Meta's Muse Spark 1.3 review: strong code and pricing, weak skills and sandbox

A developer has published a hands-on review of Muse Spark 1.3, described as Meta's latest frontier-level model, together with muse, the LLM harness the company ships for developers. Writing as gosukiwi on dev.to, the reviewer's verdict is mixed: the model itself is competitive on intelligence, readability and speed, and the subscription is good value, but the surrounding product — skill execution, output formatting and the sandbox — still needs work.

Skill execution is the weak spot

The sharpest criticism concerns skills, the scripted workflows the harness can run. According to the review, when a skill is flagged with disable-model-invocation, Muse Spark 1.3 sometimes refuses to launch it even when the developer calls it manually. The reviewer suspects this happens when the skill is invoked mid-sentence, but reports the behavior is not consistent.

More broadly, the model is described as weaker than its peers at following a skill's instructions and prone to getting confused. As a test case, the reviewer uses a skill that takes a GitHub issue all the way to a pull request: it first interrogates the issue, then autonomously plans, implements using TDD, reviews the code and opens the PR. The same skill reportedly runs perfectly on Claude (Opus 5) and Cursor (Grok 4.6). Under Muse Spark 1.3, the workflow stalls between steps; typing "continue" is enough to get it moving again, which works but is annoying in daily use.

Output formatting is a second irritation. Responses sometimes render as raw markdown and sometimes do not, with no consistent pattern.

The sandbox is too rigid

The harness ships with a sandbox, which the reviewer considers a good idea in principle but too restrictive in practice. It blocks running files outside the workspace and forbids external ports, which rules out workflows such as launching Firebase emulators. There is no mechanism to add exceptions or otherwise configure the sandbox.

That leaves two unappealing options: approve the same permission prompts repeatedly — the reviewer says the model itself can get stuck asking for the same permission over and over until it effectively does nothing — or fall back on --yolo mode and drop the guardrails.

What works

On raw capability the picture improves considerably. The reviewer subjectively rates the model's general and coding intelligence at around the level of Opus 4.8 and Grok 4.6, and says the code it produces is very similar to what other frontier models generate — to the point that outputs are hard to tell apart.

Readability earns praise as well: the reviewer jokes that Claude writes in a dialect of its own, whereas Muse Spark 1.3 rarely needs to be told to reword a response. Speed is described as solid, with an average time to first token followed by fast generation.

Pricing and usage limits

Price was the main reason the reviewer tried the model in the first place. On the $15 plan, continuous coding ran for roughly 2 to 4 hours before the 5-hour limit was reached, with heavy concurrent sessions draining it faster. The weekly allowance works out to around 4-5 of those 5-hour windows, roughly 25 hours. A $50 plan adds about 33% more usage, which the reviewer calculates is enough for approximately 11.5 hours of coding per day. Compared with rival plans, this is considered one of the best values available.

Why it matters

The review suggests Meta is playing catch-up in the coding-agent market, and that the gap lies more in product plumbing than in the model itself. The underlying intelligence, output quality and price-to-usage ratio are already competitive with Codex, Claude and Cursor; what holds the ecosystem back is unreliable skill execution, an unconfigurable sandbox and inconsistent formatting. The reviewer's bottom line is that there is no strong reason to switch today unless you want something new or a cheaper option, but given Meta's model, data centers and engineering bench, the product could plausibly mature within a few months. For teams watching AI tooling budgets, the pricing structure alone makes this one worth tracking.