# Meta Puts Its Frontier Model Behind a Paid API, Closing the Weights on Muse Spark 1.1

The multimodal agentic model comes with a one-million-token context window and metered pricing of $1.25 per million input tokens, ending Meta's open-weight approach at the frontier.

- Published: 2026-07-11T06:01:51.690Z
- Canonical: https://polylog.news/ai/2026-07-11/meta-puts-its-frontier-model-behind-a-paid-api-closing-the-w
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/), [MarkTechPost](https://www.marktechpost.com/2026/07/09/meta-superintelligence-labs-releases-muse-spark-1-1/)

Meta Superintelligence Labs, the unit led by Alexandr Wang, released [Muse Spark 1.1](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/) on July 9. It is a multimodal reasoning model built for tool use, computer control, and coding. The main improvement over the earlier Muse Spark is in agentic execution rather than raw text quality. The model has a one-million-token context window, and Meta compares its gains against Anthropic's Claude and OpenAI's frontier models.

The more important development is commercial, not technical. Unlike the Llama family, [Muse Spark 1.1 has closed weights](https://t.me/ai_machinelearning_big_data/10498) and is available only through the paid Meta Model application programming interface (API) in public preview. It is priced at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in credits, according to [Meta's launch materials and independent write-ups](https://www.marktechpost.com/2026/07/09/meta-superintelligence-labs-releases-muse-spark-1-1/). For the first time, Meta is charging for access to its own frontier model the way its rivals do, rather than releasing downloadable weights.

Independent reproduction of the model's agentic and coding claims is not yet available. The published numbers come from Meta itself, and how Muse Spark 1.1 compares with Claude and GPT-class models on public agent benchmarks such as SWE-bench and computer-use test suites has not yet been confirmed by outside researchers. What is verified today is the pricing, the closed distribution, and the one-million-token context window.

## What this means

Meta reserving its best model for a paid API ends the assumption that the company's frontier work stays open. The exposed parties are open-weight adopters who standardized on Llama for downloadable control, and the AI coding market, where Meta now competes directly with Anthropic and OpenAI on price and agentic capability rather than on openness. The channel is distribution. Meta gains a metered source of revenue and gives up the developer goodwill it earned by releasing open weights.

## What to watch

- Independent SWE-bench and computer-use scores for Muse Spark 1.1 against Claude and GPT-class models, which would show whether the agentic claims hold up outside Meta's own internal tests.
- Whether Meta keeps releasing open Llama weights alongside a closed frontier tier, which would indicate a permanent two-track strategy rather than a single exception.