Meta Joins the AI Price War With Muse Spark 1.1 and Its First Developer API
Meta's new reasoning model comes with the company's first public developer API — priced at $4.25 per million output tokens, undercutting every major US rival.
Meta has released Muse Spark 1.1, an upgraded reasoning model built for agent tasks, coding, and computer use — and for the first time, it’s letting outside developers build on its frontier model directly. The new Meta Model API launches in public preview at prices that undercut every major US competitor: $1.25 per million input tokens and $4.25 per million output tokens.
For scale: tokens are the small text chunks AI providers charge by, and a million of them is roughly 750,000 words. Anthropic’s Opus 4.8, OpenAI’s GPT-5.5, and Claude Fable 5 charge between $25 and $50 per million output tokens — many times Meta’s price. Even xAI’s Grok 4.5, which launched just a day earlier and briefly held the “cheapest near-frontier model” title, is now undercut. On the model itself: Meta says Muse Spark 1.1 is trained to orchestrate multi-agent systems — acting as a lead agent that plans and delegates to parallel helpers — and manages a context window of one million tokens, meaning how much material it can keep in mind at once. In Meta’s own benchmark table it leads four of twelve tests; in the independent VALS-AI ranking it lands fourth overall, notably fast and cheap. Like its predecessor, it ships without open weights — the Llama-era Meta that gave its models away is gone.
What’s behind this? Meta earns over $60 billion a year from advertising, so it can run an AI API as a loss-leading gateway into its ecosystem. OpenAI and Anthropic can’t: they burn billions and need healthy token margins to justify their valuations. That’s the real story here — the “pure-play” AI labs are now squeezed from two directions, with Chinese open-source models like GLM 5.2 pushing prices up from the bottom and profitable giants like Meta and Google pressing down from the top. A fair caveat: a low sticker price isn’t automatically a low bill. How many tokens a model burns per task varies widely, and Meta’s benchmark promises still have to prove themselves in production before anyone should migrate.
What this means for you: If you just chat with AI apps, this price war is quietly good news — it’s why free tiers keep getting more generous and why features like agents reach cheap plans faster. If you build things on AI APIs, Meta’s entry gives you real leverage: even if you never switch, quoting these prices to your current provider is now a negotiation tactic. For most people: nothing to do today, but the direction is clear — intelligence keeps getting cheaper.
Sources
An OpenAI Model Beat Every Human at One of Coding's Hardest Contests
At the AtCoder World Tour Finals 2026, OpenAI's system solved all five algorithm problems and doubled the best human's score — a year after finishing second.
