Meta just shipped a coding agent that beats Grok and Gemini
&imwidth=600&imheight=450&format=webp&quality=medium "Meta just shipped a coding agent that beats Grok and Gemini")
Meta just shipped a coding agent that beats Grok and Gemini Photograph: (Pixabay)
Meta has released Muse Code, a terminal-based coding agent powered by its new Muse Spark 1.2 model, scoring 59 per cent on the DeepSWE 1.1 benchmark and outperforming Grok Build 4.5 and Gemini 3.6 Flash. It arrives weeks after Meta charged for one of its models for the first time in company history.
Meta has released a coding agent that runs in a developer's terminal, and on at least one benchmark it beats the equivalent tools from xAI and Google.
What Was Released
Muse Code is a terminal coding agent — software that sits in a developer's command line and executes programming tasks rather than returning suggestions to be copied into an editor. It is powered by Muse Spark 1.2, the model Meta released on August 5.
On DeepSWE 1.1, a benchmark measuring software engineering task completion, Muse Code scored 59 per cent. That places it ahead of Grok Build 4.5 from xAI and Google's Gemini 3.6 Flash on the same measure.
The terminal is a deliberate choice of venue. It is where the current competition among AI coding tools is concentrated, because an agent operating in the command line can run tests, read errors, modify files and iterate — completing work rather than assisting with it.
The Strategic Position
Trending Stories
&im=FitAndFill=(700,400))
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
)
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
The release is notable for how recently Meta entered this market at all.
For a decade, Meta's AI strategy was to give models away — open weights, open research, commoditising rivals' advantages while monetising through its own apps. That ended on July 9, when Meta launched Muse Spark 1.1 on the new Meta Model API at $1.25 per million input tokens and $4.25 per million output tokens, the first time the company had ever charged for one of its own models.
Muse Code extends that shift from selling model access to shipping a finished developer product. It is Meta competing directly for the developer relationships that OpenAI and Anthropic have spent two years building.
Why Coding Specifically
Coding has become the most commercially valuable application of large language models, and the most competitive.
The reason is measurability. A company can quantify what an engineering team costs and estimate what an AI tool saves — a clarity absent from most enterprise AI deployments. That has made coding tools the fastest-converting AI product category, and every major laboratory now fields one.
The competition has been fierce this quarter. Anthropic released Claude Opus 5 on July 24. Alibaba claims its Qwen3.8-Max ran autonomously on a software project for more than 16 days, building its own coding harness from an empty repository. Moonshot AI's Kimi K3 topped coding leaderboards before releasing its full weights for free download on July 27. OpenAI cut prices on its lower-cost models by up to 80 per cent three days after that.
The Caveat Worth Keeping
The DeepSWE 1.1 figure is a benchmark result, and benchmark results have had a difficult year.
Meta's own former chief AI scientist, Yann LeCun, confirmed in January that Llama 4's headline results had been assembled by training multiple checkpoints and selecting the best score on each individual benchmark, publishing a composite no single model actually achieved. Meta's newer Muse Spark 1.1 has an unresolved discrepancy of its own: the company reported 80.0 on Terminal-Bench 2.1 while independent evaluator Vals AI measured 69.29 on the same test, and researchers found Meta's harness exceeded the benchmark's own per-task resource limits.
None of that means the Muse Code result is wrong. It means the number is a claim awaiting independent verification rather than a settled fact — which is the appropriate way to read any vendor-reported benchmark, and particularly this vendor's.
