· 4 min · Tools
Meta Joins the Coding Model Fight With Muse Spark 1.1, Its First Paid Model Ever
After years of giving Llama away for free, Meta is charging $1.25 and $4.25 per million tokens for a model that tops the field on workplace tool use and trails it on raw coding accuracy.
On July 9, 2026, Meta charged money for one of its own AI models for the first time. Muse Spark 1.1, the second release from Meta Superintelligence Labs, costs $1.25 per million input tokens and $4.25 per million output tokens — the small chunks of text a model reads and writes, and the unit every AI provider bills by — through a new public-preview Meta Model API, with $20 in free credits for new accounts.
The pricing is notable mostly for what it breaks. Meta's Llama models have been free to download and run since 2023, a strategy built around commoditizing the layer OpenAI and Anthropic charge for. Muse Spark 1.1 keeps that open-weight strategy alive for Meta's general-purpose models, but for this one — built specifically for agentic work, the kind where a model plans a multi-step task and delegates pieces of it to other model instances rather than answering one question at a time — Meta decided the API access itself was worth billing for directly.
What it's built to do
Muse Spark 1.1 ships with a 1-million-token context window with active compaction — compaction being the technique of summarizing or discarding older parts of a long conversation so a model doesn't run out of working memory during a task that spans hundreds of steps. Meta built it around orchestration specifically: acting as a main agent that gathers context, makes a plan, and delegates execution to parallel subagents, or as one of those subagents itself, understanding its assigned tool and knowing when to escalate a decision back up. Mark Zuckerberg, announcing the release in a post on X, called it "a strong agentic and coding model at a very low price" and said it was "strongest at agentic performance, tool use, and computer use" — a framing the benchmarks below mostly back up.
Where it actually wins
On JobBench, a benchmark measuring professional tool use — the kind of orchestration where a model has to read a CRM, query an API, update a spreadsheet, and hand a sub-task to another agent without losing the thread — Muse Spark 1.1 scores 54.7, ahead of Claude Opus 4.8's 48.4 and GPT-5.5's 38.3 by a wide enough margin that it's the standout number in the entire release. Meta also reports leading scores on three other agent- and tool-use benchmarks — MCP Atlas, Humanity's Last Exam, and Finance Agent v2 — though without publishing the same head-to-head numbers for those.
That combination — one main agent coordinating specialized subagents against a shared goal — is the same shape of problem Augment Code's Intent was built to manage from the tooling side rather than the model side: once several agents are running against one codebase or workflow, keeping track of who's doing what becomes its own job.
Where it falls behind
Pure coding correctness is a different story. On SWE-Bench Pro — a benchmark that grades a model's fix against a real, previously unseen pull request rather than a task it might have already seen in training — Muse Spark 1.1 resolves 61.5% of problems, independently confirmed on Scale's live public leaderboard, where it currently ranks first among publicly tested entries. That number still lands behind Grok 4.5's 64.7%, which itself trails Claude Opus 4.8's 69.2% and Fable 5's 80.4%:
The pattern across both benchmarks points the same direction: Muse Spark 1.1 is a stronger workplace-automation model than a bug-fixing one, at least measured against the field it's launching into.
Where this actually fits
Muse Spark 1.1 is available now in public preview, US-only at launch with no EU access date announced yet. For teams building agent workflows that touch business tools — CRMs, spreadsheets, ticketing systems, internal APIs — rather than pushing commits directly, the JobBench margin is large enough to be worth a real evaluation, especially at a quarter of Fable 5's or GPT-5.6's per-token price. For work where the deliverable is a correct code change reviewed against a real pull request, the SWE-Bench Pro gap says to keep routing that work to whichever model already tops that specific leaderboard — cheap orchestration and cheap code correctness aren't the same purchase, even when they're the same price tag.