Meta Opens Muse Spark 1.1 to Coding Tools — and Undercuts on Price

Meta spent most of this year rebuilding its AI credibility around Muse Spark, the in-house model family it introduced in April after the earlier Llama-era coding tools failed to keep pace with the field. On July 9, 2026, that effort took its most consequential turn yet: Meta opened a public API — Muse Spark 1.1 — so outside developers can wire the model directly into their own products instead of only using it inside the Meta AI app.

What actually shipped

Two things are new here, and they matter for different reasons. First is the model itself: Muse Spark 1.1, which Meta describes as a step-change from the original Spark, with a 1-million-token context window, native multimodal input (images, video, PDFs), and built-in web search with citations. Second, and more interesting for builders, is the Meta Model API — the delivery mechanism. It's in public preview for US developers, with $20 in free credits per new account, according to reporting from AI Weekly.

The API supports both the OpenAI Chat Completions format and the Anthropic Messages format. That's a deliberate compatibility play: a coding assistant, IDE plugin, or agent framework already wired for either of the two dominant API shapes can point its base URL at Meta's endpoint without rewriting its integration layer. For a fast-moving vibecoding stack — where the model behind an agent is increasingly treated as a swappable backend — that lowers the switching cost to close to zero.

The coding pitch

Meta is positioning Muse Spark 1.1 specifically for agentic coding work: diagnosing and fixing bugs in large codebases, implementing features in "enterprise-grade" systems, and running large-scale code migrations. The model adds planning mode, subagent delegation for parallel execution, and context compaction — the same category of features that have become table stakes for agent frameworks handling long-running coding sessions rather than single-turn completions. Meta says the model is competitive with leading alternatives on its own internal coding benchmark; as with any vendor-run benchmark, that claim is worth validating against your own workload rather than taking at face value.

The number that will get attention: price

Meta is charging $1.25 per million input tokens and $4.25 per million output tokens — pricing multiple outlets, including TechTimes, put at roughly a quarter of comparable frontier-model rates from OpenAI and Anthropic. Meta AI chief Alexandr Wang called it the company's "strongest model for agentic and coding work yet." Early partner commentary cited by Meta's own announcement — including Replit and Cline, two names builders will recognize from the AI-coding-tool space — praised the model's tool use and agentic coding behavior, though those are vendor-solicited endorsements, not independent evaluations.

Why it matters for builders

None of this makes Muse Spark 1.1 automatically the right model for a given coding agent — that's an empirical question you answer by running it against your own tasks, not by reading a launch post. What it does change is the calculus around model choice. A large context window, dual-format API compatibility, and aggressive pricing together mean a coding tool can now add Meta as a real fallback or cost-optimization option with a config change, not a rewrite. For teams running multi-model agent setups — routing cheap, high-volume tasks to one model and reserving a pricier one for hard reasoning — that's a meaningfully lower bar to test the tradeoff. The preview is US-only for now, and Meta is serving it from its own infrastructure rather than listing it on third-party model marketplaces like OpenRouter, so availability and long-term pricing stability are still open questions.