Meta Muse Spark 1.1 is a multimodal AI model built for agentic work, and the launch price is the point: $4.25 per million output tokens.
What Meta launched
On 9 July 2026, Meta introduced Muse Spark 1.1, described as the latest model from Meta Superintelligence Labs, with gains in tool use, computer use, coding, and multimodal understanding. Alongside the model came a public preview of the Meta Model API — the first paid developer access Meta has ever offered to its own models.
The pricing sits at $1.25 per million input tokens, $0.15 per million cached tokens, and $4.25 per million output tokens, with web search charged at $2.50 per thousand queries. The public preview runs in the United States only, behind a waitlist, with twenty dollars of initial credits. Muse Spark 1.1 is also live in Thinking mode inside the Meta AI app and on meta.ai.
Set the output price beside the field. Claude Opus 4.8 costs $25 per million output tokens. GPT-5.6 Sol costs $30. Claude Fable 5 costs $50. Meta undercuts the nearest of those by a factor of roughly six, and the cheapest read of the strategy is the correct one — Mark Zuckerberg has described rival pricing as very extreme, carrying very high margins.
The discount is real, not a headline trick
Cheap tokens mean nothing if a model burns more of them. Per-task measurement settles the question. On Artificial Analysis' Intelligence Index, a composite of nine evaluations of economically useful work, Muse Spark 1.1 set to xhigh reasoning scores 51 points — tying Z.ai's GLM-5.2 and GPT-5.6 Luna — at $0.26 per task. Among models scoring at or above 51, only GPT-5.6 Luna runs cheaper, at $0.21.
Elsewhere the picture is mixed, and worth stating plainly rather than selling. On Arena.ai's Text Arena, where humans compare outputs blind, Muse Spark 1.1 ranks sixth at 1,490 Elo, behind Claude Fable 5 and four Claude Opus versions. Tool use is where the model genuinely leads: 88.1 per cent pass rate on Scale AI's MCP Atlas, and second place on JobBench at 54.7 per cent. Strong agentic execution, second-tier raw intelligence, and a price that makes the trade attractive at volume.
Token prices decide which applications are economical at scale. An agent running a thousand steps is a rounding error at $4.25 per million output tokens and a budget line at $50. Meta moved the gate, and what opens up behind the gate — long-running agents that were previously uneconomic — will matter more than the launch itself.
Capability moving from scaffolding into weights
The engineering detail deserves attention beyond the price tag. Meta trained Muse Spark 1.1 to orchestrate multi-agent systems in both directions. Leading a task, the model gathers context, plans, and delegates execution across parallel subagents to cut end-to-end latency. Working as a subagent, the model keeps to the assigned role, understands the available tools, and escalates back to the orchestrating model when a decision would exceed its permissions.
Context management moved inside the model as well. Meta says Muse Spark 1.1 actively manages a one-million-token context window over long jobs — remembering actions, retrieving details from much earlier in a task, and compacting exchanges while keeping the steps later work will need. For computer use, Meta trained the model to choose between writing a script when automation is faster and clicking directly when interaction is simpler, and to issue batches of actions per turn rather than one action at a time.
Instruction following began as prompt engineering. Tool use began as code wrapped around a model. Both became training objectives.
A pattern repeats across the field. Behaviours developers once hand-built in scaffolding — delegation, escalation, context compaction — keep migrating into the weights. Meta also reports zero-shot generalisation to new native tools, Model Context Protocol servers, and custom skills, which is the same shift viewed from the integration side. Less external tooling to build by hand means agentic applications spread faster.
Who can afford to price this way
The structural question is not whether Meta can build a good model. The question is who can sustain the price. Meta, like Google, can subsidise model development, training, and inference from advertising revenue. Laboratories living on API margins and subscription income have no such cushion. Research into platform economics keeps producing the same finding: a competitor funded from an adjacent monopoly can price below cost for longer than a focused rival can survive.
Note also the reversal in posture. Meta built its reputation as the open alternative to OpenAI, shipping Llama weights when peers shipped endpoints. Muse Spark 1.1 is closed — no disclosed parameter count, no architecture, no training data or methods — and it is the first Meta model behind a paywall. Meta has moved from open weights to closed model, low price, and it is worth being clear-eyed that low prices and open access are not the same freedom. Meta is also spending accordingly on the substrate, as I covered in Meta's in-house AI chip and 14-gigawatt build-out.
For builders in Africa and across the Global South, cheaper inference is genuinely good news, and dependence is the cost written in smaller type. A price set by advertising economics can be unset by advertising economics. Emergent Intelligence (EI) — the dignity-first frame I use for what most people call AI — puts the question as one of agency: capability you rent on someone else's terms is capability you can lose in a quarter. Take the cheap tokens. Keep an open-weight fallback vetted and ready.
Frequently Asked Questions
These are the questions people are asking about Meta Muse Spark 1.1 and the AI price war. Short answers follow, drawn from Meta's own announcement and independent benchmarks.
What is Meta Muse Spark 1.1?
In short, Muse Spark 1.1 is a multimodal reasoning AI model from Meta Superintelligence Labs, launched on 9 July 2026 for agentic tasks including tool use, computer use, and coding. Meta released the model alongside the Meta Model API, and research-grade details such as parameter count and architecture remain undisclosed.
How does Muse Spark 1.1 pricing compare?
According to Meta, the model costs $1.25 per million input tokens, $0.15 cached, and $4.25 per million output tokens. Data from rival pricing shows $25 per million output tokens for Claude Opus 4.8, $30 for GPT-5.6 Sol, and $50 for Claude Fable 5, making Muse Spark 1.1 roughly six times cheaper than the nearest comparison.
Why is the price cut significant for AI agents?
The key is that token prices gate which applications are economical at scale. Analysis from Artificial Analysis puts Muse Spark 1.1 at 51 points on the Intelligence Index for $0.26 per task, confirming the discount survives per-task measurement rather than evaporating through higher token consumption.
Who can sustain pricing at this level?
In other words, companies with another revenue engine. Evidence points to Meta and Google subsidising model training and inference from advertising income, a structural advantage over laboratories dependent on API margins and subscriptions, which is why Mark Zuckerberg framed rival pricing as carrying very high margins.
What are the risks of building on Muse Spark 1.1?
The answer is dependence. Simply put, a price set by advertising economics can be reset by advertising economics, and the model is closed — no published weights, architecture, or training data. Analysis of the July 2026 Hugging Face incident shows what happens when hosted model policy blocks legitimate work, which is the case for keeping an open-weight fallback vetted and ready.
