Meta shipped a coding agent on August 5, and the number everyone repeated was 82.9% on Terminal-Bench 2.1. That is the wrong number to care about.
The number that matters is $0.10.
Muse Code is a terminal coding agent for macOS and Linux, powered by Muse Spark 1.2, a new coding-focused model from Meta Superintelligence Labs. It installs with one command:
curl -fsSL https://dev.meta.ai/install.sh | bash
No Windows build. No desktop app. Just a terminal, a login, and a hosted model — Meta's launch post says nothing about downloadable weights, so treat Muse Spark 1.2 as a closed dependency.
The architecture is the good part
Strip away the benchmark marketing and three design choices here are ahead of most of the field.
Persistent async background agents. Most harnesses spawn a subagent, use it, discard it, then spawn a fresh one that re-reads the same twelve files. Muse Code keeps a set of specialized background agents alive for the whole session. They accumulate context, carry out next steps on their own, and choose when to report back to the main loop. Meta's argument is that persistence cuts latency and reduces how often you have to steer.
An append-only local event log. Every model call, tool run, approval, and edit is appended to a single source of truth. Meta calls the runtime replay-exact and restart-safe: after a crash, the agent resumes precisely where it stopped instead of restarting the task. Anyone who has lost four hours of agent work to a dropped SSH session understands the value instantly. It also produces an audit trail, which matters more than most teams currently admit.
Three bundled skills. /plan converts a task into an approval-gated plan. /grill stress-tests that plan until it holds up. /goal drives toward a stated objective. /grill is the interesting one — an explicit adversarial pass over the plan before any code gets written.
Meta also co-trained the model with the harness, using rejection-sampled harness trajectories plus recipe tuning for goals, compaction, and subagents. The pitch: a purpose-built pairing beats wrapping a general-purpose model in a generic agent loop. Reasonable in principle. Untested in the wild.
About that 82.9%
Meta's chart puts Muse Spark 1.2 at 82.9% on Terminal-Bench 2.1, behind Claude Opus 5 at 86.7%. Here is the part the headlines skipped: Meta published three coding comparisons and came second or third on all three.
| Benchmark (Meta-reported) | Claude Opus 5 | Muse Spark 1.2 | Also ahead of Muse |
|---|---|---|---|
| Terminal-Bench 2.1 | 86.7% | 82.9% | GPT-5.6 Terra 81.8%, Grok 4.5 81.6% — within 1.3 pts |
| DeepSWE 1.1 | 65.0% | 59.3% | GPT-5.6 Terra 64.8% |
| Meta Internal Coding Bench | 79.4% | 70.6% | — |
Meta lost its own home fixture by 8.8 points and shipped the chart anyway. That is unusual, and it tells you the score is not what Meta is selling.
Both the 82.9% and the 86.7% come from Meta's own evaluation harness. Neither model has an independently verified entry on the public Terminal-Bench leaderboard.
Meta's methodology report is unusually candid about this. It ran each model inside that model's own vendor product — Muse Code for Muse Spark 1.2, Claude Code for Opus, Codex for GPT, Grok Build for Grok — scoring pass@1 averaged over five attempts across all 89 tasks in the official 2.1 release, in isolated Daytona cloud sandboxes. It then states outright that its agent tools and system prompts "may not be specifically tuned for proprietary third-party models."
That is honest disclosure. It is also disclosure in a linked PDF, while the number sits on the chart.
There is a precedent worth knowing. At the Muse Spark 1.1 launch in July, Meta published 80.0 on Terminal-Bench 2.1. The verified leaderboard entry for the same model on the same benchmark version — run with mini-swe-agent at xhigh effort, verified by the Terminal-Bench team on July 9, 2026 — is 76.2% ± 1.2%. The gap is 3.8 points, and Meta's published figure sits 2.6 points above the upper bound of the verified interval.
The highest score any model has posted on the verified Terminal-Bench 2.1 board is 83.8%, by Claude Fable 5 running in Claude Code. Claude Opus 5 has no verified entry at all.
Meta's comparison set — Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, Kimi K3 — is also the second tier at two of the three US frontier labs. Claude Fable 5 and GPT-5.6 Sol do not appear on the coding charts. Developers noticed within the hour.
The contributor tier is the actual news
Muse Spark 1.2 ships on two pricing tiers, and the spread between them is the most strategically significant thing in this release.
| Tier | Input / 1M | Cached input / 1M | Output / 1M | Your prompts |
|---|---|---|---|---|
| Standard | $1.25 | $0.15 | $4.25 | Not used to improve Meta's products |
| Contributor | $0.10 | $0.002 | $0.20 | Used to train future Meta models |
That is 12.5× cheaper on input, 21× cheaper on output, and 75× cheaper on cached input. Meta chief AI officer Alexandr Wang described the contributor tier to CNBC as more than ten times cheaper than the pay-as-you-go tier.
Read that table as a business model rather than a price list. Meta is not competing to have the best coding model. On the verified board, its previous model was already among the cheapest ways to solve a Terminal-Bench task while scoring below Claude Opus 4.8 on accuracy. Meta is competing on cost per finished task — and it is now willing to accept payment in your repository.
The decision this forces is not about price. It is about what your codebase is worth.
- Open-source repo, side project, throwaway prototype? The contributor tier is close to free compute. Take it.
- Client work under NDA, proprietary product code, anything with credentials in the tree? Standard tier or nothing. A 21× saving does not cover a contractual breach, and "I didn't read the tier description" is not a defense.
- Employed somewhere with a code-governance policy? This is a procurement decision, not a developer decision. It goes past legal before it goes into your terminal.
To Meta's credit, the trade is explicit. Plenty of tools make the same bargain in a terms-of-service paragraph nobody opens.
The failure mode nobody is testing
Meta's own kernel-optimization case study runs Muse Spark 1.2 through 1,000+ tool calls over as long as 24 hours, iteratively writing, compiling, and profiling Triton kernels for KDA and MLA on NVIDIA Hopper GPUs. The MLA run benchmarks against a PyTorch reference at batch size 1, 64 heads, sequence length 8192, and latent dimension 512. Impressive engineering. It also describes exactly the workload where parallel subagents get dangerous.
Picture it: one subagent validates its change at hour two and gets a green result. Another lands a conflicting change at hour five. The first agent is still holding a passing verdict computed against a tree that no longer exists. Agents then merge on expired proofs, and the run looks correct right up until it isn't.
Verification results need to carry the state they were checked against. Meta's materials do not say whether they do. If you trial Muse Code, hand it a genuinely conflicting multi-agent task and check whether validation is re-run against the merged tree or trusted from when it was computed.
One underreported detail deserves a mention: Muse Spark 1.1 was US-only. 1.2 ships with expanded global access through the Meta Model API. For developers who have been locked out entirely, that is the most consequential line in the announcement, and almost every outlet buried it.
The Bottom Line
Muse Code is a well-designed harness attached to a vendor-benchmarked model and a pricing scheme that is quietly a data-acquisition strategy. The event log and persistent subagents are good engineering worth borrowing regardless of which harness you use. The 82.9% is a claim about Meta's evaluation setup, not about the field, and Meta's last model came in 3.8 points below its own published figure when someone else ran it.
If you are productive in Claude Code or Codex, nothing here justifies switching your primary harness today. Wait for a verified Terminal-Bench 2.1 entry. If you have been geo-blocked from Muse Spark until now, or you ship open source at volume, install it this week — on a public repo, the contributor tier costs you nothing you had.
Just know which tier you clicked.
