Meta logo via Wikimedia Commons
What Muse Spark 1.3 is and when it arrived
Muse Spark 1.3 is Meta's latest update to its Muse Spark model family, and Meta announced it on September 2, 2026. It went out to developers through Meta's API immediately, and Meta also rolled it out to Meta AI and to users inside Instagram and Facebook, plus into Muse Code, the company's dedicated coding product. That spread across four surfaces at once, developer API, a standalone assistant, two social apps, and a coding tool, is itself worth noting. It means the same underlying model is now doing everything from answering a question in a Facebook chat thread to running a multi-step coding task for a professional developer.
The release itself is framed as an iteration rather than a full generational jump. It follows Muse Spark 1.2, and the version number reflects that: Meta is calling out specific, targeted improvements rather than claiming a wholesale rebuild. The improvements Meta lists center on two related areas, coding work and what the industry calls agentic behavior, plus a substantially larger context window. Pricing on the API side did not move, which suggests Meta is treating this as a capability update within the same cost tier rather than a new pricing generation.
What "agentic" and "long-horizon" actually mean
An agentic model is one that does more than answer a single question, it takes actions, checks the results, and decides what to do next, often across many steps without a person restating the goal each time. A long-horizon task is one that takes a while and involves many of those steps in sequence, like building a feature across several files, running tests, reading the errors, and fixing them, rather than answering a single prompt and stopping.
Most chat-style AI interactions are short: one question, one answer. Agentic and long-horizon work is different in kind, not just in length. The model has to hold onto what it already tried, what worked, what didn't, and what the original goal was, all while new information keeps arriving. That's a harder problem than answering a well-formed question, because small mistakes early in a long task can compound if the model doesn't manage its own state carefully. Meta's stated improvements for Muse Spark 1.3, handling long-running agent tasks and managing multiple workflows inside a single long thread, are both aimed squarely at that problem.
The coding and workflow improvements in this release
Meta lists several specific areas of improvement for Muse Spark 1.3, and they cluster tightly around coding and agent work rather than general knowledge or creative writing. The model is meant to be better at following complex instructions, the kind with several conditions or steps bundled into one request, and at asking the user for confirmation at appropriate points instead of either barreling ahead on an ambiguous instruction or stopping to ask about everything.
Coding efficiency is called out as its own improvement area, separate from the agentic gains, meaning the model is meant to write and edit code with less wasted effort even in single-step tasks. Managing multiple workflows within one long thread is arguably the most practical of these changes for a working developer: it means a single ongoing conversation with the model can track more than one task at a time, such as fixing a bug in one part of a project while also updating documentation in another, without the threads bleeding into each other or losing track of which change belongs where.
Why fewer tool calls and fewer tokens matters for cost
Meta says that, in its own internal comparisons, Muse Spark 1.3 uses about 20% fewer tool calls and about 25% fewer tokens than Muse Spark 1.2 to complete comparable tasks. These are Meta's own figures from its own testing, not numbers verified by an outside benchmark, so it's worth reading them as a company's account of its own product rather than an independent measurement. Still, the direction of the claim is a useful one to understand even without an independent number to compare it against.
A tool call is any time the model reaches out to use something, running a command, searching for information, reading a file, and each one adds latency and, on most billing setups, cost. Fewer tool calls to reach the same result means a task finishes faster and cheaper. Fewer tokens matters for a related but separate reason: API pricing is charged per token, so a model that reaches a correct answer using less text, both in what it reads and what it generates, costs less to run at the same volume, independent of any change in the price per token. For a developer running the same kind of agentic workflow repeatedly, over thousands of requests a day, this kind of efficiency gain compounds directly into a lower bill, even with the per-token price held flat.
The 1 million token context window, explained
Muse Spark 1.3 is a multimodal model, meaning it can work with more than just text, and it has a context window of 1 million tokens. A token is roughly a piece of a word, and a context window is the total amount of text, code, images, or other input the model can hold in view at once for a single request, including both what you feed it and what it generates back.
A 1 million token window is large enough to hold an entire sizable codebase, a long stack of documents, or hours of conversation history in one request, without the model losing track of earlier parts of that input. In practical terms, that means someone can hand the model a whole project's worth of files and ask it to find where a bug originates, or feed it a long research document and a full email thread at once and ask it to reconcile the two, without manually chopping the material into smaller pieces first. Combined with the coding and agentic improvements described above, that larger window is what lets Muse Spark 1.3 keep track of a long, multi-step task without repeatedly losing the thread of what it was originally asked to do.
Where it shows up: API, Muse Code, Meta AI, Instagram and Facebook
Muse Spark 1.3 reaches three distinct kinds of users. Developers building their own applications can call it directly through Meta's API, which is the version most likely to be used for custom agentic workflows, internal tools, or products a company builds on top of it. People who write code for a living, or as a hobby, can reach the same model through Muse Code, Meta's dedicated coding product, where the coding-specific improvements are the most directly relevant.
