Muse Code vs Muse Spark 1.1: What Meta's 6.7-Point Upgrade Is Actually Measuring

Do the subtraction first. Meta's launch deck for Muse Code, the terminal coding agent it shipped on August 5, 2026, reports 82.9% on Terminal-Bench 2.1 and credits that as a 6.7-point gain over Muse Spark 1.1, the model most people reading this are already paying for. Take 6.7 away from 82.9 and you get 76.2 — which is not a number Meta published, but is exactly the score sitting on the public Terminal-Bench 2.1 leaderboard for Muse Spark 1.1 at xhigh effort, submitted in July, run by Princeton's mini-SWE-agent harness.

That coincidence is the most useful fact in this comparison, because mini-SWE-agent is a deliberately minimal scaffold — a few hundred lines, no persistent subagents, no event log, no planning skills. Meta's 82.9% is Muse Spark 1.2 running inside a purpose-built agent that Meta co-trained the model with. If those two rows are what the 6.7 points span, then the headline generational gain is a model upgrade and a harness upgrade bundled into one number, and Meta has published nothing that separates them.

Meta did not state which harness it used for its 1.1 baseline, so the match is suggestive rather than proven. But it reframes the decision in front of anyone currently calling Muse Spark 1.1 through an API, which is the honest version of this matchup: you are not choosing between two models. You are choosing between a model you drive yourself and an agent that drives a newer model for you, and only one of those two things has been measured by somebody other than Meta.

These are not the same category, and the spec sheet says so

Half the confusion around this pairing comes from treating Muse Code as a model release. It isn't. It's a CLI, and Muse Spark 1.2 is what it calls.

• What it is — Muse Code is a terminal agent installed with one line (curl -fsSL https://dev.meta.ai/install.sh | bash); Muse Spark 1.1 is a model endpoint you call from your own code, your own agent framework, or any CLI that accepts a custom base URL.

• Underlying model — Muse Code runs Muse Spark 1.2, co-trained with the agent; Muse Spark 1.1 is the July generation, still served and still listed.

• Where it runs — Muse Code is macOS and Linux only, terminal only, no GUI, no IDE extension; Muse Spark 1.1 runs wherever HTTPS does, including Windows.

• Inputs — Muse Code takes a repository and a prompt; Muse Spark 1.1 accepts text, images, video, audio and PDF documents in a single 1,048,576-token context and returns text.

• Status — Muse Code is public beta and labelled as such; Muse Spark 1.1 has been generally available since July 2026 and is carrying real production traffic.

• List price — both bill through the Meta Model API at $1.25 per million input tokens, $0.15 cached input and $4.25 output on the standard tier. The rate card is the same rate card.

The last line is the one that surprises people, and it kills the obvious assumption. Adopting Muse Code does not cost more per token than what you already run. What it costs is measured in other currencies — platform support, latency, modality, and, on the cheap tier, your source code. Those are the four axes worth arguing about.

The benchmark where they diverge, read with the harness attached

Meta published two generational deltas against Muse Spark 1.1: +6.7 points on Terminal-Bench 2.1 and +6.3 on DeepSWE 1.1. Both are vendor-reported, both come from charts published as images without a methodology write-up, and neither has been reproduced by a third party. Back them out and Meta's implied 1.1 baselines are roughly 76.2% and 53.0%.

Now put Meta's 82.9% next to the board it did not cite. The public Terminal-Bench 2.1 leaderboard, at the time of writing, has Claude Code paired with Claude Fable 5 at 83.8% ± 1.2%, Codex with GPT-5.5 at 83.1% ± 1.1%, Terminus 2 with Claude Fable 5 at 80.4%, Cursor CLI with Grok 4.5 at 79.3%, and mini-SWE-agent with Muse Spark 1.1 at 76.2% in eighth place. Muse Code's self-reported 82.9% would slot into third — below two agent-model pairs that appear nowhere in Meta's deck, which instead compares against Claude Opus 5 at 86.7%, GPT-5.6 Terra at 81.8% and Grok 4.5 at 81.6% in separate runs at maximum effort.

Two boards, one benchmark, numbers that do not line up — and a third scale entirely if you look at Artificial Analysis, whose independent Terminal-Bench v2.1 evaluation tops out near 89.5% for GPT-5.6 Sol against a public-leaderboard ceiling of 83.8%. None of these parties is lying. A Terminal-Bench row is a score for a harness, plus a model, plus an effort setting, plus a compute allowance, and swapping any one of those moves the result by more than the entire spread Meta is claiming as a generation.

