Meta Starts Charging for AI: Muse Spark 1.1 Arrives With a 1M-Token Context

The first model sold through Meta's new Model API is an agentic coder priced far below OpenAI and Anthropic — and Zuckerberg returned to X after three years to claim Meta's models finally beat Google's.

Jul 10, 2026Source: TechCrunch

Meta Starts Charging for AI: Muse Spark 1.1 Arrives With a 1M-Token Context

Key Points

1.Meta released Muse Spark 1.1 on July 9 as the first model on its new Meta Model API, the company's first pay-to-use proprietary model; it also runs in 'Thinking' mode inside the free Meta AI app.
2.Meta says the model actively manages a 1-million-token context window and 'excels at computer-use workflows' — writing scripts, clicking interfaces and navigating unfamiliar software with minimal human intervention. Reuters reports pricing of $1.25 per million input tokens and $4.25 per million output, roughly a quarter of OpenAI and Anthropic rates.
3.Third-party trackers put Muse Spark at 80.0 on Terminal-Bench 2.1 — behind GPT-5.5 (83.4) and Claude Opus 4.8 (82.7) — but ahead of both on the MCP Atlas and JobBench tool-use tests; Meta's own blog publishes charts without figures.

Meta has spent years giving models away. On July 9 it started selling one. Muse Spark 1.1, an agentic coding model, is the debut offering on the new Meta Model API — the first time the company has charged for API access to a model of its own — and it lands in the same week OpenAI and SpaceXAI cut prices on theirs.

What Meta claims

Per Meta's announcement, Muse Spark 1.1 "can actively manage its context window of 1 million tokens," deciding what to retain, retrieve or compress across long sessions, and it "excels at computer-use workflows" — writing scripts, clicking through interfaces, generating batches of actions and finding its way around unfamiliar software with little supervision. The model is also live in "Thinking" mode in the free Meta AI app and on meta.ai.

The price

Meta did not publish a rate card. Reuters reports $1.25 per million input tokens and $4.25 per million output, plus $20 in free credits — roughly 25% of what OpenAI and Anthropic charge for their flagships. That is the entire pitch. Muse Spark is not the best model available and Meta is not pretending otherwise; it is trying to make the good-enough tier too cheap to ignore, a move Meta's own AI chief followed with a promise of "more to come."

The scoreboard

Meta's blog shows benchmark charts without numbers. Third-party trackers fill them in: 80.0 on Terminal-Bench 2.1, behind GPT-5.5 (83.4) and Claude Opus 4.8 (82.7), and 61.5 on SWE-Bench Pro, ahead of GPT-5.5 (58.6) but behind Opus 4.8 (69.2). Where Muse Spark leads is tool use: 88.1 on MCP Atlas for multi-step tool calls, and 54.7 on JobBench against Opus 4.8's 48.4 and GPT-5.5's 38.3. Replit CEO Amjad Masad, quoted on Meta's blog, said: "What's most impressive about Muse Spark is how much it packs into one model."

Zuckerberg breaks silence

Mark Zuckerberg posted on X for the first time in three years to announce it: "Today we're releasing Muse Spark 1.1 — a strong agentic and coding model at a very low price." In remarks accompanying the launch he went further, on Google: "This may be the first time, at least that I can remember, that Meta's models are better than all of the Google models." The release comes with Meta's AI effort reorganized under Alexandr Wang's Superintelligence Labs.

Why It Matters

Meta charging for a model ends the strategic ambiguity of the Llama era, when free weights were a weapon against rivals' pricing power rather than a business. Undercutting OpenAI and Anthropic by roughly four-to-one on an agentic coder — while leading them on tool-use benchmarks — is a bid to own the high-volume, unglamorous end of agent workloads, where cost per task, not benchmark rank, decides who gets deployed.

Filed underMeta AI AI Agents AI Models #Mark Zuckerberg #Meta #Muse Spark #agentic AI