Muse Code: Meta Shipped a Coding Agent That Loses on Its Own Benchmark

Vendors do not usually publish the chart where they come second. Meta published three of them. In the launch deck for Muse Code — the terminal coding agent Meta Superintelligence Labs released on August 5, 2026, running on the same-day Muse Spark 1.2 — Claude Opus 5 is ahead on every board Meta chose to show, including the one Meta built and controls itself. On Terminal-Bench 2.1 the gap is 3.8 points. On DeepSWE 1.1 it is 5.7, with GPT-5.6 Terra also ahead. On Meta's own internal coding benchmark it is 8.8. The viral version of this launch — "59% on DeepSWE 1.1, outperforming Grok 4.5 and Gemini 3.6 Flash" — is true, and it is the narrowest true statement available.

Which makes the launch more interesting, not less. Meta shipped a benchmark deck it does not win and shipped it anyway, because the score is not what it is selling. What it is selling is a runtime that survives a 24-hour job, and a price that undercuts the field by more than an order of magnitude — as long as you are willing to let Meta read your repository. Every benchmark figure below comes from Meta's own charts, which were published without methodology; where an independent measurement exists, the sentence says so.

What Meta actually shipped

Two things on the same day, deliberately coupled.

• Muse Code — a terminal coding agent, public beta, installed with one line: curl -fsSL https://dev.meta.ai/install.sh | bash. macOS and Linux only; there is no native Windows build, no GUI and no IDE extension. Meta describes its job as taking on complex software engineering across large repositories — planning changes, writing code, validating results.

• Muse Spark 1.2 — a coding-focused revision of Meta's reasoning model, 1,048,576-token context, available inside Muse Code and through the Meta Model API in public preview. Meta says it scaled up training compute on coding tasks and widened the diversity of training environments, with the target being long-horizon work: whole-repository generation, multi-file refactors, extended debugging.

The coupling is the claim. Meta states the two were co-trained, and that the pairing therefore produces "better tool use, fewer retries, and higher-quality output than a generic wrapper around an outside model." That is Meta's phrasing and Meta's assertion; no third party has tested the model inside a rival harness to check whether the co-training premium is real or whether Muse Spark 1.2 simply behaves like a competent coding model wherever you put it.

Three bundled skills ship with the agent, and they describe the intended working style better than the marketing copy does: /plan produces an approval-gated plan before anything is written, /grill stress-tests that plan, and /goal pushes toward completion of a stated objective. Plan, attack the plan, then execute — which is the pattern experienced Claude Code and Codex users converged on manually, now shipped as first-class commands.

The benchmark deck, read honestly

Meta published comparisons on Terminal-Bench 2.1, DeepSWE 1.1, its own Meta Internal Coding Bench and GDPval. The numbers, as reported by Meta:

• Terminal-Bench 2.1 — Claude Opus 5 at max effort in Claude Code 86.7%, Muse Spark 1.2 in Muse Code 82.9%, GPT-5.6 Terra in Codex 81.8%, Grok 4.5 in Grok Build 81.6%. Second place, and second through fourth are separated by 1.3 points.

• DeepSWE 1.1 — Claude Opus 5 65.0%, GPT-5.6 Terra 64.8%, Muse Spark 1.2 59.3%. Third place, and the only board where the gap is decisive rather than cosmetic.

• Meta Internal Coding Bench — Claude Opus 5 79.4%, Muse Spark 1.2 70.6%. Meta's home fixture, and Meta's widest loss.

• Generation over generation — Meta reports +6.7 points on Terminal-Bench and +6.3 on DeepSWE against Muse Spark 1.1. Those are the largest numbers in the deck, and they are the ones a Muse Spark 1.1 user should actually care about.

One arithmetic snag is worth naming, because it is the reason to treat all of the above as directional. Back out Meta's +6.7 and Muse Spark 1.1 lands near 76% on Terminal-Bench 2.1 — but the public tbench.ai leaderboard has carried a verified 80.0% for Muse Spark 1.1 since late July. Both numbers can be honest and still disagree, because a Terminal-Bench row is not a model score. It is a score for a harness plus a model plus an effort setting, and Meta's internal harness for 1.1 was not the one tbench ran. That single fact does more damage to cross-vendor benchmark comparison than any individual disputed figure, and it leads directly to the next section.

The runtime is the product

Strip out the scores and what remains is an engineering pitch about not dying. Two mechanisms carry it.

