SWE-2 Flash vs Muse Code: The Harness Is the Thing You Are Buying

Put the two numbers side by side and the story writes itself: SWE-2 Flash at 89.2 on Terminal-Bench 2.1, Muse Code at 82.9, therefore Cognition's model is 6.3 points better than Meta's agent. Every clause of that sentence is doing something questionable. SWE-2 Flash is a model served inside Devin. Muse Code is a terminal agent built around Muse Spark 1.2, co-trained with it, and its 82.9 was measured with the harness and the effort setting attached. Comparing them is comparing a model to a product, and the product is the part you would actually install.

That does not make the comparison useless. It changes what the interesting question is: not which model scores higher, but which of the two closed loops does more of your work for you, at what price, and with how much of the method disclosed.

What each thing actually is

Muse Code is Meta's first coding agent, launched 5 August 2026 for macOS and Linux. It is a CLI in the same category as Claude Code or Codex: it plans changes, edits files, runs tests and validates results across a repository. The design choices that distinguish it are architectural rather than cosmetic.

• Background agents — persistent async agents that stay alive across a session instead of being spawned per task, so information gathered once is not re-gathered; several can run in parallel in isolated work trees without touching the working copy.

• Event log — a local record of every model call, tool run, approval and edit, which makes a session replay-exact and restart-safe after a crash — the mechanism that makes multi-hour tasks survivable.

• Skills — built-in slash commands including a plan step, a plan-stress-test step and a goal-driving step.

• Model — Muse Spark 1.2, closed-weight, 1M-token context, $1.25 per million input tokens and $4.25 per million output.

SWE-2 Flash is not an agent. It is a model, post-trained from Inkling-Small at 276B total and 12B active parameters, served inside Devin Desktop and the Devin CLI. Cognition calls it three times faster and one fifth the price of SWE-2; access is free through 30 November 2026; there is no public API, no model ID and no downloadable weight file.

So the real shape of this matchup is: two closed systems, each pairing a model with a harness its maker controls. One publishes its harness and effort setting. The other does not.

The DeepSWE row that says they are tied

DeepSWE 1.1 is the one benchmark both reports where neither side is straining the definition of the task. SWE-2 Flash: 60.1. Muse Spark 1.2 inside Muse Code: 59.3. Eight tenths of a point.

That is not a gap. It is the same score reported twice, and the more interesting fact is that the two systems arrive there from opposite directions. Meta's model is a general flagship with 1M context and a broad tool-use profile; Cognition's is a 12B-active post-train of an open-weight base, explicitly optimised for the cheap end of the agent market. Converging on DeepSWE is a stronger result for the cheaper architecture than the raw table suggests.

On Terminal-Bench 2.1 the gap is larger — 89.2 to 82.9 — and here the methodology asymmetry bites. Meta states that its figure was produced at xhigh effort, inside the co-trained Muse Code harness, in isolated Daytona sandboxes, with pass@1 averaged over five attempts. Cognition states none of those things. Its September post for SWE-2 carried a methodology appendix; the 4 November post for SWE-2 Flash does not.

There is also a warning inside Meta's own numbers that applies to every table in this space. Meta reports Muse Spark 1.1 at 76.2 on Terminal-Bench 2.1, while the public tbench.ai leaderboard carries a verified 80.0 for the same model. The difference is that a row is not a model — it is a model plus a harness plus an effort setting. If a 4-point swing can appear between two reports of the same model, a 6.3-point gap between two differently-instrumented systems is not a decisive result. All figures in this section are vendor-reported.

One more thing to hold in view. Muse Spark 1.2 is no longer the newest model Meta has: Artificial Analysis now marks the 1.2 checkpoint deprecated and points readers to Muse Spark 1.3. Muse Code launched on 1.2, and 1.2 is the checkpoint every number in this article describes and the one still callable.

Context: 1M published against unpublished

Muse Spark 1.2 ships with a 1M-token context window, and Muse Code's architecture is built to use it — persistent agents, a durable event log, work that runs for hours. That combination is the product's actual pitch: not a better single edit, but an agent that can hold a large repository and a long task without losing the thread.

SWE-2 Flash publishes no context window at all. For a model delivered only inside an agent, the omission is less damaging than it would be for an API model, because Devin manages context on your behalf. It is still an omission: you cannot size a workflow around a budget you are not told.

A rate card against a free month

• Muse Spark 1.2 — $1.25 per million input tokens, $4.25 per million output. A discounted contributor tier exists in select countries in exchange for training-data use, and zero-data-retention requests are accepted.

• SWE-2 Flash — no per-token price. Free inside Devin through 30 November 2026, then unspecified.

• Muse Code — the agent itself, priced through model usage rather than a separate subscription in the material Meta has published.

An unpriced product is not a cheap product; it is an unknown one. If you are building a cost model for the next quarter, $1.25/$4.25 is a number you can put in a spreadsheet and SWE-2 Flash is a question mark with a deadline attached.

What you can call, and what you cannot

Here the two systems diverge in a way that has nothing to do with benchmarks. SWE-2 Flash cannot be called. There is no endpoint to point a client at, no model ID, and no date on which that changes. If your architecture needs a coding model behind an HTTP call rather than inside a desktop IDE, SWE-2 Flash is simply not a candidate this month.

Muse Spark 1.2 is a different story. The Meta side of this matchup is callable today — both Muse Spark 1.2 and 1.1 are live in the OrcaRouter catalogue, at provider list price with 0% markup, so the $1.25/$4.25 rate above is the rate you pay rather than a reseller number. That matters for this specific decision: you can run Muse Spark 1.2 against your own repository tonight and find out whether the 59.3 on DeepSWE describes your work. You cannot run the equivalent test on SWE-2 Flash without living inside Devin for the free month.

OrcaRouter also gives you the failover that a single-vendor bet lacks. If you standardise on the Meta side and the provider degrades, automatic failover keeps the call alive — which is the practical difference between "we chose a model" and "we chose a supply chain".

Which one to pick

Choose Muse Code if you want an agent you can drive from a terminal, with published effort settings, a durable event log, parallel isolated work trees and an API behind the model. The DeepSWE tie says you are not giving up capability at the frontier; the Terminal-Bench gap says you may be giving up some ceiling, on a number whose method you can at least inspect.

Choose SWE-2 Flash if you are a Devin user and the question is whether the faster, cheaper model inside the IDE you already pay for is good enough for your repositories. The free window to 30 November is a real trial, and 60.1 on DeepSWE says it will not embarrass itself.

What neither choice gets you is certainty. Both vendors measured their own systems with their own harnesses, and the one public data point that lets you check a vendor claim — tbench.ai's 80.0 for Muse Spark 1.1 against Meta's 76.2 — shows the gap between a published row and a reproduced one can be as large as the gap between the two products in this article.