Which is why the interesting figure on the public board is not the accuracy column. It's the cost column. Muse Spark 1.1 finished its 76.2% run for $198.05. Claude Code with Claude Fable 5 bought its 83.8% for $552.67, and Codex with GPT-5.5 paid $2,059.19 for 83.1%. That is 7.6 points of accuracy for 2.8x the money in the first case and roughly 10x in the second — measured by the same operator, on the same tasks, under the same rules, which is exactly the apples-to-apples that Meta's chart does not give you. The board also deducts for tasks solved by gaming the environment; Muse Spark 1.1's submission shows no such deduction, while Cursor CLI's Grok 4.5 run carries a 9-point one.

Meta published no cost figure for its 82.9%.

What Muse Spark 1.1 already does that Muse Code cannot

The pitch for Muse Code is that a co-trained agent beats "a generic wrapper around an outside model" — Meta's phrasing, Meta's assertion, untested by anyone else. It's a reasonable claim. It is also aimed at a gap that Muse Spark 1.1 was specifically built to close.

Muse Spark 1.1 speaks the Model Context Protocol natively, handles its own MCP connections without an external framework, orchestrates primary-agent and subagent roles internally, and generalises zero-shot to new tools and custom skills. Meta's own launch numbers for it were tool-use numbers: 88.1 on MCP Atlas for scaled tool use — ahead of the 82.2 it reported for Claude Opus 4.8 and 75.3 for GPT-5.5 — and 54.7 on JobBench. All vendor-reported, all from July, none independently reproduced. The point stands anyway: 1.1 is not a chat model somebody bolted an agent onto. Drop it into OpenCode, Cline, or any CLI that takes a custom endpoint and you have an agent today.

And then there is everything Muse Code structurally cannot touch. Muse Spark 1.1 reasons over images, video, audio and PDFs in one context window. A terminal coding agent has no use for that surface and no way to expose it. If your workload is a design mock turned into components, a support call transcribed and triaged, or a 400-page spec cross-referenced against an implementation, the newer thing is the less capable thing — and Meta's positioning for the 1.2 generation moved from "multimodal deep research over mixed media" to multi-file refactors and whole-repository generation, with no published video or audio result for 1.2 from anyone.

The genuine asymmetry runs the other way, and it's a runtime property rather than a capability. Muse Code appends every model call, tool run, approval and edit to a local event log, which Meta says makes a session replay-exact and restart-safe: crash it and the agent resumes where it stopped instead of re-deriving its context. Its background agents persist across the session rather than being spawned and torn down per subtask. Meta's evidence is a single case study — a 24-hour, 1,000-plus-tool-call run optimising GPU kernels on NVIDIA Hopper hardware. No speedup number was published, so the duration is the claim. If you have ever lost a nine-hour migration because the harness forgot what it was doing at step 300, that is a real thing to want, and no amount of MCP support in the model gives it to you.

Same list price, until you read the second tier

Because the standard tier is identical on both sides, the pricing argument in this matchup lives entirely in the tier Meta introduced alongside Muse Code. The contributor tier runs $0.10 per million input tokens, $0.002 cached input and $0.20 output — 12.5x cheaper on input, 21x on output, 75x on cached input — in exchange for explicit permission to use your prompts and completions to train future Meta models. Standard-tier traffic, Meta states, is not used that way.

Run a realistic agent step through it — 60,000 tokens of repository context in, 3,000 tokens of plan-and-patch out — and a thousand steps costs $87.75 on the standard tier cold, $21.75 with the repo context hitting cache, $6.60 on contributor cold, and $0.72 on contributor with a warm cache. Meta's own 24-hour kernel run lands near ninety dollars on the tier that protects your code and under a dollar on the tier that does not.

That is not a billing choice. A coding agent's prompts are your codebase — internal APIs, the comment explaining why the workaround exists, whatever your fixtures happen to contain. On an open-source tree the contributor tier is close to free money. On a proprietary repository it is a data-licensing decision wearing a discount's clothing, and it belongs in front of whoever signs off on data handling rather than whoever watches the cloud bill. Meta set the discount at 12x because the data is worth at least that much to Meta.

Staying on Muse Spark 1.1 sidesteps the question entirely, and it is worth knowing exactly what that costs and where. Muse Spark 1.1 is in the OrcaRouter catalog at $1.25 and $4.25 — Meta's list price to the cent, because we add 0% markup and pass provider pricing straight through, which is also why a Meta rate change appears on our side the day Meta makes it rather than after a contract cycle. Our seven-day production telemetry for it shows p50 time-to-first-token at 1.84 seconds and p95 at 6.00 seconds, which is a far more useful expectation than any benchmark harness will hand you. Muse Spark 1.2 is not in our catalog; it is served by Meta directly and nowhere else, so the newer half of this comparison currently has exactly one provider and no failover path by construction. If you are evaluating it, that single-supplier exposure is part of what you are evaluating.

