Muse Spark 1.1 is Meta's first paid AI coding product, priced at $1.25 per million input tokens and $4.25 per million output, and Meta is guiding 2026 AI capex to between $125B and $145B.
Meta opened its AI to paying customers for the first time on Thursday. The company launched Muse Spark 1.1, a coding-focused model aimed at software developers and the teams that employ them, priced at $1.25 per million input tokens and $4.25 per million output tokens. It is also the first time Meta has directly challenged Anthropic, OpenAI, and Google on price inside a category that has so far been sold to premium buyers.
Meta can keep that price low because ad revenue is bankrolling the effort. The company guides $125 billion to $145 billion in AI capital expenditure this year, up from a previous range of $115 billion to $135 billion. Advertising still accounts for roughly 98% of Meta's revenue, per its most recent quarterly results. That mix turns the launch into a market-structure event: a frontier AI vendor whose training and inference costs have to come back as revenue is now competing on price against a company whose AI business does not have to break even on its own.
The product is built to drop into existing developer stacks. Muse Spark 1.1 ships through a Meta Model API in public preview that mirrors OpenAI's tool-calling shape, supports structured output and parallel tool calls, and carries a one-million-token context window. New accounts get $20 in free credits, and access at launch is limited to US developers. The compatibility layer is the wedge. AI agents, the multi-step software that has become the practical interface for coding work, currently runs on OpenAI, Anthropic, and Google's interfaces; Meta is now offering the same surface, cheaper.
Replit, Cline, and Box publicly endorsed the launch the same day. Cline, which builds AI coding agents that take multi-step actions inside a developer's editor, said the new price point makes it "easy to run heavy AI coding tasks at scale." Replit CEO Amjad Masad and Box's Yashodha Bhavnani also publicly backed the model. The pattern matters: these are the companies that integrate AI directly into developer workflows, not general enterprise software buyers, which puts Meta's distribution into a channel that already routes coding spend.
Enterprise demand is already being rationed. Coinbase, the publicly traded cryptocurrency exchange, has told engineers they can spend $500 to $5,000 a week on AI tools, a band other large enterprises are starting to mirror. That puts Meta's per-million-token price directly inside the spending window a coding team can already afford. Meta is now supplying the cheapest generic input into a market where the marginal dollar is being rationed.
CEO Mark Zuckerberg has framed the pricing as "very aggressive and attractive" and called competitor chatbot pricing "very extreme." On X and in a Bloomberg interview, he said Muse Spark 1.1 offers "frontier intelligence at a much more affordable cost", and he claimed the model outperformed Google's Gemini in several categories, an assertion Bloomberg attributes to Zuckerberg rather than to an independent benchmark. Meta AI chief Alexandr Wang called Muse Spark 1.1 "our strongest model for agentic and coding work yet." Pricing is Meta's provocation, not yet an observed market outcome. The next data point worth watching is whether Anthropic, OpenAI, or Google cut their own API rates or sharpen differentiation on quality and tool-use reliability. Meta's stock closed up about 2% on Thursday, and the company is asking investors to fund a $125B to $145B AI capex push this year to find out whether that pricing is a strategy or a subsidy.