Muse Spark is Meta’s model for code tasks and agents, and its version 1.3 does the same job with 25% fewer tokens than 1.2. It was released on September 2, 2026 and reached Muse Code, Meta’s terminal agent, that same day, which had exited beta two days earlier with plans starting at $5 per month.
It’s worth separating two statements that tend to get mixed up. The 25% figure—and the 20% fewer tool calls—comes from internal comparisons by Meta’s own engineers against Muse Spark 1.2. It’s not an independent benchmark or part of the published results table. Meta did publish a separate table with eleven benchmarks, and that table has an important detail we review below.
What is Muse Spark?
Muse Spark is Meta Superintelligence Labs’ family of models oriented toward programming, reasoning, and agentic work. It’s the engine, not the tool: what you use day to day is Muse Code, and Muse Spark is the model running underneath. It’s also available directly through Meta Model API.
And what is Muse Code?
Muse Code is Meta’s code agent that runs in the terminal. You install it on your machine, point it to a repository, and work by conversing with it, just like you would with any other command-line agent.
The distinction matters this week because both pieces moved at the same time: Muse Code exited beta on August 31 and the model was bumped on September 2. Muse Code has been public since early August, so the tool isn’t new; what’s new is that Meta now treats it as a finished product rather than a preview.
Exiting beta added four things that change the workflow quite a bit:
- Cross-session messaging. Two Muse Code sessions can send each other messages. If one changes something another depends on, it sends the notification; if one solves a problem that had another blocked, it sends the answer. Messages travel over a Unix socket between your processes, so nothing goes out over the network. The agent discovers its peers and writes the message on its own.
- Workflows. You type “use a workflow” and Muse Code orchestrates a team of sub-agents in parallel on the same task. In
/workflowsyou see live progress and can stop, restart, or cancel. Workflows that work for you are saved and reused. - Rewind. Double Esc and you go back to an earlier point in the conversation. Muse Code only offers return points it considers safe according to its event log, and asks you for confirmation before undoing work.
- SDK in preview. The CLI engine—sessions, tools, permission control—as a TypeScript library. Your program spins up a local
musehost and talks to it over the Muse Session Protocol using standard input and output: no server and no network.
What changes in Muse Spark 1.3?
Three things, in order of how much you’ll notice them.
It costs less to run the same task. Fewer tool calls, fewer tokens, and fewer turns where they weren’t needed. Meta describes the code style as cleaner and less verbose, and that’s the mechanism behind the number, not a separate claim.
It asks instead of guessing. Meta trained 1.3 to ask questions when the prompt is ambiguous, ask for help when it gets stuck, and confirm before consequential actions. On long tasks it adapts to whether you prefer frequent updates or for it to work in silence. It’s a usability change more than a capability change, and it’s the kind you notice on day one of use.
It knows what it doesn’t know. According to Meta, the model was trained to recognize when it hits a wall instead of hallucinating a result, and to maintain detailed requirements throughout multi-step tasks without gradually dropping constraints along the way.
On security, Meta claims better resistance to prompt injection and better judgment about what counts as an irreversible action. Both are characterizations from the manufacturer itself.
How does it compare in price to Claude Code?
Muse Code has three monthly plans, published on August 31, that Meta positions as cheaper than pay-as-you-go on Meta Model API:
| Plan | Price | For what |
|---|---|---|
| Everyday Usage | $5.00/month | 10–50 requests every 5 hours, image and video upload, voice mode, web search |
| High Usage | $15.00/month | 3x Everyday quota, plus multimodal inputs, access to the latest models |
| Power Usage | $50.00/month | 10x Everyday quota, larger file uploads, early access to new features |
On the other side, Claude’s plans are Free at $0, Pro at $17 per month with annual billing ($20 if you pay monthly), and Max from $100 per month. The detail that matters for this comparison: Claude Code is not included in the Free plan. It appears starting with Pro, and Max adds 5x or 20x more usage than Pro.
So the honest reading of pricing is this. The cheapest entry into a code agent is Muse Code at $5, below Claude Pro’s $17. But neither one has a real free tier for agentic work, and the quotas aren’t comparable to each other: Meta counts requests per five-hour window and Anthropic doesn’t publish its quota in the same units. Comparing $5 against $17 on its own is misleading.
And there’s a detail worth checking before you subscribe. On Meta’s plans page from August 31, 2026—published two days before Muse Spark 1.3 came out—all three tiers describe their model access in terms of Muse Spark 1.2: “Access to Muse Spark 1.2” on Everyday, “More Muse Spark 1.2 usage” on High, and “Expanded Muse Spark 1.2 usage” on Power. Only High and Power also promise “access to the latest models.” As of the publication date of this article, September 3, 2026, Meta had not updated that page to mention 1.3 or clarify whether the $5 tier includes it. If you’re subscribing specifically for the efficiency improvements we describe here, the $15 plan is the only one whose wording covers them without ambiguity.
