Meta Charges for AI for the First Time: Muse Spark 1.1 API Undercuts OpenAI and Anthropic by 4x

Meta is charging developers for one of its own AI models for the first time ever. Muse Spark 1.1, announced on the Meta AI blog on July 9, launches with a paid developer API in public preview: $1.25 per million input tokens, $4.25 per million output tokens, $0.15 for cached input, and $2.50 per 1,000 web-search grounding queries. Every new API account gets $20 in free credits. The preview is US-only at launch.

The price is the strategy. Mark Zuckerberg — who returned to X after three years away to announce the model — pitched it as roughly a quarter of what OpenAI and Anthropic charge for comparable tiers. The comparison holds: GPT-5.5 runs $5/$30 per million tokens, and Claude Opus 4.8 runs $5/$25. At $4.25 output, Muse Spark 1.1 undercuts both by 4-7x where inference bills actually accumulate.

The model itself is a multimodal reasoning system with a 1-million-token context window and active context management — it compresses and summarizes as it goes, aimed squarely at long-running agentic sessions, tool use, computer use, and coding. And the benchmarks are not a budget model’s benchmarks. Muse Spark 1.1 leads MCP Atlas at 88.1, tops Humanity’s Last Exam at 62.1, and leads JobBench — beating both Claude Opus 4.8 and GPT-5.5 on those tests. The frontier labs keep the coding crown: on SWE-Bench Pro, Opus 4.8 leads at 69.2 while Muse Spark takes second at 61.5.

The strategic squeeze is the real story. Meta spent years giving Llama away to commoditize its rivals’ core product; Muse Spark 1.1 keeps the commoditization pressure but finally attaches a meter to it. Pure-play labs now face pricing pressure from two directions at once — a hyperscaler that can subsidize inference off a $160-billion-a-year ads business, and Chinese open-weight models racing the price floor to zero. OpenAI and Anthropic have to fund their compute from model revenue. Meta doesn’t.

The timing is pointed, too. The launch landed inside the most crowded fortnight in AI history — days after OpenAI’s GPT-5.6 family and xAI’s Grok 4.5, with Gemini 3.5 Pro arriving July 17. Meta chose this exact window to tell developers: same class of model, a quarter of the invoice.

If you run high-volume agent workloads, the math demands at least a bake-off. Meta’s first paid model won’t dethrone the frontier on hard coding tasks, but at these prices it doesn’t need to.

Sources

Related reading