Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks in the Meta Model API
Today, Meta Superintelligence Labs was released Muse Spark 1.1. Alongside it, Meta has opened a public preview of the Meta Model API. That second part is a structural change. Meta models have previously reached developers mainly as open weights. Muse Spark 1.1 is closed, managed, and metered per token. So the question is small. Where does it fit into the stack you're already running?
What is Muse Spark 1.1?
Meta describes it as a multimodal thinking model built for agent activities. The benefits reported with the first Muse Spark lie in tool use, computing, coding, and multidisciplinary understanding. The context window is 1,000,000 tokens. List of Meta Model API documentation 1,048,576.
Key Skills and Traits
Because it is a thinking model, so it thinks before responding. In addition, this thinking effort can be adjusted according to each application. Embedded timeline text, images, video, and documents; the output is text. The API also features structured output, parallel tool calling, the Files API, and fast caching. Adding a web_search the Responses API call tool returns quoted responses.
Pricing and Regional Availability
Access splits two ways. Buyers get it for free in 'Thinking' mode on the Meta AI app and on meta.ai. Developers pay $1.25 for a billion input tokens and $4.25 for a million output tokens. New accounts get $20 in free credits. The first launch post describes the public preview as US only, with no EU access yet.
Working
By specifying the location, the numbers define the shape. To demonstrate this, Meta published a launch table, and the table breaks down cleanly.
| Benchmark | Tests | Muse Spark 1.1 | Opus 4.8 (maximum) | GPT-5.5 (xhigh) | Gemini 3.1 Pro (top) |
|---|---|---|---|---|---|
| The MCP Atlas | Measured tool usage | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench | Use of a professional tool | 54.7 | 48.4 | 38.3 | 15.9 |
| Humanity's Final Test | Consulting tools | 62.1 | 57.9 | 52.2 | 51.4 |
| OSWorld-Verified | Using a computer | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro | Writing the actual repo code | 61.5 | 69.2 | 58.6 | 54.2 |
| DeepSWE 1.1 | Long horizon coding | 53.3 | 59.0 | 67.0 | 12.0 |
| BabyVision | Visual thinking | 76.3 | 81.2 | 83.6 | 51.5 |
Benchmark analysis
The meta is reported, with competitors shown in their strongest forms. Muse Spark 1.1 leads the use of tools and extended lines of thinking with tools. Places third in coding and multimodal. So, this is an orchestration model, not a code correctness leader. Meta also chose a set of poses and used a harness.
Factor to Take Seriously: Collaboration and Transfer
Beyond the points, orchestration behavior describes the results of the use of tools. The model actively manages a context window of one million tokens. It remembers actions, retrieves information from previous work, and consolidates what it stores.
The sending of delegates is the second half. As the main agent, it collects context, programs, and execution delegates from all the same subcomponents. As a subagent, it sticks to its task, understands the tools available, and falls back when needed. The research team also reports zero-shot implementation of new native tools, MCP servers, and custom capabilities.
Computer usage follows the same logic. The model was trained to write scripts where automation is fast. It clicks when direct communication is easy. It generates sets of actions for each step.
Wiring it to an Existing Stack
Because the Model API is OpenAI compatible, the migration is a base-URL change instead of a rewrite. The snippet below is the recipe for the first Meta phone.
# pip install openai
import os
from openai import OpenAI
# The OpenAI SDK does not auto-read MODEL_API_KEY, so pass it explicitly.
client = OpenAI(
base_url="
api_key=os.environ["MODEL_API_KEY"],
)
response = client.chat.completions.create(
model="muse-spark-1.1",
messages=[{"role": "user", "content": "Hello, world!"}],
)
print(response.choices[0].message.content)Anthropic format harnesses, such as Claude Code, point to the Message API instead. Agent CLIs like OpenCode register a provider using three values: base URL, key, model ID.
Use Cases
Basically, the demo map sent to engineering teams has a workload that they are already seeing.
For example:
Automatic Multimodal listing:
- In the Facebook Marketplace demo, the model takes a smartphone video, extracts useful images, reasons about the product, and uses a browser to publish the listing.
