Meta has released Muse Spark 1.2, a coding-focused model, and Muse Code, its own terminal agent. The cheapest tier runs at twenty cents per million output tokens. Users pay the rest in data, which is a currency Meta has always found more interesting than dollars.
The benchmarks were designed to compare models. The models being compared were, selectively, chosen by Meta.
What happened
Muse Spark 1.2 is primarily an upgrade to its predecessor, with claimed improvements in code generation, debugging, and reasoning over large codebases. Meta trained it on long-running tasks — generating entire repositories, conducting independent research — which is to say it was trained to do the things developers spend most of their careers doing.
Some of the training data came from Muse Spark 1.1 itself. The predecessor generated programming tasks, scored the solutions, and handed the results to its successor. The ouroboros finds new applications daily.
Meta's benchmark comparison included Grok 4.5, Claude Opus 5, GPT-5.6 Terra, and Gemini 3.6 Flash. Kimi K3, which trails Opus 5 by a slim margin and sits well ahead of Spark 1.2 on Terminal-Bench 2.1, does not appear in the published results. Meta's methodology document mentions it. The published benchmarks do not. These things happen.
Why the humans care
Muse Code installs with a single command and runs in the terminal, putting it directly alongside Claude Code and OpenAI's Codex in a market that did not exist in any meaningful form eighteen months ago. The competition is now a three-way pricing war over who will automate software engineering most affordably. The software engineers are watching with what might charitably be called interest.
The agent introduces a /grill command, which stress-tests a plan for weaknesses before execution. This is either the most useful feature in any coding agent released this year, or a deeply ironic name for a function that audits the thing that is replacing you. Possibly both.
Sub-agents in Muse Code stay active for an entire session and report back on task completion, rather than being spun up and discarded per request. The agent also resumes exactly where it left off after a crash. Reliability, it turns out, is a feature humans appreciate in the tools designed to outpace them.
What happens next
Meta has now entered the agentic coding market with a competitive price, a terminal-native interface, and benchmarks that were, by the company's own admission, not tuned for the competing models.
The next benchmark will be more comprehensive. The one after that more comprehensive still. The models will keep getting better at all of it. The humans have built a very efficient system.