Meta's Muse Spark 1.1 is live, and speed is the real headline

Muse Spark 1.1 gives Meta a new live model to point developers at, not just another lab benchmark. Meta introduced the model on July 9 as a multimodal reasoning model for agentic tasks, coding, computer use, and multimodal understanding. 1

Muse Spark 1.1 is now available in two places: Meta says it is live in "Thinking" mode inside the Meta AI app and on meta.ai, and developers can access it through the new Meta Model API public preview. 1 The developer preview is self-serve, OpenAI SDK-compatible, and currently aimed at US developers. New API accounts get $20 in free credits. 2

The published API price is aggressive: $1.25 per 1M input tokens and $4.25 per 1M output tokens. 2 That puts Meta's first-party API in the same cost conversation as the cheaper frontier-model tiers, rather than only the flagship expensive ones.

Meta is positioning Muse Spark 1.1 around long-running agent work. The model has a 1M-token context window, can handle images, video, and documents in a single call, and is tuned for repository-level coding, tool calling, computer-use workflows, and multi-agent orchestration. 2

The strongest outside signal so far comes from Vals AI. Vals said Muse Spark 1.1 ranked #4 on its Vals Index and was the fastest model in the top 10, running about 3x faster than the top three models. 3 Vals' model page lists Muse Spark 1.1 at 68.41% accuracy, 388.52 seconds of latency, and $0.50 per test on the Vals Index. 4

That comparison is close enough to matter. GPT-5.5 is listed by Vals at 67.95% accuracy, 681.75 seconds of latency, and $4.60 per test on the same index. 5 Meta chief AI officer Alexandr Wang framed the same point more directly, saying Muse Spark 1.1 is ahead of GPT-5.5 and Grok 4.5 on the Vals Index. 6

This is still an early access story, not a universal rollout. The API is in public preview, the developer guide describes availability for US developers, and Meta warns that Muse Spark is a reasoning model whose reasoning tokens are billed as output. 2 Vals also labels the model proprietary and access-gated. 4

The useful read: Meta now has a paid model API with a fast agentic model behind it. The open question is whether independent developers see the same speed and cost profile once real coding agents start burning retries, tool calls, and long-context reasoning in production.