Meta Muse Spark 1.1: A Low-Cost Computer-Use Agent Model

On July 9, 2026, Meta shipped Muse Spark 1.1 and, for the first time, put a frontier-class model behind a paid, self-serve developer API. If you build agents or automations, the model itself is only half the story. The other half is that Meta finally opened the door for you to call it directly.

What is Meta Muse Spark 1.1?#

Muse Spark 1.1 is Meta's multimodal reasoning model built for agentic work, released on July 9, 2026 by Meta Superintelligence Labs. It handles tool use, computer use, and parallel subagents across a 1M-token context, and it is the first Meta model you can call from a paid, self-serve developer API.

Meta describes it as a strong agentic and coding model at a low price, and it also runs in Thinking mode inside the Meta AI app and on meta.ai. The headline capabilities are the ones agent builders care about: it can plan, use tools, and drive a computer. What Meta has not published is an independent benchmark, so the capability claims are the vendor's for now.

Meta's launch spec and early reported pricing, dated July 9, 2026. Pricing is from reporting, not a Meta price page; verify before relying on it.

Why is the Meta Model API the real story?#

Because access, not raw capability, is what changed. The Meta Model API is a self-serve public preview that speaks the OpenAI format, so you can point an existing client at a new base URL and key. Meta has historically kept its best models inside its own apps, and this reverses that posture.

That matters more than any single spec. A model you cannot call is a press release; a model behind an OpenAI-compatible endpoint is something you can pilot this afternoon. Reporting also says the API accepts the Anthropic Messages format, which widens the set of existing clients that can reach it. The price is the other signal, and it reads like a deliberate move in the ongoing July 2026 model wave.

Input price per 1M tokens: Muse Spark 1.1 as reported (AI Weekly); GPT-5.6 Sol and Fable 5 per vendor pages. Dated July 2026, so verify each.

What can a computer-use, subagent model actually do?#

Meta positions Muse Spark 1.1 as an agent foundation, not a chat box. As the main agent it can gather context, make a plan, and delegate work across parallel subagents. Its computer-use mode writes scripts when automation is faster and clicks through interfaces when direct interaction is simpler, across multiple apps.

The interesting claim is orchestration: Meta says the model is trained to run multi-agent systems and optimize end-to-end latency, acting as the planner that farms work out to subagents. If that holds up on real tasks, it fits the plan-and-execute agent pattern directly. The computer-use side is the higher-risk, higher-reward part, because an agent that can click and type in live software needs a sandbox and hard guardrails before it touches anything that matters.

Reported Muse Spark pricing vs Anthropic's published Fable 5 pricing, July 2026. Both are vendor or reported figures; verify before you commit.

How much does Muse Spark 1.1 cost, and what is the catch?#

Early reporting puts the API at about $1.25 per 1M input tokens and $4.25 per 1M output, with $20 in free credits for a new account. That undercuts the premium tier heavily. The catches are real: it is a preview, benchmarks are not independent yet, and Meta has not posted a public price page.

To see why the number is loud, put it next to the models it is chasing. The spread below uses the reported Muse Spark figure against published prices for two frontier tiers; output prices, which dominate most bills, differ even more.

The access catch is the one to plan around. A public preview can change prices, rate limits, or terms with little notice, and early reports say the preview opened to US developers first. If you want the full three-tier context on the model it is undercutting, our breakdown of GPT-5.6's Sol, Terra, and Luna tiers covers it.

Should you move an agent onto it?#

Treat it as a low-cost pilot, not a migration. Because it speaks the OpenAI format, testing it against your real inputs costs little more than a config change. Keep a generally available model wired as a fallback while the preview stabilizes, and route only workloads where its price and context clearly help.

The workloads that fit first are the ones where price scales with volume or where a long context earns its keep: bulk classification, extraction over large documents, and agent loops that would otherwise burn premium tokens on routine steps. Wire it in the same way you would any new option, as a swappable model behind a fallback, using the routing shape in our multi-model fallback for n8n.

A short pre-flight for trying the preview without betting a production workflow on it.

Route by whether the step needs cheap scale or agentic computer use; every path ends at testing on your own data.

What is the practical move for builders?#

Add Muse Spark 1.1 to your shortlist, then let your own tests decide. The price and the OpenAI-compatible API make it cheap to try, and the agentic features fit real automation work. But it is a preview built on vendor claims, so verify the numbers and keep your model choice easy to swap.

None of this is a verdict on the model. It is a read on the moment: Meta priced aggressively, opened a familiar API, and shipped features aimed squarely at agent builders. The disciplined response is to pilot it cheaply, measure it on your data, and treat every figure here as volatile until you confirm it. For the cost habits that make that easy, our seven levers to cut AI API costs still apply, whatever model wins your next test.