The Llama Era Ended on Thursday

Mark Zuckerberg had not posted on X since the platform stopped being called Twitter. On Thursday he broke a three year silence to announce a model.

The model is the second release from Meta Superintelligence Labs, and it breaks with everything Llama stood for: closed weights, a paid API on Meta's own infrastructure, no OpenRouter. Muse Spark 1.1 shipped July 9, three months after the April original.

Input runs 1.25 per million tokens and output runs 4.25, in US dollars, with 20 in starter credits on every new account.

TechCrunch pegs that in line with, though slightly above, Claude Haiku 4.5 class pricing. Alexandr Wang, Meta's chief AI officer, described the rate as deliberately aggressive, built to scale with heavy consumption.

It capped a week Meta clearly choreographed.

So why charge at all. Meta spends on infrastructure like a hyperscaler, has no cloud business to show for it, and Wall Street wants the capex to produce revenue. Tokens are the first direct line from Meta's GPUs to a customer invoice, and the company says a cloud business is coming.

Put the two eras side by side and the pivot is stark.

The launch demo is worth thirty seconds of your time. You film a couch with your phone, and the model pulls the useful frames, reasons about the product, then drives your browser to post the Marketplace listing itself. In the computer use recipe it runs a real Linux desktop from one plain language goal, no coordinates, no click by click script.

For desktop work it decides when to write a script and when to click, batching actions instead of narrating every step. It runs as the lead agent that plans and delegates to parallel subagents, or as the subagent that stays in its lane and escalates. Inputs cover text, images, video and documents; output is text, no image generation.

For builders, the surface is deliberately boring, and that is a compliment. The Model API speaks the OpenAI dialect, so existing clients connect by swapping a base URL, and OpenCode ships a native Meta provider. Structured output, parallel tool calling, and built in web search grounding with citations come standard.

Here is the loop the model is trained to run.

Here's the part nobody's talking about. The agent harness wave of early 2026 ran on OpenClaw class runners, none of them Meta's, so instead of shipping a rival coding agent, Wang trained Spark to behave inside the harnesses engineers already use.

Planning mode, goal conditioning, subagent delegation, context compaction: it meets your stack where it already is. That is a distribution strategy disguised as a training detail.

Independent reality check: DataCamp's review finds long-horizon agentic work still trails Claude Opus 4.8 and OpenAI's flagship. And Meta's headline coding numbers come from Meta Internal Coding Bench, an internal eval nobody outside can reproduce. Credit where due, the safety card is real: evals under Meta's Advanced AI Scaling Framework report safe margins across chem bio, cyber, and loss of control, plus resistance to prompt injection from untrusted data.

I expected the benchmarks to be the story here. The business model is the story. But there is a research story too, and your feed probably skipped it.

Meta rebuilt its pretraining stack over nine months and fit scaling laws to a ladder of small models before committing big compute. The claim: matching Llama 4 Maverick capability on over 10x less pretraining compute.

The reinforcement learning curves are the quiet flex. Meta reports log linear growth in pass@1 and pass@16 as RL compute scales, which means reliability improves without collapsing reasoning diversity, and the gains hold on held out tasks. Test time reasoning gets the same discipline: thinking time penalties keep token spend lean, and Contemplating mode swaps one long chain for multiple agents reasoning in parallel.

That ladder is why the April numbers matter more than the July benchmarks. If the stack scales predictably, 1.1 is a waypoint, and Zuckerberg already teased more on the way.

Meta open-sourced the last era of AI and is charging for the next one.