Meta Just Became an AI API Provider — And It's Undercutting Everyone by 60%

On July 9, Meta did something it spent years telling us it would never do: it launched a paid AI API. Muse Spark 1.1 is a closed-source, proprietary model you access through Meta's developer platform at $1.25 per million input tokens and $4.25 per million output. No open weights. No community license. No download-and-self-host. Just a credit card and an API key.

If you've been building no-code AI features on OpenAI or Anthropic, those numbers should make you sit up straight.

GPT-5.5 costs $5 in, $30 out. Claude Opus 4.8 runs $5 in, $25 out. Even Claude Sonnet 4.6, Anthropic's mid-tier workhorse that most no-code builders actually use, clocks in at $3 and $15. Meta's pricing isn't just cheaper. It's a different order of magnitude.

Let me put that in a table, because the gap is worth staring at:

Model — Input (/1M tokens) — Output (/1M tokens)

Meta Muse Spark 1.1 — $1.25 — $4.25

GPT-5.5 — $5.00 — $30.00

Claude Opus 4.8 — $5.00 — $25.00

Claude Sonnet 4.6 — $3.00 — $15.00

DeepSeek V4 Pro — $1.74 — $3.48

Meta is pricing Muse Spark at roughly a quarter of what OpenAI and Anthropic charge for their flagship models. On output tokens it's about 85% cheaper than GPT-5.5. And it still undercuts DeepSeek on output, which was the previous price floor. New accounts get $20 in free credits. The API is OpenAI-compatible, so swapping providers is mostly a base URL change.

Wait, didn't Meta just kill open-weight Llama?

Yes. That's what makes this so interesting.

Back in April 2026, Meta killed off the open-weight Llama 4 line and replaced it with Muse Spark, a proprietary model built by the newly formed Meta Superintelligence Labs. The person running that division is Alexandr Wang, the 28-year-old Scale AI co-founder Meta brought aboard when it acquired a 49% stake in Scale for $14.3 billion in mid-2025. His remit, bluntly, is to turn Meta's AI research into revenue.

The original Muse Spark 1.0 had well-documented coding weaknesses. It scored 59.0 on coding benchmarks versus 80.8 for Claude and 75.1 for GPT-5.4, per the Artificial Analysis Intelligence Index. Muse Spark 1.1 is explicitly designed to close that gap: 1-million-token context window, tool calling, function calling, support for multi-agent orchestration. Early partners include Replit, Cline, and Box.

Wang claims Muse Spark 1.1 rivals GPT-5.5 and Opus-4.8 on agentic benchmarks while being 10x cheaper and twice as fast. I'd take the 10x claim with a grain of salt. Meta's comparison appears to use GPT-5.5's full output price of $30/M, which is its most expensive tier. Even so, a 4-5x saving changes the maths on per-user AI costs quite thoroughly.

So what does this actually mean for no-code builders?

Two things, and they pull in opposite directions.

The good news: your AI feature margins just got healthier. If you're running a no-code app that calls an AI model for content generation, classification, RAG, or form processing, your per-request cost just dropped by 60-75%, provided Muse Spark 1.1 is good enough for your use case. For high-volume workflows — think thousands of AI-powered form extractions or content moderation passes per day — that's the difference between a feature that burns money and one that's comfortably profitable.

I wrote recently about DeepSeek V4 Pro changing the cost equation for no-code AI features. Meta's pricing resets that equation again, from a different angle. DeepSeek's edge was being cheaper than OpenAI while matching quality on certain benchmarks. Meta's play is different: it's a company with $165 billion in annual ad revenue that doesn't actually need API income. The pricing is a weapon, not a business model.

The caution: Meta's AI strategy is not exactly coherent. Four months ago, Meta was the open-source champion, giving away Llama weights and publicly criticising closed models. Now it's selling a proprietary API. Three months from now? Anyone's guess. Building your product around a single AI provider was always a risk, but it's a particularly sharp one when that provider is still figuring out whether it wants to be a utility or a tollbooth.

This is where architecture matters more than pricing.

I've been watching no-code platforms handle AI integration in two distinct ways. Some bolt onto a specific provider: you get OpenAI features or Anthropic features baked into the platform. When that provider changes pricing, deprecates a model, or has an outage, you're along for the ride. Others, like Stacker, take a model-agnostic approach. You pick the model that makes economic sense per use case, and switching is configuration, not a rebuild.

That distinction wasn't urgent six months ago. It is now. The AI API market is fragmenting fast. Meta just entered. DeepSeek keeps compressing the price floor. OpenAI's tier structure shifts every other month. Google's Gemini 3 Flash is out there at $0.075/M input for simple tasks. If your no-code stack treats the AI model as a locked-in dependency rather than a swappable component, you're building on sand.

Why would Meta even do this?

It's not about API revenue. Meta pulled in about $165 billion last year, virtually all from advertising. A few million in API charges is a rounding error.

This is a defensive play. Every developer building on Claude or GPT is reinforcing a competitor's platform. Every AI feature built into WhatsApp, Instagram, or Facebook Messenger that runs on someone else's model is a strategic vulnerability. Meta's AI strategy under Wang is about making sure the company owns a seat at the API table, not just the open-weight one.

The $1.25/$4.25 pricing isn't designed to maximise margins. It's designed to set a floor that makes everyone else's pricing look unreasonable. Classic Meta move: commoditise the layer beneath your actual business. They did it with open-source models. Now they're doing it with API pricing.

For no-code builders, the message is simple. Cheaper AI is here, and more providers are coming. The platform you build on should let you take advantage of that without starting from scratch each time the pricing leaderboard reshuffles. Because it's going to keep reshuffling.