Muse Spark 1.1 Lands with a $1.25 API and Day-One CLI

Mark Zuckerberg dusted off his @finkd handle on July 9, 2026 — his first post on X in three years — to announce Muse Spark 1.1, Meta's first paid developer model, served through the brand-new Meta Model API. The timing was deliberate: SpaceXAI had shipped Grok 4.5 the day before, and GPT-5.6 Sol was rumored for Thursday. But where those launches led with benchmark tables, Zuckerberg led with a price tag: $1.25 per million input tokens, $4.25 per million output tokens. That is roughly 60% cheaper than Claude Sonnet 5 and within spitting distance of Haiku-tier pricing — from a company with zero margin pressure on its AI division.

View original post on X →

The model-war framing has been covered extensively. This article is about something more useful: what Muse Spark 1.1 actually ships, how to hit the API today, what Simon Willison's same-day CLI plugin tells us about the API surface, and whether the benchmarks hold up under scrutiny.

What Muse Spark 1.1 Actually Is

Muse Spark 1.1 is a natively multimodal reasoning model from Meta Superintelligence Labs — the rebranded research division that replaced FAIR. It accepts text, images, video, PDFs, and audio as input and produces text output. The key specs:

• 1 million token context window with active context management — the model compresses and retrieves from its own context mid-generation
• Native multimodal perception — not a vision encoder bolted onto a text model, but a unified architecture that reasons across modalities
• Agentic capabilities — parallel tool calling, structured output, built-in search with citations, multi-agent orchestration, and computer use across desktop, mobile, and browser
• OpenAI-compatible API — drop-in replacement for existing OpenAI SDK integrations

Zuckerberg detailed the specs in his thread:

View original post on X →

The practical upshot: this is a frontier-adjacent model that does multimodal reasoning, agentic tool use, and coding — and Meta is selling access at commodity prices.

The Meta Model API: What Developers Get

This is Meta's first serious developer API. Not a research preview, not a waitlist — a public preview with pricing, SDKs, and documentation. Here is what matters:

Access:

• Public preview for US-based developers (international rollout TBD)
• $20 in free credits per new account
• OpenAI-compatible endpoint — swap the base URL and API key, keep your existing code

Pricing (per million tokens):

Model	Input	Output	Cached Input
Muse Spark 1.1	$1.25	$4.25	$0.15
Claude Sonnet 5	$3.00	$15.00	$0.30
GPT-5.5	$2.00	$8.00	—
Claude Haiku 4.5	$0.80	$4.00	$0.08

The pricing slots Muse Spark between Haiku (the budget tier) and Sonnet/GPT mid-tier — but with frontier-class ambitions. Cached input at $0.15/M tokens is particularly aggressive for agentic workloads where the system prompt and tool definitions repeat across calls.

# Meta Model API — OpenAI-compatible
from openai import OpenAI

client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key="your-meta-api-key"
)

response = client.chat.completions.create(
    model="muse-spark-1.1",
    messages=[
        {"role": "user", "content": "Explain the CAP theorem in three sentences."}
    ],
    max_tokens=256
)
print(response.choices[0].message.content)

If you have ever integrated the OpenAI SDK, you already know how to use the Meta Model API. That is the point.

View original post on X →

Simon Willison's Same-Day Plugin: The Real Signal

Within hours of the Meta Model API going live, Simon Willison — the developer behind Datasette and the LLM CLI framework — shipped llm-meta-ai, a plugin that gives you Muse Spark 1.1 access from your terminal.

View original post on simonwillison.net →

# Install and configure in under 60 seconds
uv tool install llm
llm install llm-meta-ai
llm keys set meta-ai
# Paste your Meta Model API key

# Use it
llm -m meta-ai/muse-spark-1.1 "Generate an SVG of a pelican riding a bicycle"

This matters more than any benchmark table. When a respected independent developer can read the API docs, write a working plugin, and ship it the same afternoon the API launches, that tells you three things:

1. The API surface is clean. OpenAI-compatible means the entire ecosystem of LLM tooling — LangChain, LiteLLM, Willison's LLM framework — can integrate with minimal effort.
2. The documentation is adequate. Developers do not ship same-day integrations against poorly documented APIs.
3. The auth flow is not hostile. No six-step OAuth dance, no enterprise sales call. Get a key, set it, go.

