Meta shipped a new coding agent yesterday, and it’s aimed straight at Cursor, Codex, and Claude Code. Muse Code runs in the terminal, installs with one command, and keeps working in the background while you do something else. The pitch is price. The catch is your data.
What Meta’s coding agent actually does
Muse Code is the first coding agent from Meta Superintelligence Labs, the AI division that chief AI officer Alexandr Wang runs. Meta announced it August 5 alongside Muse Spark 1.2, the coding-focused update to its frontier model family. The tool is a terminal harness for macOS and Linux, installable with a single curl command.
The headline feature is what Meta calls async background agents. Instead of spawning helper agents fresh for each task, the coding agent keeps a set of specialized agents alive for the whole session, which Meta says avoids redundant information gathering. When a job gets big, Muse Code fans out to parallel sub-agents, each in its own isolated git worktree, so your working copy never gets touched.
There’s also an audit story. Every model call, tool run, approval, and edit gets appended to a local event log before it executes. Meta calls the runtime replay-exact and restart-safe, which is the kind of feature that wins enterprise evaluations.
The terminal choice matters more than it looks. Muse Code isn’t a plugin bolted onto an editor, it’s a command-line agent that slots into the workflow you already have. That lowers the switching cost, and it’s the same reason developers adopted Claude Code and the Codex CLI instead of waiting for their IDE vendors to catch up. If the tool is one curl command away, trying it costs almost nothing.
The price is the product
Here’s where it gets interesting. Meta is pricing Muse Code in two tiers, and the gap between them is the whole strategy.
The standard tier runs $1.25 per million input tokens and $4.25 per million output tokens, and Meta commits that prompts and completions on that tier aren’t used to train its models. That’s mid-pack pricing, roughly in line with the rest of the field.
The contributor tier is the shocker. It costs $0.10 per million input tokens and $0.20 per million output tokens, roughly 12x and 21x cheaper than standard, with cached input at nearly zero. In exchange, you opt in to let Meta use your prompts and completions to train future models. Wang says the contributor tier is “more than 10 times cheaper” than pay-as-you-go, and it’s the cheapest coding agent pricing on the market.
That trade is the whole play. The developers who save the most are the ones feeding Meta the training data it needs to close the gap with Claude and GPT-class models. For side projects and non-sensitive code, that’s a great deal. For teams with proprietary code, the cheap tier isn’t cheap at all.
The numbers behind the launch
Meta published benchmark charts comparing Muse Spark 1.2 against rivals, and they show a strong but clear second place. On Terminal-Bench 2.1, Muse Code scored 82.9%, edging OpenAI’s GPT-5.6 Terra in Codex at 81.8%, but trailing Anthropic’s Opus 5 in Claude Code at 86.7%. On DeepSWE 1.1, it posted 59.3%, third behind Opus 5 and GPT-5.6 Terra.
The generational gains are real. Muse Spark 1.2 improves on 1.1 by 6.7 points on Terminal-Bench and 6.3 on DeepSWE. Meta also co-trained the model with Muse Code itself, using rejection-sampled harness trajectories, plus a self-improvement loop where 1.1 generated coding environments and graded candidate solutions.
The demo case is a GPU kernel optimization that ran for more than 1,000 tool calls over up to 24 hours on Hopper hardware, with Zuckerberg saying it kept finding improvements well beyond the initial exploration phase.
My take on the data trade
This is Meta’s most serious entry into a category it mostly watched from the sidelines. The Llama era gave away open weights for mindshare. Muse Code gives away cheap tokens for training data, and that’s a strategy with real teeth.
The timing is brutal for rivals. Cursor is reportedly raising at a $50bn valuation, and GitHub recently froze new Copilot sign-ups as agentic usage blew through its pricing. Meta’s contributor tier lands on that sore spot: autonomous agents burn enormous compute, so whoever can serve them cheapest wins the volume.
I covered the best AI coding agents of 2026 back in July, and the pricing pressure was already visible. OpenAI’s price cuts on GPT-5.6 were the first shoe. Meta just dropped the second one.
The open question is trust. Muse Code is proprietary, with no downloadable weights, a hard turn from the Llama playbook. Zuckerberg teased that open source might be in the cards, replying “I’ll have more to share on that soon” when asked directly. Which is not a yes.
The data trade makes sense for hobbyists and startups. For anyone working on sensitive code, the contributor tier is a cost shift, not a discount. Read the terms before you take the deal. The coding agent market just got a lot more interesting, and a lot more personal.
The bigger story is where this lands in Meta’s arc. The company spent years giving away open weights to win developer mindshare, then watched the agentic era turn into a paid-token business. Muse Code is Meta admitting that the fastest path back into developer tooling is a price war, funded by the data you hand over. That’s a bet on volume over trust, and it’s going to force every other harness vendor to answer the same question: what is my data actually worth to the model maker?
