Meta Muse Spark 1.1: A Paid, Agentic Model That Actually Wants Your API Traffic

Meta has sold plenty of AI to advertisers. Muse Spark 1.1, which launched in public preview on July 9, is the first time it’s asked developers to pay for a model directly, by the token, the way OpenAI, Anthropic, and Google have for years. That shift matters more than any individual benchmark in the announcement.

What Meta actually shipped

Muse Spark 1.1 comes out of Meta Superintelligence Labs, the group Meta reorganized around after its restructuring push into frontier model development. It’s natively multimodal, meaning text, images, and audio run through a single architecture rather than bolted-together specialist components, and it works through problems step by step before producing an answer, the now-standard “reasoning model” pattern.

The context window is a full million tokens, on par with the larger context windows other frontier labs have shipped this year. More notable is what Meta built around the model rather than into it: native primary-agent and subagent orchestration, direct support for MCP servers, custom skills, and direct computer control. That’s the same feature set defining the current generation of coding-agent products, not the feature set of a chatbot.

The pricing tells you who it’s for

$1.25 per million input tokens and $4.25 per million output tokens puts Muse Spark 1.1 in the same tier conversation as OpenAI’s mid-range offerings, priced below a flagship model but well above a bargain-bin one. New Meta Model API accounts start with $20 in free credits, enough for real evaluation without a purchase order.

That pricing decision is a signal about ambition. Meta isn’t trying to out-price everyone at the bottom of the market; it’s trying to be a credible mid-tier option for the agentic coding and automation workloads that Cursor, Claude, and GPT-5.6’s Terra tier are already fighting over. Pay-as-you-go, metered by actual usage, is also a different business model than Meta AI’s free consumer product, and it’s worth watching whether Meta keeps them as genuinely separate products or eventually folds the API tier’s capabilities back into the free one.

The catch: US-only, waitlisted

The public preview is limited to the US and requires joining a waitlist. That’s a meaningful caveat for any team outside the US, and even for US teams it means Muse Spark 1.1 isn’t a same-day production option regardless of how the benchmarks read. If your team is picking a model for an agentic workflow this month, the realistic choice set is still Claude, GPT-5.6, and Gemini; Muse Spark 1.1 is worth requesting access to and evaluating in parallel, not worth blocking a decision on.

Where it sits against the field

Model tier	Input / 1M tokens	Output / 1M tokens	Positioning
GPT-5.6 Sol	$5.00	$30.00	Flagship, hardest coding and agentic tasks
GPT-5.6 Terra	$2.50	$15.00	Mid-tier, GPT-5.5 replacement
Muse Spark 1.1	$1.25	$4.25	Agentic coding, cost-conscious
GPT-5.6 Luna	$1.00	$6.00	High-volume, low-complexity calls

Lined up against OpenAI’s current three-tier lineup, Muse Spark 1.1 lands closest to Luna on input price but well above it on output price, and comfortably below Terra on both. That puts it in an odd but deliberate spot: cheaper than a mid-tier flagship competitor, but pitched at Terra’s job, agentic coding and orchestration, rather than Luna’s job of cheap, high-volume, low-complexity calls. Whether that pricing holds up depends entirely on how the model performs against Terra-class work once independent benchmarks and real production usage start coming in, which hasn’t happened yet given the waitlist.

What this means for teams choosing a model

The practical move here is the same one that applies every time a new entrant shows up in the agentic coding space: don’t adopt on the announcement, adopt on your own benchmark. Run your actual prompts, your actual tool-calling patterns, and your actual cost math against the new model once you have access, and compare it honestly against whatever’s currently in production. A model that scores well on a published benchmark and a model that handles your specific codebase’s conventions well are not guaranteed to be the same model.

This is also a useful moment to revisit how your team picks models at all. Teams that treat model selection as a one-time decision, made once and left alone, tend to overpay as pricing shifts underneath them or underuse a genuinely better fit that showed up later. The teams that do this well keep a standing evaluation harness, a fixed set of real prompts and expected outputs pulled from their own workload, and rerun it against every credible new entrant instead of relying on vendor benchmarks or launch-day hype. That harness is cheap to build once and pays for itself the first time a $1.25/$4.25 model turns out to be a better fit for a workload than the $5/$30 one it’s tempting to default to.

What Muse Spark 1.1 does confirm is where the competitive pressure in this market currently sits: not consumer chat, but agentic coding and orchestration, priced by the token, competing directly on cost-per-capability against Anthropic, OpenAI, and Google. Meta joining that fight, on those terms, with real metered pricing instead of a subsidized freebie, is the actual news here. The specific benchmark numbers will look dated within a quarter; the fact that Meta decided this market was worth charging for won’t.