On August 5, 2026, Meta Superintelligence Labs released Muse Code in beta: a terminal coding agent that directly enters the same market as Anthropic’s Claude Code and OpenAI’s Codex. The release came alongside Muse Spark 1.2, a coding-focused update to the model that powers it. For enterprise engineering teams evaluating coding agents, the launch adds a third serious competitor with a genuinely different architecture and a pricing strategy that is either unusually generous or unusually uncomfortable, depending on which tier you choose.
What Launched
Muse Code is installable on macOS or Linux with a single curl command. It handles complete software engineering tasks across large repositories, including planning changes, writing code, and validating results. Mark Zuckerberg described it on X as a tool that “takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results.”
Muse Spark 1.2, the model powering the agent, is available in the Meta Model API with expanded global access and on OpenRouter. Meta co-trained the model with Muse Code itself, using rejection-sampled harness trajectories and recipe optimizations for goals, context compaction, and sub-agents. The result is a model that was explicitly trained to perform inside this specific tool, not adapted from a general-purpose checkpoint after the fact.
This co-training approach mirrors a broader industry shift. The assumption that models and harnesses are separable products is giving way to integrated systems where the model and its runtime are developed together. Anthropic does this with Claude Code and Claude Sonnet 5, and Meta is now pursuing the same pattern.
The Architecture Bet: Persistent Background Agents
Most coding harnesses spawn helper agents fresh for each task and discard them when the task is complete. Muse Code takes a different position. It keeps a set of specialized background agents alive for the entire session. These agents carry out next steps independently and decide when to report back to the main agent, avoiding the redundant context loading that occurs when a fresh agent must re-read a repository’s structure for every subtask.
The second architectural choice is auditability. Every model call, tool run, approval, and edit is appended to a local event log before it executes. Meta calls this property “replay-exact and restart-safe”: if a session is interrupted, it can be resumed from the event log without re-running earlier steps. For enterprise teams running Muse Code on long overnight or weekend tasks, this is a meaningful operational property.
Meta’s long-horizon demo points Muse Spark 1.2 at GPU kernel optimization across more than 1,000 tool calls over up to 24 hours on NVIDIA Hopper hardware. That is the kind of workload coding agents are increasingly asked to take on, and it sets an implicit reliability bar for any competitor claiming the same use case.
Pricing: Two Tiers, One Tradeoff
The pricing structure is where Meta’s strategy diverges most sharply from rivals:
| Tier | Input (per 1M tokens) | Output (per 1M tokens) | Cached Input | Data Policy |
|---|---|---|---|---|
| Standard | $1.25 | $4.25 | $0.15 | Not used to train Meta models |
| Contributor | $0.10 | $0.20 | $0.002 | Used to train future Meta models |
The contributor tier is 12 to 21 times cheaper than standard. Meta is steering new users toward it: “It’s easy and low-cost to get started,” Zuckerberg wrote. “Install Muse Code with one line and you can start on our contributor tier.”
For enterprise teams, the relevant tier is standard. It commits that prompts and completions will not be used to train Meta’s models, which is the minimum requirement for any deployment touching proprietary code or customer data. Alexandr Wang, who leads Meta Superintelligence Labs, confirmed the company is also starting to accept requests for zero-data retention, a dedicated enterprise feature that would remove any retention of request content.
The contributor tier is effectively a distribution strategy: it gets developers using the product at near-zero cost, builds Meta’s training data flywheel, and creates a path to standard or zero-retention pricing once teams see value. It is a similar dynamic to how cloud providers offer generous free tiers to hook developers who later convert to paid enterprise plans.
How It Fits the Competitive Landscape
The terminal coding agent category has, until this release, been effectively a two-party market. Claude Code leads on benchmark performance and deep integration with the Anthropic model family. OpenAI’s Codex carries the brand weight and integrates with GPT-5.6. A set of well-funded startups including Cursor, Windsurf, and others compete on user experience and IDE integration.
Meta’s entry adds a third serious option, with three genuine differentiators: persistent session agents, an append-only local event log, and the cheapest standard pricing of any frontier-class harness.
The open question is developer trust. Enterprise teams evaluating Muse Code will need to satisfy themselves that Meta’s data commitments on the standard tier hold in practice, that the persistent agent model does not introduce unexpected behavior on complex codebases, and that the underlying Muse Spark 1.2 model matches Sonnet 5 and GPT-5.6 on real enterprise repositories, not just curated benchmarks. Meta has not yet released independent SWE-bench results for Muse Spark 1.2 in the Muse Code harness configuration.
What This Means for Enterprise AI Builders
Three practical implications for enterprise teams:
Standard tier is the only viable starting point. The contributor tier’s data policy is incompatible with most enterprise code policies. Start with standard, and monitor Meta’s zero-data retention rollout for teams with stricter requirements.
Persistent background agents reduce session cost at scale. If Muse Code’s architecture holds in production, long multi-step agent runs should consume fewer total tokens than a spawn-and-discard pattern, because background agents avoid re-loading the same context repeatedly. Track token costs per completed task, not per individual call, when benchmarking against alternatives.
The coding agent market is now three-way. Procurement and platform decisions made in 2026 will be harder to reverse than those made in 2025, because more workflows are being automated into harness-specific patterns. Evaluate Muse Code alongside Claude Code and OpenAI Codex before committing to a primary harness for your engineering teams. The architecture differences matter more than the current benchmark rankings.
Meta built Muse Spark from scratch over 14 months and shipped its first proprietary model in April 2026. Muse Code is the third release from Meta Superintelligence Labs in four months. The pace is consistent with a lab that is catching up on product capabilities, not just model performance. Enterprises that dismissed Meta’s AI strategy as an advertising-focused distraction now have a production-grade coding agent to evaluate.
Muse Code is available in beta at dev.meta.ai. Muse Spark 1.2 is live in the Meta Model API and on OpenRouter. Primary source: Meta AI Research blog. Coverage: VentureBeat, TechCrunch, CNBC.