Meta launches Muse Spark 1.1 at $1.25 per million tokens, undercutting rivals

Meta Platforms Inc. on Thursday released Muse Spark 1.1 with a public developer API priced at $1.25 per million input tokens, undercutting OpenAI and Anthropic as it pivots from open-source to paid AI services.

"The goal is to really have attractive pricing that scales with immense consumption usage," Alexandr Wang, head of Meta Superintelligence Labs, said in an interview.

The model costs $1.25 per million input tokens and $4.25 per million output tokens, below OpenAI's entry-level GPT-5 mini and Anthropic's low-cost Claude Haiku 4.5, but above Anthropic's higher-end Claude Sonnet 4.6. Every new API account receives $20 in free credits to test the model before switching to pay-as-you-go pricing. Wang characterized the pricing as "very aggressive and attractive" compared with similar offerings from rival labs. The model is now available in Thinking mode in the Meta AI app and on the website, and is expected to replace existing Llama models powering chatbots on WhatsApp, Instagram, Facebook and Meta's collection of smart glasses.

Meta shares fell as much as 4% intraday before recovering to a 0.5% decline, reflecting mixed investor sentiment over the company's massive AI spending versus its new monetization strategy. Meta does not operate a cloud infrastructure business, though it plans to start one, leaving it reliant on API revenue to justify an AI investment pace matching hyperscaler peers such as Microsoft Corp. and Amazon.com Inc.

Competitive Positioning and Strategic Shift

Meta's shift marks a departure from its Llama family of open-weight models released throughout 2024 and 2025, which anyone could download, modify and deploy. The company is now selling access to proprietary AI through Meta Model API, available in public preview to developers in the United States. Wang said MSL has "a variant of Muse Spark that is in development that we do intend to open source," though he declined to provide a timeline.

Wang said Meta trained Muse Spark 1.1 to excel in coding-related tasks because that ultimately improves the capabilities of AI agents that can autonomously perform multiple tasks. The model can write and debug code, use software and external tools, understand text, images and video, and carry out complex multi-step tasks with less human intervention, according to Meta. Wang said the company trained the model "to be able to work well with all of the most popular harnesses that developers use today," including OpenClaw, the developer tool that gained popularity in early 2026.

The Muse Spark 1.1 launch follows Tuesday's release of Muse Image, the first image-generation model from Meta Superintelligence Labs, as the company seeks to attract creators and advertisers. Meta is also training a more powerful model code-named Watermelon, Wang said, without disclosing a release date. The initial Muse Spark model, code-named Avocado, launched in April with access limited to select partners through a private API preview.

Investment Implications

Meta's API pricing strategy directly challenges OpenAI and Anthropic on cost, potentially compressing margins across the AI model market. The company trades at a discount to hyperscaler peers that operate cloud businesses, and the new API revenue stream could help close that gap if adoption scales. Chief Executive Officer Mark Zuckerberg is under pressure from Wall Street to show a return on the company's massive and growing investment in AI infrastructure and development. Wang said he has been using Muse Spark for personal health research, calling it "one of these use cases that really encapsulates the needs of these agentic systems."

This article is for informational purposes only and does not constitute investment advice.