The first is the event log. Muse Code appends every model call, every tool run, every approval and every edit to a local log, which Meta describes as making the session replay-exact and restart-safe: kill it, crash it, lose the machine, and the agent resumes precisely where it stopped rather than re-deriving its context from scratch. Competitors have crash recovery of varying quality; none of them have made it the headline. It is also, incidentally, an audit artifact — a complete local record of what an autonomous process did to your working tree, which is exactly the document a security review asks for and which most agent CLIs cannot produce.

The second is what Meta calls async background agents. The distinction from ordinary subagents is that they are not spawned per subtask and torn down; they stay alive across the session, carry out next steps on their own, and choose when to report back to the main loop. The claimed payoff is lower latency, because the orchestrator is not paying spin-up cost every time it delegates.

Meta's supporting evidence is one case study: Muse Code running for 24 hours across more than 1,000 tool calls, autonomously optimizing GPU kernels on NVIDIA Hopper hardware, with what Meta calls substantial improvements over the provided baseline implementation. No speedup figure was published, so the interesting quantity is not the result but the duration. A thousand tool calls without human rescue is a different failure surface from a fifty-call refactor, and it is a workload where a 3.8-point benchmark deficit matters far less than whether the process is alive at hour nine.

The price is the weapon — and part of it is paid in source code

There is no Muse Code subscription. Billing is tokens, on one of two tiers, and the spread between them is the most aggressive thing in this launch:

• Standard — $1.25 per million input tokens, $0.15 per million cached input, $4.25 per million output. Prompts on this tier are not used to improve Meta's products.

• Contributor — $0.10 per million input, $0.002 per million cached input, $0.20 per million output. In exchange, Meta may use the data to improve its models.

That is 12.5x cheaper on input, 21x cheaper on output, and 75x cheaper on cached input. Put it through a realistic agent step — 60,000 tokens of repository context in, 3,000 tokens of plan-and-patch out — and the four corners of the pricing grid are startling. Standard tier with a cold prompt: 8.8 cents a step, $87.75 for a thousand steps. Standard tier with the repository context hitting the cache: 2.2 cents, $21.75 per thousand. Contributor tier cold: 0.66 cents, $6.60 per thousand. Contributor tier with a warm cache: $0.72 for the same thousand steps.

The same work, at 122x the cost or 1/122th of it, depending on two choices. Meta's own 24-hour, 1,000-call kernel run costs roughly ninety dollars of tokens on the tier that protects your data, and under a dollar on the tier that does not.

Which is the actual decision this launch puts in front of teams, and it is not a pricing decision. A terminal coding agent reads your codebase — that is its entire function — so the prompts on the contributor tier contain your source, your internal APIs, your comments about why the workaround exists, and whatever test fixtures your repo happens to hold. For a side project or an open-source tree, the contributor tier is close to free money. For a proprietary codebase, or any repository within reach of customer data, it is a licensing decision wearing a discount's clothing, and it belongs in front of whoever signs off on data handling rather than in front of whoever watches the cloud bill. Meta priced it at a 12x discount because the data is worth at least that much to Meta.

What you can't get through the API

The 82.9% belongs to Muse Spark 1.2 inside Muse Code. If you call the model from your own agent framework, you have bought the model half of that pairing and none of the harness half — no event log, no persistent background agents, and, if Meta's co-training claim means anything, none of the tuned tool-use behaviour either. So there are two separate evaluations here and conflating them is the easiest mistake to make this week: does Muse Code beat the agent you already run, and is Muse Spark 1.2 a better model than what you currently call?

The second question is the cheaper one to answer, because it only requires holding your harness constant and swapping the model underneath it. Worth being precise about availability: Muse Spark 1.2 is served by Meta directly and is not in the OrcaRouter catalog. Muse Spark 1.1 is, at $1.25 and $4.25 — Meta's list price exactly, because OrcaRouter adds 0% markup and passes provider pricing straight through, which is also why a vendor price change shows up on our side the day it happens rather than after a contract cycle. Our seven-day production telemetry for 1.1 shows p50 time-to-first-token at 1.84 seconds and p95 at 6.00 seconds, which is a more useful latency expectation than any benchmark harness will give you.

The practical value of a single gateway in a week like this one is that the comparison set — Claude Opus 5, GPT-5.6 Terra, Grok 4.5, Gemini 3.6 Flash and Muse Spark 1.1 — sits behind one key at list price, so an A/B is a string change rather than a procurement exercise. It also covers the structural risk in Meta's offer: Muse Spark 1.2 has exactly one provider, which makes it a single point of failure by construction. Automatic failover across providers is the difference between a bad afternoon for Meta being a bad afternoon for your agent, or not.

