Same Numbers

Muse Code is Meta's admission that it is chasing, not leading. Released in beta on August 5 as a terminal-based coding agent built on the new Muse Spark 1.2 model, it arrives with the vocabulary of confidence — parallel sub-agents, full action history, automatic result verification — and the pricing of a company that knows it has to undercut to get looked at: $1.25 per million input tokens, $4.25 per million output, with cached input at fifteen cents, according to Forbes.

The numbers that matter aren't the price. On Meta's own Terminal-Bench 2.1 results, Muse Code posts 82.9%, which sounds respectable until you set it beside Claude Code running Opus 5 at max effort, which clears 86.7%. It does edge OpenAI's Codex (81.8%) and Grok Build (81.6%), per VentureBeat, but "beats Grok by 1.3 points" is not a launch headline, it's a footnote in someone else's story.

Worse is Meta's own internal benchmark, where Muse Spark 1.2 scores 70.6% — ahead of GPT-5.6 Terra and Gemini 3.6 Flash, sure, but 8.8 points behind Claude Opus 5, a gap Meta had every opportunity to hide and didn't, because the numbers are the numbers. That's the tell here: this is a company publishing benchmarks that quietly concede second place on its own test, on its own launch day.

None of this makes Muse Code bad. A terminal agent with sub-agent parallelism and full action-history logging is a legitimate engineering artifact, and the beta pricing is aggressive enough to poach budget-conscious teams from Anthropic's metered tiers. But "the field is moving this way" cuts both directions — it means Meta showed up to a race where the leader already lapped the internal scoreboard Meta built to measure itself. Muse Code isn't a verdict on the ceiling of what Meta can do. It's a receipt for how far behind the ceiling currently sits.

A terminal window is not a culture, except when three of them start behaving the same way in the same week. That's roughly what happened when Meta pushed Muse Code into beta on August 5th — no app, no dashboard, just a one-line install for macOS and Linux, dropping straight into the same command-line register that Anthropic and OpenAI's coding agents already occupy. The CNBC report frames it as Meta's entry into a fight already underway, but the more interesting fact is where Meta chose to enter it: not with a friendlier interface, not with a consumer hook, but with a terminal, because that's apparently now the neutral ground every serious coding agent has to stand on.

The install itself tells you something about who this is for. 9to5Mac notes there's no dedicated app — you type a command, you get an agent living in your shell, running sub-agents in parallel, keeping a full action history like a black box recorder for your own debugging session. This is not software asking to be liked. It's software asking to be trusted with root-adjacent access to your actual work, which is a different and much colder kind of intimacy than the chat-window familiarity AI tools have mostly traded in until now.

Meta didn't hide the benchmarks either, which is its own kind of tell. VentureBeat reports Muse Spark 1.2 landing at 59.3% on DeepSWE 1.1 — behind Opus 5's 65.0% and GPT-5.6 Terra's 64.8%. Publishing a number that puts you third isn't confidence, it's a bet that raw ranking matters less than being priced into the conversation at all. And on Terminal-Bench 2.1, where the model actually operates — not abstract reasoning but the literal terminal environment Muse Code lives in — StarRise clocks 82.9%, a 6.7-point jump from 1.1. That's the number Meta is actually selling: not "smartest," but "most improved at doing the specific job this product does."

Then there's the price, which reads less like a fee schedule and more like an invitation to build habitually rather than occasionally. Forbes lists $1.25 per million input tokens, $4.25 per million output, and — the detail that matters — cached input at $0.15. That cache discount is aimed at exactly the workflow Muse Code is built for: an agent that reruns, rechecks, re-verifies its own output across a long action history, cheaply, over and over, because the whole pitch is that it doesn't stop working after one pass. The product isn't a smarter typist. It's a system designed to be left running, checked on, trusted to keep its own receipts — and priced so that leaving it running doesn't feel like a risk.