Meta Starts Selling Tokens: The Open-Weights Champion Opens a Closed API

On July 9, 2026 — a day so crowded with model launches that it will probably be remembered for someone else's product — Meta did something more interesting than ship a model. It started charging for one.

Muse Spark 1.1, released by Meta Superintelligence Labs, is available two ways: free in "Thinking" mode inside the Meta AI app and on meta.ai, and metered through a new Meta Model API in public preview. The second half is the news. For the first time, Meta is renting out an in-house foundation model to outside developers, per-token, behind an API it controls.

For a company that spent years as the loudest voice for open weights, that is a considerable turn of the wheel — and it is being made quietly, in a blog post that spends most of its energy on agentic benchmarks.

What is actually in the box

Per Meta's announcement, Muse Spark 1.1 is a multimodal reasoning model aimed squarely at agentic work. Meta claims "major gains in tool and computer use, coding, and multimodal understanding" over its predecessor, and says the model handles complex projects "significantly faster than Muse Spark."

The concrete pieces:

• A 1 million token context window — MarkTechPost notes the API documentation lists 1,048,576 tokens as the maximum — with what Meta calls active context management.
• Multimodal perception across images, video, and PDFs.
• Computer use, with the model deciding on its own whether to write a script or click through an interface.
• Multi-agent orchestration for parallel execution.
• Zero-shot generalization to new tools and MCP servers — arguably the most load-bearing claim in the post, and the hardest to verify.

The API is OpenAI-compatible. MarkTechPost reports that integration amounts to changing a base URL. That is not an engineering detail; it is the entire go-to-market strategy. Meta is not asking developers to adopt a new SDK, rewrite their agent loops, or learn a new tool-calling schema. It is asking them to change one string and watch the bill go down.

Third at coding, first at tools

Here the sources diverge, and the divergence is worth sitting with.

Meta's blog post presents its evaluations as charts rather than a table of figures in the body text — the numbers are shown, not stated. AI Chat Daily, reviewing the launch, complains that Meta published "no SWE-bench score, no agent-benchmark result" to validate its claims. MarkTechPost, reading the same material, extracts specific figures:

• MCP Atlas (tool use): 88.1
• JobBench (professional tool use): 54.7
• Humanity's Last Exam (reasoning with tools): 62.1
• SWE-Bench Pro: 61.5, against Opus 4.8's 69.2
• DeepSWE 1.1: 53.3, against GPT-5.5's 67.0

Take those as reported rather than settled. They are Meta's own numbers, read off Meta's own charts, and no independent evaluation existed at publication time. But if they are approximately right, they tell a clear and rather honest story: Muse Spark 1.1 is third-best at coding and first at using tools.

That is a flank attack, and a smart one. Anthropic owns paid coding workloads. OpenAI owns the developer base. Meta is not pretending to beat either head-on. It is betting that the next contested surface is not writing the patch but running the agent — orchestrating tools, driving interfaces, holding a million tokens of state across a long session. So it optimized for MCP Atlas and JobBench, benchmarks far younger and far less gamed than SWE-Bench, and it led with them.

The strategic logic is sound. The epistemics are shakier: a model leading on benchmarks that few people have learned to distrust yet is a weaker signal than a model placing third on one everybody has been attacking for years. Meta's SWE-Bench Pro number, the one where it loses, is the number I trust most.

The price is the product

Muse Spark 1.1 costs $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits on new accounts, according to MarkTechPost.

You will see this framed as a fraction of what OpenAI and Anthropic charge. Be careful with that framing — it depends entirely on which competing tier you pick. AI Chat Daily places the $1.25 input price slightly above Claude Haiku 4.5 and GPT-5.6 Luna — that is, above the cheap tiers, not below them. Against flagship tiers the discount is real and large. Against the fast, cheap models developers actually reach for in high-volume pipelines, Meta is not undercutting anyone.

So the pitch is not "cheapest." It is frontier-adjacent capability at near-budget-tier prices: a million-token context, multimodal input, parallel tool calling, and computer use, at roughly what you would pay for a small model. Whether that holds depends on whether the agentic benchmark lead survives contact with real workloads.

Early access partners are talking like it does. Meta's post quotes Replit's Amjad Masad calling it a "complete agentic foundation," Box's Yashodha Bhavnani citing "enterprise capabilities competitive with today's leading frontier models," Cline's Saoud Rizwan praising tool use at pricing viable for coding workloads, and OpenClaw's Dave Morin calling it an "awesome model for running agents." These are design partners describing a product they helped shape. Weight accordingly.

What Meta is not saying

Three silences stand out.

Geography. The preview is US-only, with no EU access yet, per MarkTechPost. For a model whose selling point is enterprise agentic workflows, locking out the EU on day one is a meaningful constraint — and a quiet acknowledgment of regulatory drag.

Llama. Neither Meta's post nor the coverage I read addresses what happens to the open-weights line. Muse Spark is closed, hosted, and metered. Llama was none of those. Meta has not said Llama is over; it has simply stopped talking about it in the announcement that matters. Read that as you like.

Why now. Meta has never needed an API business. It has Instagram, WhatsApp, and default distribution to billions. The move to sell tokens is best read as defense of the tens of billions it has sunk into AI infrastructure — converting compute into developer revenue rather than waiting for it to pay off through ad ranking. But the API market is the one arena where Meta's distribution advantage evaporates completely. Switching costs there are one base URL wide, which is exactly the door Meta just walked through, and exactly the door developers can walk back out of.

The hype-to-substance ratio

Real: the model exists, the API is live, the context window is large, the OpenAI compatibility is genuine, the pricing is aggressive against flagships, and the agentic positioning identifies a real gap.

Unproven: every benchmark figure is self-reported; "zero-shot generalization to new tools and MCP servers" is a sweeping claim with no published methodology; the partner quotes are from design partners; and "public preview" is doing real work in a US-only launch.

Unconfirmed: the widely repeated claim that Zuckerberg pegged the pricing at roughly a quarter of what Anthropic and OpenAI charge does not appear in any source I was able to fetch directly, and one of them explicitly notes its absence. Treat it as unverified.

The takeaway

The model is the second story. The first is that Meta — the company that made open weights a competitive weapon and a political identity — now has a metered endpoint, a price sheet, and design partners. It entered on the axis where the incumbents are least entrenched, priced against their flagships while quietly sitting above their budget tiers, and made switching cost exactly one line of configuration.

That last choice cuts both ways, and Meta knows it. An OpenAI-compatible API is a bet that you win on price and capability, because you have surrendered every other kind of lock-in. It is the most confident thing about this launch, and the most exposed.

Watch two numbers over the next month: whether anyone outside Meta reproduces the MCP Atlas and JobBench results, and whether the EU gets access. The first tells you if the product is real. The second tells you how much Meta believes in it.