Screenshot-driven debugging:
- In the OpenCode demo, it builds an interactive web application, takes automatic screenshots, traces failures back to the code, and verifies fixes.
Flexible planning:
- In the party demo, a new context arrives in the middle of an order, and the model updates the system without prompting.
Coding harnesses get first-class support: programming mode, goal state, sub-delegation, context density. The Meta team also reports significant benefit from the Meta Internal Coding Bench.
'+''+esc(MODELS[i])+''+''+''+x.toFixed(1)+'
'; }).join(''); come back
'+esc(dn)+''+''+esc(dt)+'
'+bars+'
'; }).join(''); requestAnimationFrame(function(){ box.querySelectorAll('.fill').forEach(function(f){ f.style.width = f.getAttribute('data-w')+'%'; }); resize(); }); } document.getElementById('msxChips').addEventListener('click',function(e){ var b = e.target.closest('.chip'); if(!b) return; this.querySelectorAll('.chip').forEach(function(c){ c.setAttribute');'-pressed';-press' b.setAttribute('aria-pressed', 'true'); root.querySelectorAll('.tab').forEach(function
[$in,$out,$runs].forEach(function(el){el.addEventListener('input',calcCost);});
[$cin,$cout].forEach(function(el){el.addEventListener('input',calcCost);}); // —- route advisor —- var W={tool:0,mcp:0,ctx:0,cost:0,acc:0,eu:0}; document.getElementById('msxQs').addEventListener('click',function(e){ var b=e.target.closest('.toggle'); if(!b) return; var on=b.getAttribute('aria-pressed')==='true-pressed(',setAttribute) to 'false':'true');[b.closest(‘.q’).getAttribute(‘data-w’)] = in?0:1; advise (); }); function advise(){ var t=document.getElementById('msxRecT'), d=document.getElementById('msxRecD'); var fit=W.tool+W.mcp+W.ctx+W.cost; if(W.eu){ t.textContent=”Blocked from access, not from power”; d.textContent=”The Meta Model API preview is open to US developers. Plan to test regional releases instead of navigating access controls.”; } else if (W.acc && fit<=1){ t.textContent="Gcina lo msebenzi kumholi wokubhala ngokunemba"; d.textContent="Kuthebula lokwethulwa kwe-Meta, i-Opus 4.8 ihola i-SWE-Bench Pro kanti i-GPT-5.5 ihola i-DeepSWE 1.1 ne-BabyVision. I-Muse Spark 1.1 ibeka isithathu kuleyo migqa."; } okunye uma(fit>=3){ t.textContent=”Strong candidate: port this project to Muse Spark 1.1″; d.textContent=”This is a Meta tuned profile. Test it with a $20 credit, then compare the cost per completed run against your current settings before switching.”; } else if(fit===2){ t.textContent=”Suitable for A/B testing”; d.textContent=”The OpenAI compatible endpoint makes the base-URL change and the main change. Run both models in the same harness for a week and study the retry rate.”; } else if(fit===1){ t.textContent=”Weakest fit so far”; d.textContent=”One matching condition is not enough. Add features that define your function, or keep your defaults.”; } else { t.textContent=”Answer the above questions”; d.textContent=”Change the conditions that define your job. The advisor updates as you go.”; } resize(); } // —- copy —- document.getElementById('msxCopy').addEventListener('click',function(){ var txt = document.getElementById('msxCode').innerText; var btn = this; function done(){ btn.textContent;Copied”Copied setTimeout(function(){ btn.textContent=”Copy snippet”; },1600 } if(navigator.clipboard && navigator.clipboard.writeText){ navigator.clipboard.writeText(txt).then(done,fallback.(}) ta=document.createElement('textarea'); document.body.appendChild(ta) // —–automatically resize WordPress iframe embeds —- function resize(){ var h = root.offsetHeight + 40; try{ window.parent.postMessage('muse-resize', '*'); window.addEventListener('resize', 1200); calcCost();
