Meta Puts a Price on Its AI: Muse Spark 1.1 Opens to Paying Developers

The company that built its AI reputation on free weights is now metering tokens. Alexandr Wang says the rate is aggressive. The market will decide.

Meta released Muse Spark 1.1 on Thursday and attached something to it the company has never attached to an AI model before: an invoice.

The model is priced at $1.25 per million input tokens and $4.25 per million output tokens, with new accounts receiving $20 in free credits. Access runs through a developer portal in public preview, where some early partners are already onboarded and everyone else joins a waitlist to be added over time. The model also went live the same day in "Thinking" mode inside Meta AI.

That pricing line is the story. Everything else about the launch is a capability claim that will be litigated by developers over the next few weeks.

The Reversal

The first Muse Spark, internally called Avocado, shipped in April under a closed partner program with no public-facing API. Before that, Meta's models reached developers as open weights under the Llama banner, a strategy that bought the company enormous goodwill and, eventually, an embarrassment. Llama's spring 2025 release landed poorly, and Zuckerberg responded by essentially starting over, poaching Wang from Scale AI to run the effort, cutting staff and reshuffling teams.

Muse Spark 1.1 is closed, hosted, and billed per token.

Wang insists the open-source commitment survives. He told CNBC that Meta Superintelligence Labs has a variant of Muse Spark in development that it intends to open source, and declined to say when. No date, no weights, no repository. For now, the free-model era at Meta exists as a promise about a model nobody outside the building has seen.

The strategic logic is not hard to follow. Open weights established Meta's credibility with developers, but hosted APIs generate recurring revenue and keep inference workloads inside Meta's own infrastructure rather than routing them to cloud providers or rival platforms. Meta's 2026 capital expenditure sits at record levels, with hundreds of billions committed to compute and data centers, including a $10 billion facility in Canada. Somebody has to pay for that.

A Number Designed to Undercut

Wang called the pricing "very aggressive and attractive" against comparable offerings from Anthropic and OpenAI. Zuckerberg went further, telling Bloomberg the rate amounts to roughly 25% of what those two labs charge for equivalent models.

The comparison depends heavily on which model you pick as the equivalent.

TechCrunch placed Meta's rate slightly above Anthropic's Claude Haiku 4.5 and OpenAI's GPT-5.6 Luna, both of which sit at the cheap end of their respective catalogs. Against frontier-tier models, the discount is real. Against the budget tier Meta is actually priced near, it disappears. Zuckerberg's 25% figure is doing quiet work by choosing its denominator.

The more useful metric will not be cost per token at all. It will be cost per completed task. An agentic model that burns four times the turns to close a bug ticket is not cheaper than an expensive model that closes it once, and agentic workloads are where Meta has staked this release.

Wang framed the target directly: pricing that scales with heavy consumption. Volume, in other words. Meta is betting it can absorb thin margins on inference in exchange for developers who stop reaching for a competitor's endpoint by reflex.

Coding as Load-Bearing Infrastructure

Ask why a social media company suddenly cares about code migrations and the answer Wang gives is that it does not, particularly.

Strong coding ability, in his telling, is a prerequisite for the multi-step autonomous behavior that defines agentic AI, which makes the two capabilities inseparable. He described the end state as agents that autonomously handle multiple tasks "like a fleet of human interns."

Coding is the training substrate. The agent is the product.

That framing explains an otherwise odd set of engineering decisions. Meta says the model can actively manage a context window of one million tokens, retrieving information from earlier work and compacting it in a way that preserves the steps needed later. It also says the model orchestrates multi-agent systems to cut end-to-end latency, acting as a main agent that gathers context and delegates to parallel subagents, or as a subagent that knows when to escalate.

Wang said MSL deliberately optimized for the harnesses developers already use, on the theory that compatibility would drive adoption faster than raw benchmark scores. That is an unusually candid admission that the fight is over ecosystem lock-in.

The Benchmark Problem

Here is where a reader should slow down.

Meta's headline coding claim rests on Meta Internal Coding Bench, its own primary internal evaluation, on which the company says Muse Spark 1.1 improves substantially over its predecessor and is competitive with leading alternatives. A private benchmark, scored by the vendor, published by the vendor.

Wang was bolder on X, writing that across many agentic evals the model "rivals gpt-5.5 and opus-4.8." Zuckerberg, per Bloomberg, offered a version aimed at a different rival: "Meta's models are better than all of the Google models," a first, he said, as far as he could remember.

None of that is independently verified yet.

The partner testimonials carry the same caveat. Replit, Cline and Box appear as early access partners, and each supplied a favorable line for Meta's announcement post. Replit CEO Amjad Masad called the model "A complete agentic foundation." Cline's chief executive praised the tool use at a price point workable for production coding loads. These are companies with pre-release access and a commercial interest in a cheaper inference option existing. Their enthusiasm is data. It is not an audit.

On safety, Meta says it evaluated the model under its Advanced AI Scaling Framework and found it operating within safe margins across frontier risk categories covering cybersecurity and loss of control, alongside chemical and biological risk. The company also reports strong resistance to direct jailbreaks and to indirect attacks arriving through untrusted data, with lower hallucination rates and reduced sycophancy. The supporting evaluation report is Meta's own.

Distribution With a Wall Around It

For a launch pitched as an opening, Meta has kept the doors narrow.

The company is limiting API access to its own properties for now, declining to list the model on third-party marketplaces such as OpenRouter. "This is going to be served on top of the computer infrastructure that we've built," Wang said.

Undercutting on price while restricting where the model can be bought is a hedge rather than a full commitment to the developer market Anthropic and OpenAI have spent years cultivating. Meta wants to find out whether cheap tokens pull developers across before it commits to wider distribution.

There is a second distribution channel, and it is the one nobody else can copy. Muse Spark already powers Meta's assistant across the standalone app, with rollouts into Facebook, Instagram, WhatsApp, Messenger and the Ray-Ban Meta AI glasses. Whether or not a single developer signs up for the API, the model reaches billions of people through products they already have installed.

Zuckerberg Broke a Three-Year Silence for This

A small detail that says more than the benchmarks.

The launch was significant enough to prompt Mark Zuckerberg to post on X for the first time in three years, his last post dating to July 2023, around the platform's rebrand from Twitter. He kept it short: "a strong agentic and coding model at a very low price."

Reading the market reaction, the monetization angle is what analysts fixed on. Shay Boloor, chief market strategist at Futurum Equities, told Reuters that a genuinely competitive coding model would give Meta a clearer "monetization bridge from AI models to paid developer tools."

Meta has spent years converting AI research into advertising performance. It has never converted AI research into a line item customers pay for directly.

A Loud Week, and a Bigger Model Behind It

Muse Spark 1.1 did not arrive in quiet air. SpaceXAI shipped a new version of Grok on Wednesday, OpenAI dropped its GPT-5.6 family the same Thursday, and Meta itself had released Muse Image, originally code-named Mango, two days earlier.

Four frontier announcements in five days.

And Meta has already told everyone this is not the release that matters. A substantially more powerful successor, code-named Watermelon, is still in training on vastly more compute, expected later this year. Wang confirmed it will succeed Muse Spark 1.1.

Which recasts what happened Thursday. The pricing page, the waitlist and the developer portal are scaffolding, built now so that the model Meta actually expects to compete with has somewhere to land when it arrives.