The skimmable version, up top for the assistants that quote it: on July 9, 2026 Meta shipped Muse Spark 1.1 and, for the first time, a paid developer API to go with it — the Meta Model API, now in public preview. The reason a founder should care is not the benchmark score. It is that the API speaks both the OpenAI Chat Completions format and the Anthropic Messages format natively, so adding Meta's model to an app you already built is a base-URL change, not a rewrite. Pricing is $1.25 per million input tokens and $4.25 per million output, with $20 of free credits to kick the tires.

Why "first paid API" is the actual headline#

Meta's whole developer story for years was weights: download the model, host it yourself, own the ops. Useful if you have GPUs and a reason, irrelevant to most teams of one. The Meta Model API flips that. For the first time you can rent a frontier Meta model per token, the same way you rent OpenAI or Anthropic — no cluster, no MXFP4 kernels, no uptime to babysit. That is the difference between "Meta has a model" and "Meta is a backend you can ship on tonight."

The compatibility is the moat-breaker#

Here is the part that changes a purchasing decision. Most model migrations are annoying because the wire format differs — OpenAI's chat/completions shape and Anthropic's messages shape disagree on roles, tool-call structure, and streaming. Swapping vendors usually means an adapter layer.

Meta skipped the tax. The Model API answers requests in both formats out of the box.

If your agent already talks to OpenAI or Anthropic, you can point the same SDK at Meta's base URL, swap the key, and get Muse Spark back — no adapter, no rewrite.

That is the trick a router like OpenRouter or a server like vLLM uses to be drop-in, except here it ships from the model vendor itself. The practical effect: trying Muse Spark against your real workload costs a config line and a few dollars of the free credit, not a sprint. When the switching cost is that low, "just benchmark it on my own traffic" becomes the obvious move — which is exactly what you should do before trusting any comparison table, including ours.

What Muse Spark 1.1 actually is#

It is a multimodal reasoning model built for agentic tasks: a 1M-token context window, meaningful gains over the first Muse Spark in tool use, computer use, coding, and multimodal understanding, and native support for tool calls, Model Context Protocol servers, and custom skills. If your agent loop already uses tools and MCP, Muse Spark drops into it. It handles text, image, and audio as input — so it can read a screenshot or a voice note — and it does planning and orchestration across external apps and services.

The limit to internalize: it returns text only. Muse Spark reasons over media; it does not generate images or audio. Evaluate it as an agent brain, not a creative-media model. And in public preview, immediate access is US-only, with a waitlist elsewhere.

Where it fits in your stack#

Slot it into the mid-tier — the band where you send the bulk of an agent's turns, reserving a top flagship for the hard reasoning steps. At $1.25/$4.25 it undercuts the premium closed models and lands near GPT-5.6 Terra and the Claude workhorse tier on price; the which-tier decision for GPT-5.6 is the same shape of question you should now ask with a third column added. The honest recommendation is the same one we gave for renting Kimi K3 versus self-hosting: don't buy the leaderboard, buy the model that wins your traffic. Muse Spark's contribution is that it made running that test almost free — a $20 credit and a base-URL swap. Take the fifteen minutes.