What changed
Meta launched Muse Spark 1.1 and opened the Meta Model API in public preview on 2026-07-09 14. It takes up to 1M tokens of multimodal input (text, images, and other media) and returns text only 13, and it is built for agentic work: Meta says its tool use zero-shot generalizes to native tools, MCP servers, and custom skills, that it runs computer-use workflows across multiple applications, orchestrates multi-agent systems, and codes on large, complex codebases including bug fixing and large migrations 1. The API is OpenAI-compatible; US developers get immediate access, with a waitlist for everyone else 13. This is public preview, not GA.
The price, in context
At $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits 23, Muse Spark 1.1 lands next to Claude Haiku 4.5 and GPT-5.6 Luna; TechCrunch puts it “in line with (albeit slightly above)” that pair 2. Read the number, not the launch copy: this is the fourth cheap-tier agentic coder to sit at the same price floor. What makes it a routing decision rather than a project is the surface. The API is OpenAI-compatible with native MCP tool use 13, so adding it to a cost-aware router is a base-url swap and one routing-table row, not a client migration.
The strategy flip
Here is the part the spec sheet buries. Meta made open-weight Llama the free default and used it to commoditize the base-model layer; now it is metering tokens and, per Fortune, “competing on Anthropic and OpenAI’s turf” 4. Zuckerberg frames the focus as “strong agentic and multimodal models at very low cost” 4. This is Meta’s first paid, closed model, which is a bigger signal than one more row on the price list: the free-weights-as-strategy era at Meta is the thing that changed.
Impact on your team
If you run a cost-aware model router, this is a candidate cheap tier you can trial today: US-only, on $20 in free credits, reachable through your existing MCP tools with a base-url swap 13. The concrete move is to benchmark it on your own agentic, bug-fix, and migration workload before trusting Meta’s framing, because Meta dodged the flagship comparison 4. What to wait on: do not treat the third-party SWE-Bench or Terminal-Bench numbers as official, and do not build on GA assumptions, since it is preview, US-only, and text-out. The strategic read matters more than the spec: the era of free open weights from Meta is what ended here, not just the arrival of one more model on the price list.