Muse Spark 1.1 is now open to outside developers through a paid API, the same season Meta is telling its own engineers to stay away from Anthropic's Claude and OpenAI's Codex.
Meta introduced Muse Spark 1.1 on Thursday and opened the model to developers through a new Meta Model API in public preview, according to Reuters. Bloomberg framed it more bluntly: Meta is starting to charge for AI for the first time, putting Muse Spark 1.1 directly against the business models Anthropic and OpenAI have run for years. Developers in the United States can sign up now, test prompts, and build with the model directly. New accounts get $20 in free credits before moving to pay-as-you-go pricing of $1.25 per million input tokens and $4.25 per million output tokens, according to Reuters reporting picked up by U.S. News and Yahoo Finance. That rate sits above OpenAI's entry-level GPT-5 mini and Anthropic's Claude Haiku 4.5, but below Anthropic's higher-end Claude Sonnet 4.6.
Muse Spark first arrived in April as Meta Superintelligence Labs' debut frontier-class model, built for multimodal understanding and parallel agent workflows. Version 1.1 pushes further into coding and agentic tasks specifically. It carries a one-million-token context window, large enough to hold an entire codebase in a single session, and is built to orchestrate multi-agent systems, functioning as either a primary agent or a subagent. Meta says the new version shows sharp gains over its predecessor on its own Internal Coding Bench, particularly in cybersecurity and agentic coding, and now competes with the leading models from Anthropic and OpenAI. If you're a developer who defaults to Claude Code or Codex for those tasks, Meta is betting you'll at least try the alternative once the free credits run out.
The timing is hard to ignore. Meta restricted its own engineers from freely using Claude Code and Codex back in May, a policy first detailed by the-decoder.com and later covered across outlets including Cryptobriefing and AI Weekly. The concern inside Applied AI Engineering wasn't security in the usual sense. It was distillation. If engineers used Claude or Codex to generate test cases, hunt for bugs, or shape benchmarks for MetaCode, Meta's internal coding assistant, that assistant would effectively be learning from a competitor's model instead of Meta's own data. The policy specifically bars placing AI-generated content from those tools into any environment accessible by the model being trained, while still permitting engineers to use them for scaffolding and organizing test frameworks, as long as a human reviews every line that comes out.
So Meta is running two policies at once: the same policy, really, pointed in opposite directions. Internally, it's walling off MetaCode's training from rival AI output. Externally, it's inviting developers to build on top of its own model instead. Frankly, that's the whole strategy laid bare: keep Anthropic and OpenAI's fingerprints out of what Meta trains, and put Meta's own model in front of the developers who'd otherwise be paying Anthropic and OpenAI.
It also fits a pattern Meta has run all year with the Muse name. Two days before the Spark 1.1 API opened, Meta Superintelligence Labs launched Muse Image, its first image generation model, across the Meta AI app, meta.ai, Instagram Stories in the U.S., and WhatsApp in limited countries. Muse Image doesn't map a prompt straight to pixels. It leans on Muse Spark as a reasoning layer underneath, planning composition and pulling in real-time web context before anything renders. By Meta's own comparisons, Muse Image already beats Google's Nano Banana 2 on several editing benchmarks, while trailing OpenAI's latest GPT Image model on overall quality.
Muse Spark, in other words, isn't one product. It's the layer sitting under most of what Meta Superintelligence Labs ships, from Instagram's photo editor to a coding API developers can now pay to use directly. Whether outside developers actually switch from Anthropic or OpenAI will come down to the ordinary things: latency, reliability, and whether those coding benchmarks hold up once they're not being graded on Meta's own test. But the bigger story here is control. Meta spent the first half of 2026 deciding what its engineers could feed into its own models. It's spending the second half deciding, for the first time, to sell what comes out the other end.
Also read: China finally lets its AI giants buy Nvidia's H200 chip on its own terms • Mercor's Revenue Hit $2 Billion in June and the AI Data Race Isn't Slowing • Ollama Raises $65 Million as It Grows to Nearly 9 Million Developers
