Meta Muse Code: The Third Major Terminal Coding Agent of 2026
On 5 August 2026, Meta Superintelligence Labs released Muse Code in public beta, making it available for macOS and Linux developers. Muse Code is a terminal-based agentic coding tool powered by Muse Spark 1.2, designed to plan, write, and validate code across large codebases. The launch makes Meta a direct competitor in the agentic software engineering tools market for the first time with a shipping commercial product, joining Anthropic's Claude Code and OpenAI's Codex as the three major terminal coding agents now in general developer access.
Muse Spark 1.2: The Model Behind the Agent
Muse Spark 1.2 is a coding-centric update to Muse Spark 1.1, Meta's developer-focused model series. On Meta's own evaluation harness, Muse Spark 1.2 scores 82.9 per cent on Terminal-Bench 2.1 — the benchmark designed to measure agentic coding performance in terminal environments — and 59.3 per cent on DeepSWE v1.1. Meta has not published independently verified leaderboard entries on these benchmarks as of the launch date, a caveat that third-party evaluators have flagged. Standard-tier pricing for Muse Spark 1.2 is unchanged from Muse Spark 1.1: 1.25 US dollars per million input tokens, 0.15 US dollars per million cached input tokens, and 4.25 US dollars per million output tokens. Web search grounding — used by Muse Code to look up documentation and package references — costs 2.50 US dollars per 1,000 queries.
The Contributor Tier: Steep Discount for Data Permission
Meta introduced a second pricing tier at launch called the contributor tier, priced at 0.10 US dollars per million input tokens and 0.20 US dollars per million output tokens — a discount of 12.5 times on input and 21.25 times on output versus the standard tier. In exchange, contributors explicitly permit Meta to use their prompts and completions to train future Muse models. The contributor tier is capped at 60 requests per minute, compared to the standard tier's 3,000. For individual developers doing exploratory work or building prototypes who are comfortable with the data permission terms, the contributor tier significantly lowers the cost of extended evaluation.
How Muse Code Works: Plan, Code, Validate, Resume
Muse Code operates as a terminal agent that accepts a task description and works through a structured sequence: planning the set of changes required, writing the code, running tests or build commands to validate the result, and iterating on failures. The agent can run multiple sub-agents in parallel to speed up tasks that can be decomposed — for example, running tests across different modules simultaneously while writing code elsewhere. Muse Code maintains a persistent action log: a record of every step it takes during a session. If the tool crashes or the user disconnects, it can read its action log on restart and resume from its last verified state, rather than starting over from scratch. The resume capability addresses a practical failure mode in long-running agentic tasks, where a session interrupted late in execution previously required a full restart.
The Competitive Landscape: Claude Code, Codex, and Muse Code
Three frontier AI companies now offer terminal coding agents in public access: Anthropic's Claude Code, OpenAI's Codex, and Meta's Muse Code. The convergence on terminal-based agents as the primary delivery format for agentic coding reflects a shared design conclusion — the terminal is where developers' existing tools live (version control, test runners, build systems, package managers), and an agent native to that environment integrates into existing workflows without requiring a new IDE or browser extension. Each agent has taken a different approach to pricing and data terms. Claude Code charges at Anthropic API rates. Codex has its own tier structure. Muse Code introduces the contributor tier as a novel mechanism to subsidise access in exchange for training data — an approach new in a commercial coding agent context.
What Indian Development Teams Should Know
For Indian software development teams evaluating agentic coding tools, Muse Code's arrival has two practical dimensions. The first is tooling choice: at 1.25 US dollars per million input tokens for the standard tier, Muse Spark 1.2 is priced competitively with comparable coding models. At the contributor tier's 0.10 US dollars per million input tokens, it is substantially cheaper than alternatives for developers prepared to accept the data terms — a calculation each team must make based on whether their codebases contain proprietary logic they prefer not to share for training purposes. The second is competitive context: three well-resourced terminal coding agents in the market in a single year compresses the timeline at which agentic coding becomes routine rather than experimental. Indian development teams that have not integrated agentic coding tools into their workflow are now operating in an environment where these tools can handle production-grade repository tasks at standard commercial pricing.
The Bottom Line
Meta Superintelligence Labs launched Muse Code on 5 August 2026 — a terminal coding agent in public beta for macOS and Linux, powered by Muse Spark 1.2. The model reports 82.9 per cent on Terminal-Bench 2.1 and 59.3 per cent on DeepSWE v1.1 in Meta's own evaluations. Standard-tier pricing is 1.25 US dollars per million input tokens and 4.25 US dollars per million output tokens. The contributor tier at 0.10 US dollars input and 0.20 US dollars output offers a 12.5 times input discount in exchange for data permission to train future models, with a 60 requests-per-minute cap. Muse Code plans changes, writes code, runs tests, operates multiple sub-agents in parallel, and resumes after a crash using its persistent action log. For Indian software teams, the launch completes a three-agent terminal coding field and establishes new benchmarks for what agentic coding tools can deliver at standard commercial pricing.
Frequently Asked Questions
What is Meta Muse Code and what can it do?+
Muse Code is a terminal-based agentic coding tool released in public beta by Meta Superintelligence Labs on 5 August 2026, available for macOS and Linux. Powered by Muse Spark 1.2, it accepts a task description and autonomously plans the required code changes, writes the code, runs tests or build commands to validate the result, and iterates on failures. It can run multiple sub-agents in parallel to speed up tasks that can be decomposed, maintains a persistent action log so it can resume after a crash without restarting from scratch, and integrates with web search to look up documentation and package references during a coding session.
What are Muse Spark 1.2's benchmark scores and how is Muse Code priced?+
On Meta's own evaluation harness, Muse Spark 1.2 scores 82.9 per cent on Terminal-Bench 2.1 and 59.3 per cent on DeepSWE v1.1. Independent verification on official leaderboards had not been published as of the August 5 launch date. Standard-tier pricing is 1.25 US dollars per million input tokens, 0.15 US dollars per million cached input tokens, and 4.25 US dollars per million output tokens. Web search grounding costs 2.50 US dollars per 1,000 queries. A contributor tier offers 0.10 US dollars per million input tokens and 0.20 US dollars per million output tokens — a 12.5 times input discount — in exchange for permission to use prompts and completions to train future Meta models, with a 60 requests-per-minute rate cap.
How does Muse Code compare to Anthropic's Claude Code and OpenAI's Codex?+
All three tools — Anthropic's Claude Code, OpenAI's Codex, and Meta's Muse Code — are terminal-based coding agents designed to plan, write, and validate code across large codebases. The convergence on terminal delivery reflects a shared design conclusion that working natively within the developer's existing terminal environment integrates into workflows without requiring a new IDE. Muse Code's contributor tier at 0.10 US dollars per million input tokens is a novel pricing mechanism not offered by the others, enabling much lower-cost access in exchange for training data permission. The three agents now compete across benchmark performance, pricing, data policy, and supported platforms.
What should Indian software development teams consider when evaluating Muse Code?+
Indian development teams evaluating Muse Code should consider two factors. First, pricing: at the standard tier of 1.25 US dollars per million input tokens, Muse Spark 1.2 is competitively priced with similar frontier coding models. The contributor tier at 0.10 US dollars per million input tokens significantly lowers evaluation cost, but requires explicit permission for Meta to train on your prompts and completions — teams working with proprietary codebase logic should review the data terms carefully before opting in. Second, maturity: Muse Code launched in beta on 5 August 2026, and teams should validate its resume-after-crash behaviour and multi-sub-agent coordination on their specific codebase types before committing it to production workflows.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.