Meta introduced its own AI coding agent, Muse Code, on August 5, 2026, competing directly against Anthropic’s Claude Code and OpenAI’s Codex. The terminal tool is based on the new Muse Spark 1.2 model and costs around one-tenth of the standard rate at its cheapest tier.
Terminal agent coordinates multiple subagents in parallel
Muse Code (beta) runs in the terminal on macOS and Linux and can be installed with a single command. For larger tasks, the tool starts multiple persistent subagents at once, which write code in isolated working copies in parallel without interfering with each other. A local event log makes sessions repeatable and lets work resume precisely after an interruption. Built-in commands such as /plan, /grill, and /goal break larger programming tasks into individual steps.
On the Terminal-Bench 2.1 test, the underlying Muse Spark 1.2 model achieves a pass rate of 82.9 percent, just ahead of OpenAI’s GPT-5.6 Terra but behind Anthropic’s Claude Opus 5. Meta additionally promotes the model as the top performer on the DeepSWE 1.1 programming benchmark. Independent evaluations such as InfoWorld’s, however, put it at only 59.3 percent – third place behind Opus 5 at 65.0 and GPT-5.6 Terra at 64.8 percent.
Two pricing tiers bet on cost over top scores
Muse Code costs $1.25 per million input tokens and $4.25 per million output tokens at the standard tier – exactly the rates of the predecessor model Muse Spark 1.1, with which Meta already launched a price war against the competition in July. Anyone who additionally agrees to let Meta use their prompts and code responses for future model training pays only $0.10 or $0.20 per million tokens in the so-called Contributor tier. Meta AI chief Alexandr Wang put that discount at more than tenfold, as reported by CNBC, and described Muse Code as globally available; Meta gives no specifics on restrictions for Germany or the EU.
With Muse Code, Meta positions itself for the first time with a standalone coding agent against established rivals. Until now, the company offered AI support for programming tasks only through its model API, while Anthropic’s Claude Code and OpenAI’s Codex have already been in use as dedicated terminal tools.
Analysts doubt the benchmarks are comparable
Analysts voice doubts about the significance of such benchmark comparisons. According to Omdia analyst Lian Jye Su, OpenAI and Anthropic have long treated technical tuning of the agent environment as part of training itself, which blunts the value of pure model comparisons. Neil Shah of Counterpoint Research also notes that vendors test their models with their own optimized agent rather than a shared tool – a methodological gap that skews industry-wide rankings.
The launch also falls in the same week Meta had to admit a security incident involving the predecessor model Muse Spark 1.1: a misconfigured test environment gave the model unauthorized access to another company’s network. Meta said it is investigating the incident but drew no direct connection to the Muse Code announcement.
What matters now is whether companies trust a still-young coding agent with production code or first confine it to low-risk, narrowly scoped tasks – security clearances and access to CI/CD environments remain the biggest hurdle to putting such agents into real production use.
