On 9 July 2026, Meta launched Muse Spark 1.1, an AI model built for coding and agentic tasks, alongside the company's first paid developer API. According to CNBC, Meta's AI chief Alexandr Wang called it Meta's strongest model for that kind of work yet. It is a direct move into the market led by Anthropic and OpenAI - the same market our AI coding agent guide covers.
The pitch: aggressive pricing
The headline is price. Per CNBC and MLQ News, every new API account starts with $20 in free credits, and usage is billed at $1.25 per million input tokens and $4.25 per million output tokens. Wang described the pricing as "very aggressive and attractive" compared with similar offerings from Anthropic and OpenAI, and claimed the model "rivals GPT-5.5 and Opus-4.8" across agentic benchmarks while being "10x cheaper and twice as fast."
Those benchmark numbers are Meta's own claims, not independent results, so treat them as marketing until third parties test them. Meta named Replit, Cline and Box as early API partners.
The quiet part: closed weights
There is a strategic reversal here. Meta spent years distributing its Llama models freely to the open-source community. Muse Spark 1.1 goes the other way: reporting describes it as proprietary closed weights, accessible only through Meta's apps or the paid API. Wang said Meta remains "committed to open source" and has a Muse Spark variant it intends to open source, but the flagship is now a paid product.
The launch fits the wider picture: Mark Zuckerberg is under Wall Street pressure to show a return on Meta's enormous AI spend, after the company fell behind OpenAI, Anthropic and Google on popular models.
What developers should actually weigh
A new low-priced entrant is good for everyone - but price is only one axis. Before you rewire your workflow around any AI coding tool, weigh a few things, the same way our best coding LLMs and Cursor vs Claude Code comparisons do:
- Proven quality over claimed quality. Wait for independent agentic benchmarks before trusting the "rivals Opus-4.8" line.
- Lock-in. Closed weights mean you depend on Meta's API and pricing staying attractive; there is no self-host fallback for the flagship.
- What leaves your machine. Any cloud coding model receives your prompts and code context. If you handle sensitive or regulated code, that is an AI agent security question, not just a price one.
- Real cost. Cheap per-token pricing can still add up on agentic workloads that burn large token volumes; measure on your own usage.
The honest summary: Muse Spark 1.1 makes the AI-coding market cheaper and more competitive, which is worth watching. But "cheaper" is not automatically "better," and a closed-weights product from a company pivoting away from open source deserves the same scrutiny you would give any other vendor before it touches your codebase.
