On August 5, 2026, Meta entered one of the fastest-moving segments of the AI industry: autonomous coding agents. The company released Muse Code in beta, a command-line coding agent powered by its new Muse Spark 1.2 model, positioning it as a lower-cost rival to Anthropic’s Claude Code and OpenAI’s Codex. Five days later, Mark Zuckerberg announced on Threads that Meta would open-source Muse Spark 1.2’s weights entirely.

The move puts a third heavyweight into a market that Anthropic and OpenAI have controlled for roughly a year. It also continues a pattern Meta has used before: release a capable open model, drive down costs across the ecosystem, and force rivals to respond on price.

What Muse Code Actually Does

Muse Code is not a plugin for an existing code editor. It runs entirely from the terminal on macOS and Linux, installed with a single command. That makes it closer in concept to Claude Code, which also operates as a CLI agent, than to browser-based tools like GitHub Copilot.

When given a substantial task, Muse Code breaks it into parallel workstreams. It spins up multiple sub-agents, each working in an isolated worktree, so the main branch remains clean while agents experiment and iterate concurrently. Every model call, tool run, approval, and edit gets appended to a local event log that Meta describes as making sessions “replay-exact and restart-safe.” Crash the process mid-session, restart it, and Muse Code picks up from where the log left off.

The underlying model, Muse Spark 1.2, was built by Meta Superintelligence Labs, a unit the company assembled after a high-profile talent acquisition push in 2025. Spark 1.2 is specifically optimized for long-context coding tasks rather than general conversation.

The Benchmark Numbers

Meta published results on Terminal-Bench 2.1 and DeepSWE 1.1, two independent evaluations widely used to rank coding agents.

On Terminal-Bench 2.1, Muse Code scored 82.9%. Claude Code running on Anthropic’s Opus 5 model scored 86.7%, and OpenAI’s Codex with GPT-5.6 Terra scored 81.8%. Muse Code lands almost exactly between the two incumbents, slightly behind Claude Code and slightly ahead of Codex.

On DeepSWE 1.1, the gap widens. Claude Code scored 65%, Codex scored 64.8%, and Muse Code scored 59.3%. DeepSWE is a harder benchmark focused on resolving real GitHub issues from open-source repositories, so the 5 to 6 point gap is meaningful for teams doing deep repository work.

What the numbers show is that Muse Code is a genuinely competitive product, not a demonstration piece. It does not lead the field in accuracy, but it is close enough that the delta is unlikely to matter for most development tasks.

Why the Pricing Changes the Conversation

This is where Meta’s entry becomes significant for anyone running software development at scale.

Muse Code is priced at $1.25 per million input tokens and $4.25 per million output tokens on the standard pay-as-you-go plan. That is already aggressive. But the contributor tier takes it further: developers who consent to having their prompts and completions used to train the model pay just $0.10 per million input tokens and $0.20 per million output tokens. Compared to Claude Pro at $20 per month for individual use, teams processing high volumes of code through an API are looking at dramatically lower costs per token with Muse Code on the contributor tier.

The context here matters. In late July 2026, OpenAI cut prices on GPT-5.6 Luna by 80%, dropping input token costs to $0.20 per million. That cut was widely attributed to competitive pressure from Chinese AI providers who had captured 46% of US enterprise AI spend by that point. Meta’s pricing on Muse Code continues the same trend: the cost of running AI on code is falling fast, and any business that locked in higher rates a year ago should be reviewing those contracts now.

The Open-Source Dimension

Zuckerberg’s announcement on August 10 that Muse Spark 1.2 weights would be released as open source adds another layer to this story. Open weights mean that any company with sufficient GPU infrastructure could run the underlying model without paying Meta anything at all. For businesses in Cyprus and across Europe that are cautious about sending proprietary code to third-party APIs, a self-hosted version of Muse Spark 1.2 may represent the most cost-effective and privacy-compliant path to AI-assisted coding.

Meta has used this playbook before with its LLaMA family of models. Open-sourcing a competitive model creates adoption, reduces the reputational risk of being a closed model company, and applies pricing pressure on rivals. It also builds a community of developers who extend and fine-tune the model, which feeds back into future training data.

What This Means for Businesses

If your development team is not yet using an AI coding agent, the Muse Code launch is a reasonable moment to run a pilot. The barrier to entry is low: the CLI installs in seconds, there is no platform lock-in, and the pay-as-you-go pricing means you can test it on real tasks without committing to a subscription.

If your team is already using Claude Code or Codex, the competitive dynamics created by Meta’s entry are worth watching. Both Anthropic and OpenAI are likely to respond with price adjustments or capability improvements in the months ahead. Businesses with annual contracts should check renewal terms, and teams running large-scale automated workflows should model out whether Muse Code’s pricing changes the economics of their existing setup.

The broader picture is that the coding agent market has moved from “two serious options” to “three serious options” within a month. All three tools can now handle planning, writing, debugging, and validating code end-to-end. The differentiation is increasingly about ecosystem depth (Claude Code leads here with hooks, subagents, and MCP support), context-handling for large codebases, and per-token cost at scale.

For businesses in Cyprus and the wider region that are exploring AI-assisted software development, this competition is good news. The tools are getting more capable, the prices are falling, and the option to self-host on open weights gives enterprises a privacy-safe path that did not exist at this quality level twelve months ago. The right moment to start evaluating these tools is now, before your competitors have already built internal workflows around them.