The Benchmarks: Strong, but Read the Fine Print

Meta's evaluation report paints a nuanced picture. Muse Spark 1.1 excels at agentic tasks while trailing on raw coding benchmarks:

Where Muse Spark 1.1 leads:

• JobBench: 54.7 (vs. Opus 4.8: 48.4, GPT-5.5: 38.3) — multi-step agentic task completion
• MCP Atlas: 88.1 — tool orchestration and structured output
• Humanity's Last Exam (with tools): 62.1 (vs. Opus 4.8: 57.9) — complex reasoning with tool access

Where it trails:

• SWE-Bench Pro: 61.5 (vs. Opus 4.8: 69.2) — real-world software engineering
• DeepSWE 1.1: 53.3 (vs. GPT-5.5: 67.0) — deep code understanding
• Terminal-Bench 2.0: 59.0 (vs. GPT-5.4: 75.1, Gemini 3.1 Pro: 68.5) — terminal and shell tasks

View original post on Handy AI →

The honest read: Muse Spark 1.1 is genuinely strong for agentic orchestration and multimodal understanding. It is not the best coding model — that crown stays with Opus 4.8 and Codex. But at this price point, "good enough at coding plus best-in-class at agent orchestration" is a compelling package for builders running multi-agent systems.

Zuckerberg's Margin Compression Play

"The pricing from some of the other AI labs is very extreme and the margins are very high," Zuckerberg told Bloomberg on launch day. That is the thesis statement for the entire release.

Meta does not need AI API revenue to survive. Its advertising business prints money. The Meta Model API is a strategic weapon, not a profit center. The playbook:

1. Price below cost for pure-play AI labs — at $1.25/$4.25, Meta can afford to operate at break-even or a loss on API revenue indefinitely
2. Force competitors to match or lose developer share — OpenAI and Anthropic cannot subsidize API pricing with ad revenue
3. Commoditize the model layer — if frontier-quality models are available at commodity prices, the value shifts to distribution (Meta's apps: WhatsApp, Instagram, Facebook, Ray-Ban Meta glasses) and the platform ecosystem

One Hacker News commenter put it bluntly: this is a "spoiler strategy" — commoditize coding models via aggressive pricing to deflate competitor revenue. Meta can afford to run the API as a loss leader because it monetizes AI through its consumer products, not through developer API margins.

What the Community Is Saying

The Hacker News thread on Muse Spark 1.1 hit 333 points and 174 comments within hours — high engagement but heavily skeptical.

View on Hacker News →

The debate breaks into three camps:

The enthusiasts point to the pricing as transformative. Cached input at $0.15/M tokens makes agentic workloads — where system prompts and tool definitions repeat across hundreds of calls — dramatically cheaper. For teams running multi-agent pipelines, switching from Sonnet at $3/$15 to Muse Spark at $1.25/$4.25 could cut API costs by 60-70%.

The skeptics focus on two issues. First, the benchmark controversy: did Meta game Terminal-Bench by exceeding resource limits? If so, the agentic performance claims need independent verification. Second, the closed-weights pivot: Meta built its AI developer community on Llama's open weights, and Muse Spark 1.1 ships with zero download option. The community that made Llama a standard is being asked to trust a proprietary API.

The pragmatists note that Meta said it has "a variant of Muse Spark that is in development that we do intend to open source." But no timeline was given, and "intend to" is not "will." As one commenter put it: "Meta's open-source goodwill is a depreciating asset. Every month without open weights draws down the balance."

Elon Musk replied to Zuckerberg's announcement with a single word: "jinx" — SpaceXAI had shipped Grok 4.5 the day before with a similar "cheaper than the competition" pitch. The billionaire price war is real, and developers are the beneficiaries.

What This Means for You

If you are building agentic workflows: Muse Spark 1.1 is the strongest contender for multi-agent orchestration at this price tier. The 1M token context with active management, parallel tool calling, and native multimodal perception make it purpose-built for agent pipelines. Start with the $20 free credits and benchmark against your actual workloads — do not rely on Meta's published numbers.

If you are an API-first developer: The OpenAI-compatible endpoint means zero switching cost. Willison's LLM plugin lets you test from the CLI in under a minute. If you already use OpenRouter or LiteLLM, expect Muse Spark 1.1 integration within days.

If you depend on open weights: Do not switch from Llama yet. Muse Spark 1.1 is API-only, US-only in preview, and Meta has not committed to an open-weights release timeline. Keep running your Llama infrastructure and evaluate the API as a supplement, not a replacement.

The bottom line: The model is good. The pricing is disruptive. The API surface is clean enough for same-day third-party tooling. But the benchmarks need independent verification, and the closed-weights pivot is a trust deficit Meta has not yet addressed. Use it, benchmark it, but do not bet your stack on it until the numbers are confirmed by someone other than Meta.

---

Meta Model API access: ai.meta.com | Simon Willison's plugin: llm-meta-ai | HN discussion: 333 points, 174 comments