Meta launches Muse Code terminal agent and Muse Spark 1.2
Meta shipped two things at once: Muse Code, a terminal-based AI coding agent now in beta, and Muse Spark 1.2, a coding-focused model update that powers it. The combination puts Meta in direct competition with Anthropic’s Claude Code, OpenAI’s Codex, and Cursor.
What Muse Code Does
Muse Code installs on macOS or Linux with a single curl command and handles complete software engineering tasks across large repositories: planning changes, writing code, and validating results. Three bundled slash commands cover the core workflow: /plan creates an approval-gated plan, /grill stress-tests that plan, and /goal drives the agent toward a stated objective.
The architectural differentiator is persistent background agents. Rather than spawning fresh sub-agents per task, Muse Code keeps specialized agents alive for the entire session, avoiding redundant repository exploration. For large jobs, it fans out to parallel sub-agents each working in an isolated git worktree, so the developer’s working copy is never touched. Zuckerberg claimed it built six game features simultaneously with no collisions in testing.
Every model call, tool run, approval, and edit is appended to a local event log before it executes. If a 20-hour run crashes, the agent resumes exactly where it stopped.
Benchmark Position
Muse Spark 1.2 scores 82.9% on Terminal-Bench 2.1 running inside Muse Code, ahead of GPT-5.6 Terra in Codex (81.8%) and Grok 4.5 in Grok Build (81.6%), but behind Anthropic’s Opus 5 at max effort in Claude Code (86.7%). On DeepSWE 1.1 it posts 59.3%, third behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). On Meta’s own internal coding benchmark, Muse Spark 1.2 scores 70.6% versus Opus 5’s 79.4%. Claude leads all three charts. The gains over 1.1 are real: +6.7 points on Terminal-Bench and +6.3 on DeepSWE, though the 1.1 scores used a different harness, so some of that jump belongs to Muse Code rather than the model alone.
Pricing: the data tradeoff
The Meta Model API offers two tiers. The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens (cached input: $0.15). Prompts and completions on this tier are not used to train Meta’s models. Rate limits run to 3,000 requests and 4 million tokens per minute per team.
The contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens (cached input: $0.002), roughly 12x and 21x cheaper than standard. The tradeoff: Meta uses your prompts and completions to train future models. Rate limits drop to 60 requests per minute, signaling this tier targets individuals and small experiments rather than production. Zuckerberg steered new users toward this tier on launch. VentureBeat’s own testing confirmed the installer works but requires a payment method on file before the agent will run anything — low-cost is accurate, free is not.
Enterprises with proprietary codebases will need to consciously opt into standard pricing to keep their code out of Meta’s training pipeline.
️ The open source question
No Llama mention appears anywhere in the announcement. Meta’s Llama family accumulated roughly 1.2 billion downloads before the Llama 4 rollout drew criticism and benchmark controversy. The Muse Spark line, which launched April 8 as Meta’s first proprietary model, carries no downloadable weights and no self-hosting option. Asked directly on X whether Muse Code would be open source, Zuckerberg replied: “I’ll have more to share on that soon.”
Muse Code is available in beta now. Muse Spark 1.2 is live in the Meta Model API with expanded global access.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