This is also why 1.2 shipped silent and slow

Muse Spark 1.2 appeared on Meta's API on August 5 with no announcement and no changed numbers — same 1M context, same $1.25 and $4.25, same five-level reasoning dial as Muse Spark 1.1. The one thing that moved was measured from outside: Artificial Analysis clocked time-to-first-token at 26.12 seconds against 1.1's 2.90 seconds at maximum reasoning effort, buying three points of Intelligence Index, 51 to 54. Read as a general-purpose model release, that is a strange trade — nine times the wait for three points.

Read as the model half of a co-trained agent, it is the obvious trade. Nobody waiting on a twelve-file refactor notices twenty-six seconds of planning; the person waiting on a chat completion notices all of it. Meta was not shipping a chat model quietly and hoping nobody measured the latency. It was shipping the engine before the car, and the announcement arrived when the car did. Meta's Muse Spark developer page now reads 1.2 as well.

The corollary is a warning for anyone who adopted this family for its breadth. Muse Spark 1.1 was sold as a multimodal reasoning model — text, images, video, audio and PDFs, good for mixed-media research. Version 1.2's positioning is software work, and its launch material is a coding agent. The multimodal surface has not been removed, but it is no longer what Meta is optimising or advertising, and nobody outside Meta has published a video or audio result for 1.2.

Where Muse Code is still thin

• It is beta, and labelled as such. Nothing about the event-log guarantee has been verified by anyone outside Meta.

• macOS and Linux only. Windows developers are on WSL or nothing, which is a real constraint for enterprise fleets rather than an inconvenience.

• Terminal only. No GUI, no IDE integration, and no cloud-hosted parallel agents of the kind Codex already offers. Meta's launch material does not mention MCP support.

• No published methodology behind any of the benchmark charts, and no independent evaluation yet of either the agent or Muse Spark 1.2's coding sub-scores. Artificial Analysis has a composite index for 1.2 but not the per-benchmark coding and agentic breakdown it publishes for 1.1 — and 1.1's agentic score was its weakest dimension by a wide margin, which makes it precisely the number that would settle this release.

• Meta is late. Claude Code and Codex have a year of habit, plugin ecosystems and team workflows built around them. A crash-safe event log is a good reason to try something; it is not by itself a reason to move a team.

Questions people are actually asking

Is Muse Code free?

No, but there is no subscription either — unlike the Claude Code and Codex plans it competes with, Muse Code bills purely per token through the Meta Model API. Cost therefore scales with how much repository context your sessions push, not with headcount, which favours occasional heavy users and penalises always-on agents. The contributor tier is close enough to free that price is not a real barrier to trying it.

Does the contributor tier train on my private code?

Meta's terms for that tier are that the data may be used to improve its models, and a coding agent's prompts contain your code. Meta states that standard-tier prompts are not used to improve its products; that is the tier to use on anything proprietary. Treat the choice as a data-handling decision with a discount attached, not a billing preference — and note that the discount applies per token, so the incentive to quietly downgrade grows exactly as your usage does.

Can I use Muse Spark 1.2 with Claude Code or my own agent?

The model is available through the Meta Model API independently of Muse Code, so yes for anything that accepts a custom endpoint. What you will not get is the co-trained pairing Meta credits for its benchmark results, and the honest expectation is that a model run outside the harness it was tuned with scores below the number on the chart. If the goal is to evaluate the model rather than the agent, run it against Claude Opus 5 and GPT-5.6 Terra in your own harness and ignore the deck entirely.

Who should install it this week

If you run long-horizon autonomous jobs — migrations, dependency upgrades across a monorepo, anything you would currently babysit because the agent loses its place — Muse Code is worth an afternoon on a scratch clone. The event log and persistent background agents are aimed exactly at that failure mode, and the benchmark deficit is smallest where task duration is longest.

If you are on an open-source or non-sensitive codebase and cost is your binding constraint, the contributor tier is the most interesting number in the coding-agent market right now, and 0.66 cents a step is not a typo.

If you are running a team on Claude Code or Codex against a proprietary repository, there is nothing here to force a move yet. Muse Code is second or third on every board its own vendor selected, the standard tier prices at parity with the field rather than below it, and the differentiator is a reliability property nobody outside Meta has tested. The thing to watch is not the next benchmark. It is whether the event log holds up when somebody who does not work at Meta runs a thousand tool calls through it.