Meta Launches Muse Spark 1.1: A Closed-Weights Agentic Model with Aggressive Pricing

TL;DR
Meta Muse Spark 1.1 is Meta's first paid, closed-weights API model, with $1.25/M input tokens and 1M context. The pricing, benchmarks and HN debate.
Last updated: July 9, 2026
Meta announced Muse Spark 1.1 today alongside the public preview of the Meta Model API - marking the company's first paid, closed-weights model offering. The release signals a strategic shift for a company that built its AI reputation on open-weights releases like Llama.
What Muse Spark 1.1 Actually Is#
Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs designed specifically for agentic tasks. The headline specs:
- 1 million token context window (no premium for long-context)
- Multimodal input (text, images, video)
- Structured output and parallel tool calling
- Search with citations built in
The model focuses on what Meta calls "agentic performance" - the ability to use tools, coordinate multi-step workflows, and operate autonomously. According to Meta's blog post, Muse Spark 1.1 "zero-shot generalizes to new native tools, MCP servers, and custom skills."
The Pricing That Got HN's Attention#
The Meta Model API pricing is notably aggressive:
| Token Type | Cost per 1M |
|---|---|
| Input | $1.25 |
| Output | $4.25 |
| Cached input | $0.15 |
| Web search | $2.50/1K queries |
For context, that $0.15 cached input price is lower than most competitors' standard input rates. The rate limits are generous too: free tier gets 60 requests/minute with 2M tokens/minute, paid tier gets 3,000 requests/minute with 4M tokens/minute.
As one HN commenter noted: "Very strong pricing, cheaper than Grok 4.5, particularly the cached reads."
Where It Ranks on Benchmarks#
Muse Spark 1.1's benchmark story is nuanced. It excels at tool use but trails top models on pure coding and reasoning:
Strong performance:
- MCP Atlas (scaled tool use): 88.1 - ahead of Opus 4.8 and GPT-5.5
- JobBench (professional tool use): 54.7 vs Opus 4.8's 48.4 and GPT-5.5's 38.3
Competitive but trailing:
- Terminal-Bench 2.1 (coding): 80.0 vs GPT-5.5's 83.4 and Opus 4.8's 82.7
- OSWorld-Verified: 80.8 vs Opus 4.8's 83.4
One commenter raised a valid concern about benchmark selection: "A lot of these benchmarks are unfamiliar. Are labs just choosing the ones that make them look best?"
What HN Is Saying#
The HN discussion (113 comments at time of writing) centers on a few themes:
The closed-weights elephant in the room. Multiple commenters expressed disappointment that Meta, known for open-weights Llama releases, is launching a closed API model. "This is not open-weights, right?" asked one. Another noted: "I missed the fact that Meta was developing and releasing closed-weights models... bummer."
Trust issues persist. Some commenters remain skeptical after previous benchmark controversies: "My trust factor is gone with Meta right now. Has there been any independent analysis to confirm they didn't cheat on benchmarks again?"
But competition is competition. The prevailing sentiment acknowledges that more options benefit developers: "Competition for cheaper and efficient models is a good thing, regardless of if you don't like SpaceX, Meta, etc. Especially from US based labs."
One commenter connected the release to Meta's recent acquisition: "Everyone has been loving to shit on the Alexander Wang acquisition but this seems legitimately impressive to me? Meta's AI org went from a total mismanaged dumpster fire for multiple years to delivering a competitive model in less than a year."
The Developer Take#
A few observations for developers evaluating this:
The tool-use focus is real. If you're building agentic systems that need to call many tools reliably, the JobBench and MCP Atlas scores suggest Muse Spark 1.1 might outperform more expensive alternatives. The parallel tool calling and structured output support reinforces this positioning.
Pricing makes experimentation cheap. At $1.25 input / $4.25 output, you can run extensive agentic workflows without budget anxiety. The $0.15 cached input is particularly attractive for systems with repetitive context.
It's not on OpenRouter yet. Several commenters noted they're waiting for OpenRouter availability before testing. If you want to try it now, you'll need to use the Meta Model API directly.
The coding story is secondary. For pure coding tasks, Opus 4.8 and GPT-5.5 still benchmark higher. Muse Spark 1.1 seems optimized for orchestration and tool use rather than raw code generation.
The Bigger Picture#
Meta releasing a closed-weights paid API is strategically interesting. The Llama series established Meta as the open-weights champion, giving developers free access to frontier-capable models. Muse Spark 1.1 represents a different bet: that some developers will pay for a managed API experience, especially for agentic workloads where reliability and tool integration matter more than model weights.
Whether this signals a shift in Meta's AI strategy or just a parallel product line remains to be seen. The HN consensus seems cautiously optimistic: more competition is good, even if it comes with Meta's baggage.
Continue Reading#
- Your AI Session Is No Longer Yours: How Providers Seal Reasoning, Search, and Subagent State
- Inkling-Small: Thinking Machines Ships a 12B-Active Open Model That Beats Its Big Sibling on Agent Work
- Beyond the Pelican Test: Opus 5 Renders the Lord of the Rings With a 1M-Token Budget
Sources#
Get the next comparison like this in your inbox
One email a week on News and the rest of the AI dev stack. Free.
Read next
Meta Muse Spark 1.1 Developer Guide: First Paid Meta API for Agentic Tasks
Meta Muse Spark 1.1 developer guide: Meta's first paid model API, with 1M-token context, OpenAI-compatible endpoints, pricing and setup for agentic tasks.
7 min readMeta Ships Muse Code and Muse Spark 1.2: A Terminal Agent With a 12x Cheaper Contributor Tier
Meta Muse Code is a terminal coding agent co-trained with Muse Spark 1.2. How the harness works and what the $0.10/$0.20 contributor tier asks in return.
7 min readMuse Ember: Meta's $0.10 Subagent Model vs Haiku 5.5 and Luna
Muse Ember lists at $0.10 in and $0.50 out per million tokens, the same sticker as Haiku 5.5 and GPT-6 Luna. Here is what actually separates them.
7 min readNew here? Start with
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.