How do you install Muse Code?
One command, on macOS or Linux, just as Meta publishes it:
curl -fsSL https://dev.meta.ai/install.sh | bash
Then you log in with your Meta Model API account. There’s no installation path for Windows in Meta’s instructions: the documented command covers macOS and Linux only. If you work on Windows, WSL is the obvious solution, but Meta doesn’t document it, so treat it as your own configuration and not as a supported path.
What do Meta’s benchmarks say?
Meta published its own comparison table, so it’s worth reading with caveats up front. These are benchmarks chosen by the manufacturer and run by the manufacturer, and they compare Muse Spark 1.3 at maximum reasoning against Opus 5 and GPT 5.6 Sol also at maximum. Muse Spark 1.2 appears at xhigh, one level lower, which favors the generational jump.
| Benchmark | Muse Spark 1.3 (max) | 1.2 (xhigh) | GPT 5.6 Sol (max) | Opus 5 (max) |
|---|---|---|---|---|
| GDPVal-AA v2 | 1754 | 1615 | 1710 | 1824 |
| JobBench | 64.9 | 61.6 | 45.4 | 65.7 |
| OSWorld 2.0 (partial/binary) | 66.9 / 32.0 | 47.6 / 17.9 | 62.7 / 27.3 | 68.3 / 31.4 |
| DeepSearchQA | 89.4 | 85.9 | 93.0 | 90.4 |
| Agentic IF Index (internal) | 57.8 | 46.2 | 60.5 | 59.1 |
| AutomationBench | 49.4 | 38.2 | 46.7 | 50.3 |
| MRCR 256K–512K | 98.5 | 66.3 | 91.5 | – |
| MRCR 512K–1M | 98.1 | 55.5 | 73.8 | – |
| DeepSWE v1.1 | 75.4 | 55.0 | 73.0 | 74.0 |
| SWEAtlas CodeBase QnA | 59.4 | 46.2 | 53.5 | 52.7 |
| Terminal-Bench 2.1 | 88.8 | 82.9 | 88.8 | 86.7 |
In agentic work, Meta loses its own table. Opus 5 leads in knowledge work, professional tool use, computer use, and end-to-end business flows. GPT 5.6 Sol leads in agentic navigation and instruction following. Muse Spark 1.3 takes the OSWorld 2.0 binary score by six tenths and little more.In code and long context the photo flips, and that’s where this article’s argument lives. The 1.3 leads the three rows of code and the two of long context, with 98.5 and 98.1 in MRCR versus 91.5 and 73.8 of GPT 5.6 Sol. Anthropic did not report MRCR, so those two rows have no comparison with Opus 5.
The honest summary is narrower than “Meta wins” or “Meta loses”: with their own numbers, Meta built a better model at reading a large codebase and working in the terminal than operating a computer or browsing the web. If your work is the former, that’s the relevant finding. If it’s the latter, the table says buy something else, and Meta is the one saying it.
What’s still not available?
The maximum reasoning mode. Meta’s announcement says that previous reasoning modes work today and that the maximum arrives “shortly after completing additional safety testing”. At the time of publishing this note, September 3, 2026, Meta had not given a date.
And here’s the promised detail about the table: the Muse Spark 1.3 column says (max). All the numbers Meta published for the 1.3 come precisely from the mode you still can’t use. They’re not the results of the model that installs today.
The open weights release of Muse Spark is also announced without a timeline. If you’re reading this weeks later, both deserve a revisit.
Is it worth switching?
It depends on where your account is, not on which model scores better. The shapes of the two tools converged: both are terminal agents, both orchestrate subagents in parallel, both bring SDKs and both sell subscription in addition to API usage. The only real differentiator of Muse Code today is messaging between sessions over local socket; multiple sessions coordinating without you playing messenger between terminals is not something alternatives document.
If you already work over Meta Model API, the 1.3 is a free upgrade that cheapens each task. If not, installation is a command and evaluating it costs you an afternoon. What you shouldn’t do is decide with the benchmark table in hand, because that table describes a mode Meta still hasn’t delivered.
If you’re interested in the rest of the family, we covered Meta’s open model in Meta says 3.1x but doesn’t say how much: running Muse Glimmer 30B local with llama.cpp. And on how the price per task changes your model decision, check out Almost Fable 5 at half the price: why Opus 5 changes your model decision.