Meta today made a decisive entry into the competitive AI coding agent space with the beta release of Muse Code, a terminal-based agent, and the launch of Muse Spark 1.2, a coding-focused update to its frontier model family. The coordinated release positions Meta directly against Anthropic’s Claude Code, OpenAI’s Codex, and a rapidly expanding field of agentic development tools that are increasingly defining how professional software is built.
The announcement, made by Meta CEO Mark Zuckerberg on X, introduces a tool that can take on complete software engineering tasks across large code repositories, from planning changes to writing code and validating results. For a company that has largely observed the rise of dedicated coding agents from the sidelines—while rivals Anthropic and OpenAI turned theirs into flagship products and startups like Cursor built billion-dollar valuations on the workflow—the launch marks a significant strategic pivot.
Meta’s developer story has for years centered on Llama, the open-weight model family that surpassed a billion downloads. Muse Code, installable on macOS or Linux with a single command, is a different proposition entirely: a proprietary harness co-trained with the model that powers it. Yet Zuckerberg has hinted that open source may still be in the cards, responding to a direct question on X with “I’ll have more to share on that soon.”
What is Meta Muse Code and How Does Its Architecture Work?
Muse Code is a terminal-based AI coding agent that operates as a full harness for software development. Unlike many rivals that spawn helper agents fresh for each task, Muse Code employs a novel architecture built around async background agents. These specialized agents remain active for an entire session, avoiding redundant information gathering and reducing latency. Instead of re-exploring a repository with every new request, the agent maintains persistent context, carrying out next steps autonomously and choosing when to report back to the main agent.
For large-scale tasks, Muse Code fans out to separate sub-agents working in parallel. Each sub-agent operates in its own isolated git worktree, ensuring the developer’s working copy is never disturbed. Zuckerberg reported that in testing, the system built six features for a game simultaneously without any collisions. While worktree isolation and parallel sub-agents exist in competing tools, Meta differentiates its offering through the combination of persistence plus parallelism.
The system also prioritizes auditability. Every model call, tool run, approval, and edit is appended to a local event log before it executes. This creates a single source of truth that makes the runtime “replay-exact and restart-safe.” If Muse Code crashes 20 hours into a complex task, it resumes precisely where it stopped, with no lost work and no re-prompting. For engineering leaders who have been burned by opaque agent runs, this complete local audit trail may prove to be the most critical feature in enterprise evaluations.
Muse Code ships with bundled skills familiar to users of rival tools: /plancodecodecodecode transforms a task into an approval-gated plan, /grillcodecodecodecode stress-tests that plan until it holds up, and /goalcodecodecodecode drives the agent toward completion of a stated objective.
Muse Spark 1.2: Co-Trained with Its Own Harness
At the heart of Muse Code lies Muse Spark 1.2, which Meta describes as a coding-focused update with significantly scaled-up training compute on coding tasks and broader training environment diversity. The update directly addresses what was arguably the original Muse Spark’s weakest flank. When the model debuted in April, it vaulted Meta back into the top five on frontier reasoning and vision benchmarks but trailed on the agentic coding evaluations that matter most in this market.
Four months later, a coding-specialized checkpoint paired with a purpose-built harness reads as Meta’s direct answer to that gap. Two training details stand out as particularly significant. First, Meta co-trained the model with Muse Code itself, using rejection-sampled harness trajectories and recipe optimizations tailored for goals, context compaction, and sub-agents. This means the model was explicitly tuned to perform best inside this particular tool—mirroring an industry-wide shift away from treating models and harnesses as separable products.
Second, Meta employed a self-improvement loop: Muse Spark 1.1 generated challenging coding environments and instruction-following templates, then graded candidate solutions against those requirements, producing a scalable training dataset for its successor. Meta credits this loop with making version 1.2 measurably better at following complex instructions.
Benchmark Performance: A Clear but Honest Second Place
Meta published benchmark charts comparing Muse Spark 1.2 against other coding models, though the announcement notably avoids touting any specific placements. The charts explain why: they show a strong but clear second place, with Anthropic’s Claude systems leading across all evaluations.
On Terminal-Bench 2.1, Muse Spark 1.2 running in Muse Code scored 82.9%, edging OpenAI’s GPT-5.6 Terra in Codex (81.8%) and xAI’s Grok 4.5 in Grok Build (81.6%) but trailing Anthropic’s Opus 5 at max effort in Claude Code, which leads at 86.7%. On DeepSWE 1.1, Muse Spark 1.2 posted 59.3%—third, behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). Most striking is Meta’s own internal coding benchmark, where Muse Spark 1.2’s 70.6% comfortably beats GPT-5.6 Terra (65.4%) and Gemini 3.6 Flash (63.9%) yet still sits nearly nine points behind Opus 5’s 79.4%—an unusually candid admission that even on the test Meta designed itself, Anthropic’s model wins.
The generational gains are real, however: Muse Spark 1.2 improves on version 1.1 by 6.7 points on Terminal-Bench and 6.3 on DeepSWE. One caveat worth noting is that the 1.1 scores were recorded in a generic harness while 1.2 ran in Muse Code, meaning some of that improvement belongs to the new harness rather than the new model.
Long-Horizon Autonomy: A 24-Hour GPU Kernel Optimization
Meta’s most striking demonstration involved pointing Muse Spark 1.2 at GPU kernel optimization on NVIDIA Hopper hardware. Working in Triton and barred from simply wrapping existing third-party kernel libraries, the agent wrote, compiled, and profiled its way to what Meta calls “substantial improvements” over baseline implementations of KDA and MLA kernels—including genuinely non-obvious optimizations like re-centering gated cumulative decay at a chunk midpoint. The agent ran for more than 1,000 tool calls over up to 24 hours, with Zuckerberg noting that “it kept finding substantial improvements well beyond the initial exploration phase.” Sustained improvement over a full day of autonomous operation, if it holds up outside Meta’s controlled demos, addresses one of the most persistent criticisms of coding agents: that they plateau or drift once past their initial burst of progress.