Then there's the much larger group: ordinary users who never touch an API or a coding tool at all. Meta rolled Muse Spark 1.3 out to Meta AI, its general-purpose assistant, and into the AI features already built into Instagram and Facebook. For that group, the update is mostly invisible as a version number, it shows up as the assistant behind a chat in Messenger, a suggestion inside Instagram, or a response from Meta AI simply getting somewhat better at following what was asked. Three very different audiences are drawing on the same underlying model, just through different doors.
Pricing and what stays the same
API pricing for Muse Spark 1.3 is unchanged from 1.2: $1.25 per million input tokens and $4.25 per million output tokens. Meta did not use this release to raise or lower the sticker price on the API, which is a meaningful choice in itself. Combined with Meta's claim of fewer tool calls and fewer tokens per task, holding pricing flat means the effective cost of running a given agentic workflow could come down even though the per-token rate did not change, since the same task is reported to need less of both.
That's a different kind of improvement than a price cut, and it's worth keeping the distinction clear. A price cut lowers the cost of every token regardless of how the model is used. An efficiency gain like the one Meta describes here only lowers cost if it actually reduces the number of tokens and tool calls a given task consumes, which will vary by use case and hasn't been independently verified outside Meta's own numbers.
Max mode and why it's still limited
Muse Spark 1.3 includes what Meta calls a max mode, a more compute-intensive setting than the standard version of the model. Meta is keeping that mode limited in availability while it finishes more safety testing on it. That's a fairly common pattern across the industry this year: a more capable or more resource-hungry mode gets held back from general release until a company is satisfied with how it performs under closer scrutiny, even as the standard version of the model ships broadly.
There's no public detail from Meta yet on exactly what max mode changes about the model's behavior or when the additional testing will wrap up, so it's best treated as a feature to watch for rather than one to plan around today. For most developers and everyday users, the standard release of Muse Spark 1.3, the one already available through the API, Muse Code, Meta AI, Instagram, and Facebook, is the version that matters right now.
How this fits the rest of September 2026's model releases
Muse Spark 1.3 landed in the middle of a busy month for AI model releases. September 2026 also brought Claude Opus 5.5, Gemini 3.8 Flash, DeepSeek V4.1 Flash, and OpenAI Astra, each from a different lab and each with its own emphasis. Without inventing benchmark comparisons between these models, since none of the companies involved have published numbers that would let a reader compare them directly, it's fair to describe the general shape of what each release focused on based on its own vendor's framing: efficiency, cost, and specific capability gains rather than a single sweeping leap for any one of them.
Coverage of Muse Spark 1.3, including reporting from Bloomberg, has framed the release as part of Meta narrowing the gap with OpenAI and Anthropic rather than pulling ahead of them outright. That framing lines up with what Meta itself is emphasizing: targeted gains in coding and agentic reliability, a larger context window, and unchanged pricing, not a claim of leading the field on general capability. Positioned alongside the other releases from the same month, Muse Spark 1.3 reads as Meta's entry in a broader pattern where every major lab is shipping incremental, cost-conscious updates rather than dramatically repositioning its lineup.
Privacy considerations of an assistant built into social apps
Because Muse Spark 1.3 now sits inside Meta AI and inside the AI features of Instagram and Facebook, it's worth thinking about it not just as a coding or API tool but as an assistant woven into apps that already hold a lot of personal information, photos, messages, contacts, and browsing behavior on those platforms. That's a different privacy situation than using a standalone chatbot on its own website, because the assistant can potentially draw on context from an app someone already uses for personal communication.
This isn't a claim about any specific data practice Meta has stated for Muse Spark 1.3, and nothing here should be read as describing a documented policy that Meta hasn't actually published. The reasonable general advice is the same advice that applies to any AI feature built into a social platform: check what permissions and settings are attached to Meta AI in your own account, review what Meta's own privacy settings say about how your activity feeds an AI assistant, and decide deliberately what you're comfortable sharing with it, the same way you would with any other feature on Instagram or Facebook that touches your personal data.
Frequently asked questions
When did Meta release Muse Spark 1.3?
Meta released Muse Spark 1.3 on September 2, 2026. It became available to developers through Meta's API right away, alongside a rollout to Meta AI and to users on Instagram and Facebook, and it also rolled out in Muse Code, Meta's coding product.
What is new in Muse Spark 1.3 compared to 1.2?
The focus is coding and agentic tasks: handling long-running agent work, managing multiple workflows inside a single long thread, following complex instructions, asking the user for confirmation at the right moments, and general coding efficiency. Meta also says it uses about 20% fewer tool calls and about 25% fewer tokens than 1.2 on comparable tasks, based on its own internal testing.
How big is the context window?
Muse Spark 1.3 is a multimodal model with a 1 million token context window, which is large enough to hold long documents, large codebases, or extended conversation history in a single request.
Did the API pricing change?
No. Pricing is unchanged from Muse Spark 1.2, at $1.25 per million input tokens and $4.25 per million output tokens.
What is max mode and why is it limited?
Max mode is a more compute-intensive setting for Muse Spark 1.3. Meta is keeping it limited in availability until the company finishes more safety testing on it.
Who can actually use Muse Spark 1.3?
Developers can use it through Meta's API, coders can use it inside Muse Code, and everyday users can reach it through Meta AI and through the AI features built into Instagram and Facebook.