The latency trade nobody mentions in the launch coverage

Artificial Analysis, measuring independently at each model's maximum reasoning effort, clocked Muse Spark 1.1's time-to-first-token at 2.90 seconds and Muse Spark 1.2's at 26.12 seconds. Output speed fell from 213.5 to 165 tokens per second. The purchase was three points of Intelligence Index, 51 to 54, and a jump from #22 to #13 of 185 models.

Read as a general-purpose model release that is a bad trade — nine times the wait for three points. Read as the engine of a coding agent it is obviously the right one, because nobody waiting on a twelve-file refactor notices twenty-six seconds of planning, and everybody waiting on a chat completion notices all of it. That is the clearest statement of who each side of this comparison is for, and it was measured by a third party rather than asserted by Meta.

It also means the switch is not free even if you only care about code. Anything interactive you had pointed at Muse Spark 1.1 — inline completion, a review bot, a chat surface over your docs — gets materially worse if you move it to the 1.2 generation at high effort. The upgrade is targeted, and targeting has a direction.

Which one you should actually be running

Stay on Muse Spark 1.1 if your work is not a repository. Multimodal pipelines, long-context analysis, mixed-media research, anything interactive, anything on Windows, anything already wired through MCP and working. Nothing in the Muse Code launch improves those, and the latency measurement makes at least one of them worse.

Try Muse Code if your failure mode is duration. Monorepo migrations, dependency upgrades, week-long refactors — jobs you currently babysit because the agent loses the plot. The event log and persistent background agents target exactly that, and a benchmark deficit of a few points matters least where the run is longest. Beta, on a scratch clone, on a repository you would be comfortable losing.

Do the honest experiment if you are trying to decide on the model. Hold your harness constant and swap Muse Spark 1.1 for Muse Spark 1.2 underneath it. That isolates the one variable Meta's chart bundles, and it is the only way to find out whether the co-training premium is real or whether 1.2 is simply a competent coding model wherever you put it. A single gateway makes this a string change rather than a procurement exercise — Muse Spark 1.1, Claude Opus 5, GPT-5.6 Terra, Grok 4.5 and Claude Fable 5 all sit behind one OrcaRouter key at list price with automatic failover across providers, so the comparison set costs you nothing to keep warm. Muse Spark 1.2 is the exception and has to come from Meta.

Questions worth asking before you install anything

Can I point Muse Code at Muse Spark 1.1 instead of 1.2?

Meta's launch material does not say, and the two were shipped as a co-trained pair, which is an argument against it being supported or useful. The reverse direction is well documented though: Muse Spark 1.1 works in any agent that accepts a custom endpoint, which is how most people are running it as an agent today. If your goal is a controlled comparison, run 1.1 and 1.2 in your harness rather than trying to bend Meta's.

Does the contributor tier apply to Muse Spark 1.1 as well?

Meta introduced the two-tier structure with the Muse Code and Muse Spark 1.2 launch, and its published rate card attaches the contributor pricing to that release. Assume the 12x discount is a 1.2-generation offer until Meta says otherwise, and price any 1.1 workload at $1.25 and $4.25. If you are on OrcaRouter you are on list price by definition, so there is no data-sharing tier to opt into or out of.

Is the 6.7-point claim wrong, then?

No — it is unaudited and under-specified, which is different. Meta may well have run its 1.1 baseline in a harness comparable to Muse Code's, in which case the gain is a clean model-over-model result. But no methodology was published, the implied baseline lands precisely on a minimal third-party scaffold's score, and the same benchmark yields three different scales depending on who runs it. Treat +6.7 as directional and get your own number before you plan around it.

What would settle this

One submission. If Meta puts Muse Code on the public Terminal-Bench 2.1 leaderboard under the same rules everyone else submitted under — no modified timeouts or resources, cost column filled in, hack deductions applied — the 82.9% becomes comparable to the 76.2% next to Muse Spark 1.1's name and to the 83.8% at the top, and this entire argument resolves in an afternoon. Until then the only independently verified number in the matchup belongs to the older model, and it says Muse Spark 1.1 solved three-quarters of Terminal-Bench 2.1 inside a scaffold small enough to read in one sitting, for under two hundred dollars.

The second thing worth watching is narrower and arrives sooner: whether Artificial Analysis publishes the coding and agentic sub-scores for the 1.2 generation that it already publishes for 1.1. Agentic performance was 1.1's weakest published dimension by a wide margin. It is precisely the number Meta claims to have fixed, and precisely the one nobody outside Meta has yet measured.