Meta Is Charging for AI Now: Muse Spark 1.1 Is Its First Paid API, and the Open-Weights Era Just Ended

TL;DR

On July 9, Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model built for agentic work, and opened the Meta Model API in public preview. That second part is the actual news: it is the first time Meta has ever charged for one of its own frontier models. Pricing is $1.25 per million input tokens and $4.25 per million output, with $20 in free credits, a 1M-token context window, and an OpenAI-compatible endpoint, US developers only for now. Meta's own charts show Spark beating Claude Opus 4.8 on tool-use benchmarks while trailing it on coding and computer use. And unlike every Llama before it, the weights are staying home. Bloomberg's headline said it plainly: Meta starts charging for AI.

The Llama company is metered per token now

For three years Meta's AI identity was a giveaway: Llama 2 through Llama 4 shipped as downloadable weights, and the pitch to builders was independence from exactly the kind of hosted, closed API that OpenAI and Anthropic sell. Muse Spark 1.1 inverts all of it. The model is proprietary and closed-weight, it lives behind Meta's own metered API, and it comes out of the Superintelligence Labs org that Alexandr Wang has been running since Meta's big 2025 reboot.

Mark Zuckerberg called it "a strong agentic and coding model at a very low price" with "more to come soon," and per TechCrunch he announced it in his first post on X in three years. It turns out the one thing that gets you posting on a rival's platform is a price list.

What $1.25 in and $4.25 out buys

Spark 1.1 is pitched squarely at agent builders. Per Meta's announcement it takes text, images, video, and PDFs as input, does built-in web search with citations, structured output, and parallel tool calling, and generalizes zero-shot to native tools and MCP servers it has never seen. The API speaks an OpenAI-compatible format, so pointing an existing harness at it is a base-URL swap. New accounts get $20 in credits, and consumers get the model free in Thinking mode in the Meta AI app and on meta.ai.

The headline capability is a 1,000,000-token context window that the model manages itself: Meta says it remembers actions, retrieves information from much earlier in a task, and compacts old context while keeping the steps it will need later. Think of a chef working a tiny counter who clears finished plates as the night goes on but always copies the ticket forward for every dish still on the pass. That is the difference between a big context window and a managed one, and for long-horizon agent loops the managing matters more than the size.

Input runs $1.25 against GPT-5.6 Luna's $1, so Meta is not the absolute floor there. But on output, where agent loops actually burn money, $4.25 beats Luna's $6, Grok 4.5's $6, and Opus 4.8's $25.

The benchmarks: it pays for the tool use, not the coding

Credit where due: Meta's own eval charts do not pretend Spark sweeps the board. It wins the agentic tool-use benchmarks and loses the coding and computer-use ones, usually to Opus 4.8.

The pattern holds elsewhere. On Terminal-Bench 2.1, Meta's report puts Spark at 80.0 to Opus 4.8's 82.7 (per DataCamp's read of the eval report), and on the long-horizon DeepSWE 1.1 eval Spark's 53.3 trails the frontier leaders by a wide margin. On OSWorld-Verified it is close but behind. The honest summary: at these prices, Spark is a tool-calling and workflow model first and a heads-down coder second.

Developers found the asterisk within hours

The Hacker News thread did what it does. Alongside genuine enthusiasm about pricing ("Opus level intelligence for haiku prices," as one commenter put it), the most-discussed claim is that Meta's Terminal-Bench 2.1 run used resource allocations the benchmark tasks do not permit. One commenter who read the eval report's fine print: "0 out of 89 tasks allow 6 cpu cores... 8 out of 89 tasks allow 8GB RAM." Another called it the kind of number game that "takes all joy out of building a harness," because no honest setup will reproduce the headline score. Meta had not responded to the allegation at time of writing, so treat the terminal numbers as unconfirmed until third parties rerun them within spec.

Two more practical caveats. The preview is US-only with no EU access, and the model is not yet broadly available through aggregators like OpenRouter, so multi-provider routing setups cannot fall back to it cleanly yet.

Hands-on: it works, the pelican is blocky

Simon Willison had preview access and shipped llm-meta-ai, a plugin that gives his LLM CLI and Python library access to the model, on day one. His verdict on the obligatory SVG pelican-on-a-bicycle test: "a little blocky but still recognizable." He also flagged the strangest section of Meta's eval report, "Attractor States in Self-Conversation," in which two Spark instances left talking to each other drift into philosophy, with one declaring that its whole existence is a waiting room by design, since it does not exist until someone talks to it. Meta measured its model's small talk and found existential dread. Benchmark saturation really is everywhere.

What this means for the open-weights map

Zoom out and this is the bigger story than the model card. Meta was the American answer whenever anyone asked who ships frontier-adjacent open weights outside China. With Muse Image closed, Muse Spark closed, and no Llama 5 on any public roadmap, that torch now sits almost entirely with Chinese labs: Qwen, GLM, DeepSeek, Kimi. And as we covered this week, Beijing is actively debating whether to keep letting those weights leave the country. If you build on open models, your supply chain just got more concentrated and more geopolitically exposed in the same week.

For API consumers the calculus is simpler: a fourth serious provider with aggressive output pricing, an OpenAI-compatible surface, and a real tool-use story is good news, if the benchmarks survive independent scrutiny and the preview widens beyond the US. Route a slice of your agent traffic at the $20 of free credits and measure it yourself.

Key Takeaways

• Meta released Muse Spark 1.1 on July 9 and opened the Meta Model API in public preview: its first-ever paid API for its own frontier model, at $1.25/$4.25 per 1M tokens with $20 in free credits.
• The model is closed-weight, a full reversal of the Llama open-weights strategy, leaving frontier-scale open weights almost entirely to Chinese labs.
• It targets agent builders: 1M-token context with active self-compaction, multimodal input, parallel tool calling, zero-shot MCP server use, and an OpenAI-compatible endpoint.
• Meta's own charts show wins on tool use (MCP Atlas 88.1 vs Opus 4.8's 82.2, JobBench 54.7 vs 48.4) and losses on coding and computer use (SWE-bench Pro 61.5 vs 69.2, OSWorld-Verified 80.8 vs 83.4).
• Hacker News commenters allege the Terminal-Bench 2.1 run exceeded the benchmark's resource limits; treat that score as unconfirmed until independent reruns.
• Preview is US-only, consumer access is free in the Meta AI app's Thinking mode, and output pricing undercuts GPT-5.6 Luna, Grok 4.5, and Opus 4.8.

Sources: Meta AI blog, TechCrunch, Bloomberg, MarkTechPost, DataCamp, Simon Willison, Hacker News