Key Takeaways:
Meta Platforms Inc. entered the AI coding-agent race on Aug. 5 with Muse Code, a terminal-based tool priced at $0.10 per million input tokens on its contributor tier — roughly 12 times cheaper than its own standard rate and a fraction of what Anthropic and OpenAI charge, a bid to pull developers away from Claude Code and Codex.
"Releasing Muse Code in beta today," Chief Executive Officer Mark Zuckerberg wrote on X. "It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results."
The agent, overseen by Alexandr Wang, head of Meta Superintelligence Labs, runs on the newly released Muse Spark 1.2 model, which Meta co-trained with the harness itself. The standard pay-as-you-go tier costs $1.25 per million input tokens and $4.25 per million output, with cached input at $0.15; the contributor tier, where developers let Meta train on their session data, drops to $0.10 input and $0.20 output, with cached input near-free at $0.002. Rate limits run to 3,000 requests and 4 million tokens per minute on standard pricing versus 60 requests per minute on the contributor tier, a signal the discount is aimed at individuals and small experiments rather than production workloads.
The pricing wedge is the sharpest in the category. Anthropic's Claude Opus 5 charges $5 per million input tokens and $25 per million output, while OpenAI's GPT-5.5 runs $5 and $30 respectively — putting Meta's contributor tier at roughly 2 percent of Claude Opus 5's combined per-token cost. Meta is competing on price and architecture rather than benchmark headroom, and it did not headline specific leaderboard scores in the announcement, a notable omission in a field where rivals trumpet placement. Muse Spark 1.1 scored 77.4 on SWE-Bench Verified against Claude Opus 4.6's 80.8 and Gemini 3.1 Pro's 80.6, the gap the coding-specialized 1.2 checkpoint is meant to close.
Co-training and the long-horizon demo
Muse Code's design differs from rivals in two ways. It keeps a set of specialized background agents alive for the entire session rather than spawning helpers per task, cutting redundant repository exploration, and it appends every model call, tool run, approval, and edit to a local event log that makes the runtime replay-exact and restart-safe — if the agent crashes 20 hours into a job, it resumes where it stopped. When a task is large enough, Muse Code fans out to parallel sub-agents in isolated git worktrees, so the developer's working copy is never touched. Bundled skills include /plan, an approval-gated planning step, and /grill, which stress-tests a plan.
Meta's most striking demonstration is a long-horizon case study: Muse Spark 1.2 ran more than 1,000 tool calls over up to 24 hours on NVIDIA Hopper GPUs, writing, compiling, and profiling Triton kernels for KDA and MLA workloads while barred from importing third-party kernel libraries. The agent found non-obvious optimizations, including re-centering gated cumulative decay at a chunk midpoint, and kept improving well past the initial exploration phase, Zuckerberg said. The test measures sustained autonomous performance over an extended session, distinguishing it from one-shot code generation benchmarks.
The launch marks a strategic turn for Meta, whose developer story long centered on the open-weight Llama family, downloaded roughly 1.2 billion times. Muse Code is entirely proprietary — no weights, no license — and the contributor tier reads as the successor to the Llama flywheel: cheap tokens in exchange for training data rather than free weights in exchange for mindshare. Meta is also starting to accept requests for zero-data retention, a feature common in enterprise agreements with OpenAI and Anthropic that gives Muse Code a path into corporate environments where data-handling compliance is a procurement requirement.
What's at stake for investors
META shares ended the launch day up 0.1 percent and are down 10.5 percent year-to-date. The coding-agent category has been a two-horse race between Anthropic and OpenAI, with Google's Gemini CLI and startups like Cursor in pursuit; Meta's entry adds a third well-capitalized player with a genuinely different architecture and an aggressive price point. For developers, the contributor tier removes the per-token friction that has slowed trial of competing services, though it sends session data into Meta's training pipeline and requires a payment method on file even at the discounted rate. Muse Code currently lacks a Windows client and a dedicated GUI, a meaningful barrier for teams outside macOS and Linux environments. Anthropic and OpenAI are likely to respond with their own developer-access programs, and whether Muse Spark 1.2 can close the capability gap while holding its pricing advantage will determine whether Meta's land-and-expand bet pays off.
This article is for informational purposes only and does not constitute investment advice.
