Meta has entered the crowded market for terminal-based AI coding agents, launching Muse Code in beta on August 5, 2026, powered by its updated Muse Spark 1.2 model. The tool is available through the Meta Model API and installs via a single terminal command, positioning it as a direct rival to Anthropic's Claude Code and OpenAI's Codex.
Muse Code's headline feature is a crash-safe runtime: the agent coordinates persistent background subagents and maintains a replay-exact event log, so that, according to Meta, "after a crash, the agent can resume precisely where it stopped." The tool ships with default commands including "/plan" for breaking down tasks, "/grill" for stress-testing code, and "/goal" for autonomous objective completion, and it also demonstrates multimodal capability, converting video inputs like a home flythrough into a functional web application with a booking interface.
Benchmarks Show Meta Trailing, Not Leading
On Terminal-Bench 2.1, Muse Spark 1.2 scored 82.9%, behind Claude Code running on Anthropic's Opus 5 at 86.7%, but ahead of OpenAI's GPT-5.6 Terra on Codex at 81.8% and Grok Build at 81.6%. On DeepSWE 1.1, a benchmark focused on agentic coding capability, the gap narrowed: Muse scored 59.3% against 65.0% for Opus 5 and 64.8% for Codex. On Meta's own internal benchmark, Muse scored 70.6% versus Opus 5's 79.4%. Across more than 1,000 tool calls, Muse Spark 1.2 showed roughly a 61–69% speedup in sustained performance, placing it in the middle of the competitive field rather than at the front.
In effect, Meta's own published numbers show Muse Code trailing Claude Code on every headline metric it chose to release, while edging out Codex and Grok Build on most of them.
Price as the Real Differentiator
Rather than compete purely on raw capability, Meta has leaned into affordability. Meta AI chief Alexandr Wang said on X that Muse Code is one of the most affordable coding agents on the market and is available globally. Pricing for the standard tier is $1.25 per million input tokens and $4.25 per million output tokens, with a discounted contributor tier at $0.10 and $0.20 respectively for developers willing to share their usage data with Meta, an option that trades cost savings for reduced privacy.
What Meta Is Actually Betting On
Meta's own framing leans on the crash-safe runtime and subagent architecture as the differentiators that matter most for day-to-day developer workflows, rather than outright benchmark supremacy. Whether developers weigh Muse Code's lower cost and resilience against its clear benchmark gap versus Claude Code will likely determine how much traction the tool gains against two competitors with a considerable head start in the agentic coding space.
