Meta launches paid subscriptions for its Muse Spark 1.1 AI model, undercutting OpenAI and Anthropic on developer API pricing by roughly 75%.
Meta launched a paid API for developers on Thursday, opening its Muse Spark 1.1 model to businesses willing to pay for access. It is the first time the company has charged outside developers for one of its AI models, and the move puts Meta squarely in competition with Anthropic and OpenAI for the developer tools market.
What Meta Launching Paid Subscriptions for Its AI Model Actually Means
Meta launched paid subscriptions for Muse Spark 1.1 through a developer portal that is currently in public preview. Pricing runs at $1.25 per million input tokens and $4.25 per million output tokens, with new accounts getting $20 in free credits to start. CEO Mark Zuckerberg told Bloomberg that figure amounts to roughly a quarter of what Anthropic and OpenAI charge for similar models. Alexandr Wang, who runs Meta Superintelligence Labs, called the pricing "very aggressive and attractive" in comments to CNBC.
Some developers have already been onboarded as early partners. Others will need to join a waitlist and wait for an invite. Meta says it has no plans right now to distribute the model through third party platforms like OpenRouter, keeping usage tied to its own infrastructure for the moment.
How It Stacks Up Against Anthropic and OpenAI
The comparison Meta wants people to make is on price and capability together, not just cost. Muse Spark 1.1 is built for agentic and coding tasks, the kind of work where an AI system carries out a string of actions on its own rather than answering a single prompt. It handles a context window of one million tokens and can run several subagents at once. Zuckerberg described its agentic reasoning and tool use as state of the art or close to it, and said internal testing showed it beating Google's Gemini on benchmarks covering agents, coding, and multimodal tasks.
Wang argued that solid coding skill is basically a precondition for agentic behavior, since multi step autonomous tasks and programming ability tend to rise and fall together in these models.
Early Partners and Internal Use
- Replit, Cline, and Box are named as early access partners
- Meta employees have reportedly already used the model to build new features across the company's apps
The Road From Llama to Avocado and Watermelon
This release traces back to April, when Meta stepped away from its open source Llama approach and introduced the original Muse Spark as a proprietary model limited to select partners through a private preview. Coding was a known soft spot in that first version. Thursday's model, developed internally under the codename Avocado, addresses that gap. Wang said an open source version is being worked on, though he would not say when it might arrive.
Meta is also training a further model under the codename Watermelon. Zuckerberg and Wang both confirmed its existence but declined to give any sense of timing.
Why Meta Is Pushing So Hard on AI Right Now
Thursday's API rollout arrived two days after Meta released Muse Image, a consumer and advertiser facing tool for generating images. Together the two launches read as an attempt by Meta to show tangible payoff from its AI spending, at a moment when Zuckerberg is fielding pointed questions from investors about the company's ballooning infrastructure budget.
Will Aggressive Pricing Be Enough to Win Developers Over
Meta's bet is that undercutting Anthropic and OpenAI on price, while claiming comparable or better performance on agentic and coding benchmarks, will be enough to pull developers toward its ecosystem. Keeping distribution limited to Meta's own platforms for now suggests the company wants tight control over the early rollout before deciding whether to widen access. Whether outside developers embrace a closed distribution model, even at a lower price, remains the open question hanging over this launch.
