At the beginning of September 2026, Meta released Muse Spark 1.3—a reasoning model for agentic programming. Vendor tables rank it above Claude Opus 5 and GPT-5.6 Sol in coding and long-context tasks. At the same time, a way to run the model for free through a third-party service appeared. We examine which claims are confirmed by independent measurements and which are worth checking yourself.
What is known about the model
Muse Spark 1.3 continues the line that replaced Llama. It is a proprietary model: the weights are not published, and access is provided through Meta’s API and partner platforms. Its stated purpose is long autonomous coding sessions: analyzing large repositories, fixing bugs, and working in the terminal.
Two reasoning modes are available: xhigh and max. The more advanced mode is noticeably stronger on some tests, but at launch it was available only in a limited preview. API pricing for xhigh is 1,25 dollars per million input tokens and 4,25 per million output tokens.
Test results
All the figures below come from Meta’s own materials and benchmark aggregators. There had been no independent full audit of version 1.3 at the time of writing, so the table should be read as a claim, not a verdict.
| Benchmark | Muse Spark 1.3 | Competitors |
|---|---|---|
| DeepSWE 1.1 (agentic coding) | 75,4 | Opus 5 — 74,0; GPT-5.6 Sol — 73,0 |
| Terminal-Bench 2.1 (terminal) | 88,8 | Sol — 88,8; Opus 5 — about 86 |
| MRCR v2, 256–512K context | 98,5 | top positions in its group |
| MRCR v2, 512K–1M context | 98,1 | top positions in its group |
| SWE-Atlas (codebase questions) | 59,4 | comfortably ahead of Sol and Opus 5 |
Independent analyses add a note of caution. The Terminal-Bench comparison uses model-plus-agent combinations rather than standalone models, so the figure answers the question “how did the product perform?” more than “how does the model reason?” In Meta’s table, the competitor listed is Opus 5, even though the more powerful Fable 5.1 had already been released by then. Confidence in the figures is moderate: they are a reason to test, not a reason to start buying.
Where the model is strong
The signals converge in three areas. The first is long context: document-retrieval accuracy barely drops even across a range of half a million to one million tokens. For analyzing someone else’s repository, this is the key argument—the model keeps the entire project in mind rather than just a fragment.
The second is autonomous error correction over long runs. The improvement over version 1.2 was more than 20 points, the largest jump in the table. The third is the terminal: running commands, configuring environments, and writing scripts. This is where the model outperforms its competitors most convincingly.
Where it is weaker
The other side is also clear. Graphical-interface control, browser-based agentic workflows, and web searches are not its territory; it trails its competitors there. Following strict instructions and rule-based tasks comes less naturally to it than open-ended reasoning.
The practical conclusion is simple: this is a model for codebases and the terminal, not for carefully following procedures. If the task is “analyze the repository and fix it,” it is a candidate. If it is “follow the instructions literally,” it is worth comparing against competitors on your own examples.
How much it costs
There are two official options. The standard API tier costs 1,25 dollars per million input tokens and 4,25 per million output tokens; the data is not used for training. The Contributor tier is several times cheaper—0,10 and 0,20 respectively—but prompts and responses are sent for training Meta’s models.
Meta also sells subscriptions to its Muse Code terminal agent, starting at 5 dollars per month. The subscription limits the number of requests per unit of time, and the limits may not be sufficient for active development.
Free access through OpenCode Zen
At the beginning of September 2026, the OpenCode Zen model catalog added Muse Spark 1.3 Contributor Free: input, output, and cache reads are available at no cost. This is not free Muse Code, but a free quota for a specific model through an intermediary.
The process is as follows: install OpenCode using the script from the project website or a package manager, open it in the project folder, connect a provider with the /connect command, and select OpenCode Zen. Find Muse Spark 1.3 Free in the model list and make sure the Free label appears on the right. Switching to another model without that label activates its standard pricing.
Three caveats. First, the free version is Contributor—everything entered and received may be used for training. Do not send client code, secrets, or personal data there. Second, this is a temporary offer; the terms have already changed for previous versions. Third, the intermediary may ask for payment details to connect the service, depending on the account.
What to do in practice
If you need the model for evaluation, the free quota is a convenient entry point: you can run your own tasks before buying a subscription. Focus on what the table says the model does well: analyzing an unfamiliar repository, finding where a fix belongs, and performing terminal operations.
Measure it with your own runs, not the vendor table. Run the same bug fix through Spark 1.3 and your current production model, recording the time and number of iterations. For confidential code, use only paid tiers that do not train on your data.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.