Meta Muse Spark 1.1: The Everything-Input AI Model

Most "multimodal" models mean text plus images, and maybe a PDF if you ask nicely. Meta Muse Spark 1.1 takes text, images, video, audio, and PDF documents — all in one model, all in one conversation, with a 1,048,576-token context window to hold it. Meta introduced it on July 9, 2026 as the second release from Meta Superintelligence Labs, and it is now live on CoreAI as a standard-tier reasoning model. That last part matters: the widest input support of any current frontier model, priced like a mid-tier workhorse.

What Is Meta Muse Spark 1.1?

Muse Spark 1.1 is a multimodal reasoning model built for agentic work — the kind where a model does not just answer a question but operates tools, navigates software, and carries a task across many steps. Meta shipped it through a public preview of its new Meta Model API, and per Meta's launch post, the 1.1 release brings major gains in tool use, computer use, coding, and multimodal understanding over the original April model.

The pitch is persistence. Plenty of models can describe a video frame or transcribe audio. Muse Spark 1.1 is designed to inspect visual and audio input, retain the details that matter, and keep using them many steps later while it is simultaneously running tools and writing code. That is a different skill from captioning, and it is the one agent builders have been asking for.

Why Does Everything-Input Actually Matter?

Because real work is not text-shaped. A support ticket arrives with a screen recording. A contract is a scanned PDF. A bug report is a screenshot plus a voice memo from a stressed colleague. With most models, each of those needs a separate preprocessing step — transcribe the audio, OCR the PDF, extract frames from the video — before the model ever sees it. Muse Spark 1.1 ingests the mess directly.

The 1M-token context is what makes that practical rather than theoretical. Video and audio are token-hungry; a window that fits an hour of mixed media plus your whole conversation is the difference between a demo and a tool. If you want the background on why million-token windows changed workflows, we broke it down in our guide to 1M-token context models.

Muse Spark 1.1 Specs and Pricing vs the Field

Here is how Meta's model stacks up against the other big multimodal releases of mid-2026, using live per-token pricing:

Read that table twice. The model with the widest input support is also the cheapest per token of the four. Meta is clearly buying its way into a crowded agent market — TechCrunch called it Meta's entry into the AI coding battle against OpenAI and Anthropic — and buyers get to enjoy the subsidy while it lasts.

One more wrinkle worth knowing: availability is uneven. Meta's Model API preview launched US-only with no EU access yet, and consumers get the model free only inside the Meta AI app's Thinking mode. So while the specs are global-frontier, the official access story is narrower than the launch headlines suggest — which is part of why third-party access matters more than usual for this release. If your team is outside the US, the API waitlist is not your only door in.

When Should You Pick Muse Spark 1.1?

Our take: make it your default for anything involving mixed media. Summarizing a recorded meeting, auditing a video ad against a brief in a PDF, turning a screen recording of a bug into a reproduction script — this is the model for those jobs, because it is the only one that sees all the evidence natively. It is also a serious option for agentic coding, where its computer-use gains and long-context persistence show up.

Where we would not pick it: pure text work with no media in sight. For straight writing, chat, and analysis, Claude Sonnet 5's output quality still justifies its higher price, and premium reasoning models keep the edge on the hardest problems — the same tier logic we mapped in our GPT-5.6 tier guide. Muse Spark 1.1 wins when the input is messy, multimodal, and long.

How to Try Muse Spark 1.1 Without the API

Meta's own preview is developer-focused, initially US-only, and lives behind an API key. The simpler route: Muse Spark 1.1 is already live on CoreAI across iOS, Android, and the web — same subscription that covers 300+ other models, with plans from $57/month. Attach a PDF or an image, toggle thinking mode, and see how it handles your real files. Better yet, open Compare and run the same prompt through Muse Spark 1.1 and Claude Sonnet 5 side by side before you commit a workflow to either.

Frequently Asked Questions

What inputs does Muse Spark 1.1 accept?

Text, images, video, audio, and PDF documents, with text output. That is the broadest input set of any major frontier model as of July 2026, and it removes the preprocessing pipeline — transcription, OCR, frame extraction — that other models require for mixed-media work.

How big is the Muse Spark 1.1 context window?

1,048,576 tokens — a full million-token window. That is enough for hours of mixed media, entire codebases, or long agentic sessions where the model needs to remember details from early inputs hundreds of steps later.

How much does Meta Muse Spark 1.1 cost?

API pricing is $1.25 per million input tokens and $4.25 per million output tokens, which places it in the standard tier — notably cheaper per token than Claude Sonnet 5 or Kimi K3. On CoreAI it is included under one flat subscription with no per-token billing.

Is Muse Spark 1.1 good for coding?

Yes. Meta positioned the 1.1 release around agentic coding, with major gains in tool use and computer use over April's original Muse Spark. It is built to carry long engineering tasks, though for the hardest pure-code problems it competes with dedicated coding flagships rather than clearly beating them.

Where can I use Muse Spark 1.1 today?

Meta's Model API preview is US-only and developer-oriented. For everyone else, it is live in CoreAI's model library on iOS, Android, and the web — no API key, no code, and you can compare it against 300+ other models.