Meta’s Pricing Strategy for Muse Spark 1.2: Your Data for a Discount
The pricing structure may prove to be the most consequential—and most scrutinized—part of the launch. Meta is offering Muse Spark 1.2 through its Meta Model API in two distinct tiers that represent fundamentally different value propositions.
The standard tier is priced at $1.25 per million input tokens and $4.25 per million output tokens (with cached input at $0.15). Meta commits that prompts and completions on this tier are not used to train its models. There is no long-context premium, and rate limits run to 3,000 requests and 4 million tokens per minute per team. This positions it about mid-range compared to other leading AI models available over API.
The contributor tier is where Meta’s strategy diverges sharply from its rivals. At $0.10 per million input tokens and $0.20 per million output tokens—roughly 12x and 21x cheaper than standard, respectively, with cached input at a near-free $0.002—it is the cheapest available on the market. The trade-off is explicit: users grant Meta permission to use their prompts and completions to train future models.
This is the tier Zuckerberg is steering new users toward, promoting it as an easy and low-cost way to get started. However, in VentureBeat’s own testing on a Mac mini, the one-line installer worked as advertised but stopped short of running anything, reporting that payment was “required to finish setting up your account.” Even the heavily discounted contributor tier requires a payment method on file before Muse Code will do any work.
Meta frames the contributor tier as lowering the barrier for prototyping and experimentation “where training on your data is acceptable.” But this means the default on-ramp for Muse Code sends developers’ code and prompts into Meta’s training pipeline—a tradeoff that enterprises with proprietary codebases will need to consciously opt out of by moving to standard pricing. The contributor tier also carries much tighter rate limits (60 requests per minute versus 3,000), a clear signal it is aimed at individuals and small experiments rather than production workloads.
The approach is classically Meta: subsidize access, harvest data at scale, and use it to close the gap with the frontier. The question for developers and enterprises is whether the steep discount justifies the privacy compromise, particularly for legally protected or commercially sensitive code.
Meta and Open Source: The Llama Era Has Ended
What today’s announcement conspicuously lacks is any mention of open source—a striking omission from the company that spent three years positioning itself as the standard-bearer of open AI. From the original LLaMA’s debut in February 2023, through Llama 2’s commercially usable license, the coding-specialized Code Llama, and the 405-billion-parameter Llama 3.1, Meta’s entire pitch to developers was that frontier-class weights should be free to download, self-host, and fine-tune.
That strategy worked spectacularly for a time. By early 2026, the Llama family had been downloaded roughly 1.2 billion times, averaging about a million downloads a day. Self-hosting offered enterprises cost reductions reported at as much as 88% versus proprietary API providers. But the unraveling came swiftly. Llama 4 debuted to mixed reviews and, eventually, admissions that its benchmark results had been misleading. Meanwhile, Chinese open-weight rivals from DeepSeek, Alibaba, and Zhipu AI surged to account for roughly 41% of downloads on Hugging Face by late 2025, eroding Llama’s claim to leadership of the movement it started.
The Llama era effectively ended on April 8, 2026, when Meta shipped the original Muse Spark as its first proprietary model: cloud-only, with no downloadable weights and no self-hosting. Asked at the time whether Llama development would continue, a Meta spokesperson told VentureBeat only that “our current Llama models will continue to be available as open source”—pointedly silent on future ones. Four months later, today’s release does nothing to advance any promise of open-source availability: no weights, no license, and neither the blog post nor Zuckerberg’s thread so much as uses the word “open.”
The reversal is all the sharper because Meta’s rivals have been moving in the opposite direction. OpenAI released its Codex CLI as open source under the permissive, enterprise-friendly Apache 2.0 license and followed with its gpt-oss open-weight models. Google’s Gemini CLI harness is likewise Apache-licensed. With Muse Code, Meta lands closest to the posture of Anthropic, whose Claude Code remains proprietary, while the company that once argued open source was the path forward now asks developers to pay per token for a model they cannot inspect, or to subsidize that access with their own data.
Seen in that light, the contributor tier reads as the successor to the Llama strategy itself: the ecosystem flywheel is no longer free weights in exchange for mindshare, but cheap tokens in exchange for training data. However, Zuckerberg’s reply on X—”I’ll have more to share on that soon”—does keep hope alive that Meta will return to the open-source AI game.
The Strategic Significance of Meta’s AI Coding Push
Terminal coding agents have become the fastest-growing surface in enterprise AI, and until today the category has effectively been a two-horse race between Anthropic and OpenAI, with Google and a crowd of startups in pursuit. Meta’s entry brings a genuinely different architectural approach with persistent background agents and an append-only local event log, a credible long-horizon demonstration spanning 24 hours of autonomous optimization, and an aggressive pricing wedge that undercuts virtually every competitor.
Yet the open questions are precisely the ones benchmark charts cannot answer. Whether Muse Spark 1.2 actually matches Claude and GPT-class models on real-world repositories is something only extensive developer testing will reveal. Whether developers trust Meta with their code, particularly in light of the contributor tier’s data-harvesting model, will determine adoption rates. And whether the contributor tier’s discount is enough to make developers stop asking those hard questions remains to be seen.
Muse Code is available in beta today. Muse Spark 1.2 is live in the Meta Model API with expanded global access. For developers willing to install a terminal agent and log in with a Meta account, the future of AI-assisted coding now includes a third major player—one that brings both a compelling technical architecture and a set of tradeoffs that the developer community will need to weigh carefully.
