educational wrappers on ChatGPT. Sam Altman even joined the initiative's AI advisory board earlier this year.

Meta declined to comment on the arrangement.

The productivity catch-22

The reliance on GPT-4 is telling because Meta's latest Llama release was trained on massive datasets, including four times more code than its predecessor. Zuckerberg himself claimed mid-year that Llama would be the most advanced model in the industry by next year.

Yet on the internal front lines, the raw horsepower of closed frontier models still appears necessary to patch the gaps.

Meanwhile, the tool's effectiveness has introduced a familiar anxiety inside Meta. One source pointed out that as Metamate handles more grunt work, leadership increasingly concludes that fewer humans are needed on the payroll, serving as a subtle driver for the ongoing rolling job cuts this year.

Why it matters

Catching Meta using OpenAI on its home turf sounds like a PR headache, but we think it is mostly a masterclass in enterprise pragmatism.

Publicly, tech leaders have to beat their chests about model supremacy. Privately, running an engineering org means prioritizing developer speed over ideological purity. If GPT-4 still handles edge cases better than Llama, waiting for the next training run to fix your internal developer friction makes zero business sense.

For enterprise teams, this validates a multi-model routing strategy. Sticking to a single family of models out of brand loyalty is a trap. Routing easy prompts to open-weight models and sending thorny reasoning to frontier proprietary APIs is precisely what the smartest companies are doing, even if they run their own model labs.

The darker takeaway is the intern comparison. When an internal tool can reliably perform the work of an entry-level software engineer, the pipeline for training junior developers starts to dry up, and executive appetites for head-count reductions only grow sharper.