Meta Launches Paid Muse Spark 1.1 API, Its First Charge for an AI Model

Meta opened a paid developer API for its Muse Spark model on Thursday and released an upgraded version, Muse Spark 1.1, aimed at coding and agentic work. It is the first time Meta has charged businesses to use one of its AI models, and it puts a low-priced new supplier in front of enterprise teams that have mostly been choosing among OpenAI, Anthropic and Google.

US developers can now test Muse Spark in public preview on the Meta Model API, prototype integrations and compare outputs, Reuters reported. New accounts get $20 in free credits before moving to pay-as-you-go pricing of $1.25 per million input tokens and $4.25 per million output tokens.

What can Muse Spark 1.1 do?

Meta calls Muse Spark 1.1 its most capable model for real-world coding and agentic tasks. According to Reuters, it can write and debug code, operate software and external tools, understand text, images and video, and carry out multi-step tasks with less human intervention.

The model has a 1-million-token context window and can run several subagents at the same time, Quartz reported. Chief executive Mark Zuckerberg said it outperformed Google's Gemini on agent, coding and multimodal benchmarks, and that Meta staff already use it to build features across the company's apps. Those are Meta's own claims, not independent evaluations.

Early access partners named by Quartz include Replit, Cline and Box. Others can join a waitlist. Meta said it will not offer the model through outside platforms such as OpenRouter for now.

How does Meta's pricing compare with OpenAI and Anthropic?

Price is the pitch. "Our focus is on delivering strong agentic and multimodal models at very low cost," Zuckerberg wrote in a post on X, as quoted by Reuters.

Meta's rates sit above entry-level models such as OpenAI's GPT-5 mini and Anthropic's Claude Haiku 4.5 but below Anthropic's Claude Sonnet 4.6, Reuters reported. Zuckerberg told Bloomberg the price is roughly a quarter of what Anthropic and OpenAI charge for comparable models, according to Quartz. Whether models are truly comparable depends on the task, so enterprise teams will want to test on their own workloads rather than rely on list prices.

The launch landed in the same week that OpenAI made its GPT-5.6 family generally available in three price tiers, which makes the summer of 2026 a buyer's market for teams willing to evaluate more than one provider.

Why is Meta moving away from open Llama models?

The paid API completes a shift that began in April, when Meta introduced the original Muse Spark as a proprietary model available only to select partners through a private preview. That was a break from the open-weight Llama strategy that made Meta popular with enterprise developers who wanted to run models on their own infrastructure. Coding was a relative weakness of the first release, Quartz noted.

Meta has not abandoned openness entirely. Alexandr Wang, who leads Meta Superintelligence Labs, said an open-source version of Muse Spark is in development, without giving a date, according to Quartz. A next-generation model is also in training. Muse Spark 1.1 is expected to replace the Llama models behind Meta's chatbots on WhatsApp, Instagram, Facebook and its smart glasses, Reuters reported.

The commercial motive is plain. Zuckerberg faces pressure from investors over Meta's infrastructure spending, and charging for developer access is one way to show a return on it.

What should enterprise AI teams do with a new model provider?

A credible, cheaper model is good news for buyers, but it is a preview from a company new to selling models to businesses. Before moving production workloads, teams should check data handling and retention terms, regional availability (access is US-only for now), rate limits, support commitments and how the API fits their existing security reviews.

The bigger lesson is architectural. Enterprises that hard-wired one provider into their agents now face a choice between switching costs and paying more than they need to. Stefan Leichenauer, VP of engineering at SandboxAQ, warned in a AI Connect webinar that a chosen vendor "may not even exist two years from now," and urged teams to use open frameworks and test several vendors, as covered in our piece on future-proofing AI infrastructure against lock-in.

Model choice should also follow the job. Mistral AI's Aaron Delp told a AI Connect webinar audience that the best return comes from "the right model at the right time in the right place," a point developed in our coverage of right-sizing enterprise agents.

What comes next?

Watch for independent benchmark results, the terms Meta offers enterprise customers once the preview ends, and whether the API expands beyond the United States. If Meta's pricing holds up against real coding and agent workloads, it will add pressure on rivals' prices for the second half of the year.

Frequently asked questions

How much does the Muse Spark API cost?

$1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new accounts, according to Reuters.

Who can use Muse Spark 1.1?

US developers can access it in public preview on the Meta Model API; early partners include Replit, Cline and Box, and others can join a waitlist.

Is Muse Spark open source like Llama?

No. Muse Spark is proprietary, though Meta Superintelligence Labs head Alexandr Wang said an open-source version is in development without giving a date.

Filed under: Agentic AI, AI Strategy & ROI