▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=-Gj0-EIyx6g
• Original title :: Meta's Claude Code clone is INSANELY cheap

Theo takes Meta's newly released Muse Code CLI (a close Claude Code clone) and its Muse Spark 1.2 coding model for a live spin. The model benches around 54 on the artificial-analysis intelligence index — roughly tied with GPT-5.5 and Grok 4.5, behind Opus 5, Fable 5 and Kimi K3 — but it is absurdly fast (~190 tok/s average, 316 peak) and, on Meta's data-sharing "contributor" tier, 10-20x cheaper than its own list price. His verdict: not a daily driver for real code, but a genuinely useful and near-free tool for cheap, parallel, code-adjacent analysis.

What Meta shipped

Muse Code is a terminal coding agent powered by Muse Spark 1.2, announced by Zuckerberg. Claimed features: background agents that accumulate context across a whole session instead of restarting per task; automatic fan-out into parallel sub-agents in isolated git worktrees; and an auditable local event log written before every model call, tool run and edit so a crashed run resumes exactly where it stopped without re-prompting. Meta also claims a kernel-optimization run of ~1000 tool calls over 24 hours on Nvidia Hopper that kept finding improvements well past the exploration phase.

Theo notes Meta is unusually well positioned for large-repo agentic work: they abandoned git for custom Mercurial-based tooling, rewrote PHP into Hack rather than rewrite their codebase, and generally fix the tools instead of laying off the teams that build them (a jab at Google). The CLI is closed source, which he calls a mistake for an "open source company at heart", though Meta does document setup with open code, Claude Code, Codex, curl and Python via what looks like an OpenAI-compatible API.

Benchmarks and pricing

Meta's own benchmark tables conspicuously omit Fable and Soul, comparing only against Opus 5 and 5.6 Terra. On Terminal Bench 2.1 Muse Spark 1.2 slightly beats Terra and trails Opus 5; on DeepSWE it edges past Grok 4.5 but stays behind Terra and Opus; on Meta's internal bench it lands second. Artificial analysis puts it at 54 on the intelligence index — up 3 points from 1.1 and 11 from 1.0 in April — which Theo finds embarrassing for a closed-weight lab given the open-weight Kimi K3 shipped earlier and scores higher. Hallucination numbers allegedly improved (abstention up, hallucination rate down 10 points), which his own testing flatly contradicts.

The real story is price. List is $1.25/M in, $0.15/M cached, $4.25/M out. The contributor tier — where Meta reads and trains on everything you send — is $0.10/M in, $0.20/M out, $0.002/M cached: a 10-20x gap that makes it the cheapest model on the chart at roughly 2-3 cents per benchmark task. Token efficiency is middling (~30k tokens/task vs ~36k for Fable 5 and ~17k for 5.6 Soul), but throughput is the standout: ~191 tok/s average on OpenRouter against ~30 tok/s for Soul.

Hands-on findings

The CLI picked up his existing Claude Code skills and rules directory rather than the shared agents/AGENTS.md standard — consistent with Meta being, by his account, a heavy Anthropic shop internally. It supports worktrees and a YOLO mode, has vision, and shipped the first genuinely good streaming voice-to-text he has seen in any coding CLI. He calls the harness more stable and less annoying than Claude Code itself.

On a codebase audit of the T3 code event-sourcing model it produced findings and a synthesized HTML report in under a minute while Fable and DeepSeek V4 Flash were still working after four. But cross-review was harsh: Fable rated its own plan meaningfully better, and Soul scored Muse's integration plan 4.8/10 against Fable's 7/10. Asked to remake his fish game, it built 2D and 3D versions in under five minutes total (Opus 5 took over an hour) with pleasant sprite animation, shadows and a distinct visual "flavor" — but broken mouse look, backwards-swimming fish and no collision understanding, and repeated fix attempts kept missing.

The worst moment: asked to integrate Muse as a provider into T3 code, it hallucinated the entire investigation as being about "anti-gravity", ran the whole thing on that assumption, then asked Theo what Muse even was — while running inside Muse Code. The eventual integration it produced did not work at all. Rate limits were a constant irritant, hitting even after he moved off the contributor tier onto paid per-token, which he considers unreasonable.

The use case that landed

Where the model earned its keep was cheap, parallel, code-adjacent analysis. Asked to audit every open PR on T3 code, fan out across sub-agents and produce a prioritized mergeability list, it indexed and reviewed 222 pull requests in under five minutes for ten cents on the contributor tier. The output page included clickable PR links (which Codex and Claude routinely omit unless told), confidence scores, and clean/dirty merge status — a layout he liked better than what his usual models produce and might adopt as a template, possibly automated as a daily job.

Cost comparison for the session: roughly $0.40 on the contributor tier versus nearly $32 of Fable and Opus spend in Claude for less work — about a 100x gap. Even at list price the same audit would be around $2.

Verdict

Muse Spark 1.2 "knows how to act like a modern smart model but doesn't know what the modern smart models know" — excellent at decomposing work into non-colliding sub-agents, poor at seeing complex work through, comparable in character to Grok Code Fast. Theo would rather wait 2-3x longer for output he can actually merge than retry five times fast and still get a mess. Nobody should make this their day-to-day coding model, and he would still pick Grok 4.5 over it. But for pulling signal out of noise cheaply — title generation, categorization, report summaries, log digging, PR triage — it is solid, and the different aesthetic flavor makes it fun to experiment with. He frames the broader picture as big companies (Meta, Google, Apple) fighting over fifth place while startups like Moonshot and Z.ai race ahead, and notes the wider trend of good cheap models returning: OpenAI cut Luna's price 80%, and DeepSeek V4 Flash shipped a strong new snapshot.