Model Drop: Muse Spark 1.1
A surprise release from Meta with some strong agentic performance claims
Meta’s original Muse Spark landed in April to a collective shrug (I skipped it too). Muse Spark 1.1 one is different for a reason that has nothing to do with benchmarks: it’s the first model Meta has ever asked anyone to pay for.
Model: Muse Spark 1.1, served through the new Meta Model API (public preview, OpenAI-compatible)
Model type: Natively multimodal reasoning model. Text, image, video, PDF, and audio input; text output
Ship date: July 9, 2026
Maker: Meta (Menlo Park), out of Meta Superintelligence Labs.
Pricing: $1.25 / $4.25 per million input / output tokens on the Meta Model API, with $20 in free credits at signup. Hacker News commenters flagged a cached-input rate around $0.15. Free in the Meta AI app
Available on: The Meta AI app and meta.ai in Thinking mode, and the Meta Model API in public preview for US developers. Meta says it will replace the Llama models behind WhatsApp, Instagram, Facebook, and its smart glasses. No open weights
Headline benchmarks: The wins are agentic. MCP Atlas 88.1, JobBench 54.7 (Opus 4.8: 48.4, GPT-5.5: 38.3), Humanity’s Last Exam with tools 62.1 (Opus 4.8: 57.9), Finance Agent v2 57.2. The losses are coding: SWE-Bench Pro 61.5 (Opus 4.8: 69.2), Terminal-Bench 2.1 80.0 (GPT-5.5: 83.4, Opus 4.8: 82.7), DeepSWE 1.1 53.3 (GPT-5.5: 67.0, Opus 4.8: 59.0). That DeepSWE number is still a leap from the original Muse Spark’s 10.0.
Other info: 1M-token context window with built-in context compression (the original shipped with roughly 260K per Artificial Analysis). Multi-agent orchestration is native: the model is built to run as a primary agent delegating to subagents, or as a subagent itself. Computer use decides on its own when to write a script versus click through a UI. Parameter count, architecture, and knowledge cutoff undisclosed. Safety evals ran under Meta’s Advanced AI Scaling Framework, which reports strong jailbreak resistance and “safe margins” on chem/bio, cyber, and loss-of-control categories.
More details: Introducing Muse Spark 1.1.
What shipped
Meta released Muse Spark 1.1 alongside the public preview of the Meta Model API, the first time developers at large can buy Meta inference directly. The pitch, straight from Zuckerberg’s Threads post, is “a strong agentic and coding model at a very low price.” It’s a multimodal reasoning model built for tool use, computer use, and multi-agent orchestration, running a 1M-token context window, and it’s already live in the Meta AI app’s Thinking mode with the WhatsApp, Instagram, and Facebook assistants next in line. The subtext: Llama is over. Meta trained a closed frontier model, put a meter on it, and joined the market it spent three years undercutting with free weights. It’s a significant pivot.
The evidence Meta stacked up is selectively strong. On agentic benchmarks the model beats the frontier in places: MCP Atlas 88.1, JobBench 54.7 against Opus 4.8’s 48.4 and GPT-5.5’s 38.3, Humanity’s Last Exam with tools 62.1 over Opus 4.8’s 57.9. On coding, the launch’s other headline, it trails everyone that matters: seven points under Opus 4.8 on SWE-Bench Pro, and a distant third on DeepSWE 1.1 behind GPT-5.5 and Opus 4.8. The blog post shows charts without publishing a system card’s worth of numbers, pricing appeared in press briefings rather than launch materials, and within hours a Hacker News commenter alleged Meta ran Terminal-Bench outside the benchmark’s resource limits. For a lab that reorganized after the Llama 4 benchmark-contamination mess, the eval hygiene questions started early.
What’s new
Muse Spark 1.1 is a real upgrade over April’s Muse Spark, but the most consequential changes are business decisions.
A price tag, for the first time ever. Meta has never charged for a model. The Meta Model API at $1.25 / $4.25 with $20 in starter credits ends the free-weights era and puts Meta in direct commercial competition with Anthropic, OpenAI, and Google. The Batch called the pivot “a significant loss for the developer community” that built on Llama.
Agentic tool use that beats the frontier. JobBench at 54.7 against Opus 4.8’s 48.4 and GPT-5.5’s 38.3 is the standout, and MCP Atlas 88.1 and Finance Agent v2 57.2 back it up. The original Muse Spark won nothing.
Coding it can show in public. DeepSWE 1.1 went from 10.0 to 53.3 in one release. Still third place, but April’s model couldn’t be mentioned in the same sentence as Opus.
1M context with compression. Roughly 4x the original’s window, with built-in compression the model manages itself. Combined with video and PDF input, that’s a lot of enterprise document sludge in one call.
Orchestration as a first-class feature. The model is explicitly built to be both the orchestrator and the subagent in multi-agent systems, and its computer use picks between scripting and clicking on its own. Meta is designing for agent swarms.
How and where to use it
Where it runs, what it’s for, and where you should keep your current model.
Where it’s available
Meta AI app and meta.ai in Thinking mode for free
Meta Model API in public preview for US developers, OpenAI-compatible with $20 in credits
It’s headed for WhatsApp, Instagram, Facebook, and Meta’s smart glasses as the Llama replacement
What it’s good at
Cheap, high-volume agentic work
Multi-step tool-use jobs (JobBench, MCP Atlas), financial-analysis agents (Finance Agent v2), tool-assisted research (HLE with tools), and long-context multimodal grinding through video, PDFs, and images
What it’s bad at / shouldn’t be used for
Serious coding, and especially long-horizon DeepSWE work
Anything that needed open weights: local inference, fine-tuning, air-gapped deployments, the entire Llama ecosystem use case
It’s US-only in preview, so non-US products wait
If your workload involves data you wouldn’t hand Meta, that’s not a technical limitation but it should be on this list anyway
First impressions
The positives
Techzine framed the launch as Meta finally reaching the pack it’s been chasing:
“…fits into the now-familiar lineup of LLMs that fall just short of Claude Fable 5 but are much more affordable.”
Meta isn’t claiming the frontier. It’s claiming the price-performance shelf where Grok 4.5, GLM-5.2, and discounted Sonnet live.
On Hacker News, commenter greenavocado gave the launch the most honest compliment it got all day:
“Meta is back in the game, albeit not at the top.”
After the Llama 4 contamination scandal, the lab reorg, and an April release nobody noticed, “back in the game” is a real status change. Six months ago the question was whether Meta Superintelligence Labs would ship anything at all.
OfficeChai’s benchmark writeup called the JobBench result the standout, and it’s worth sitting with the spread: 54.7 against Opus 4.8’s 48.4 and GPT-5.5’s 38.3.
The negatives
The sharpest critique on Hacker News wasn’t about capability, it was about eval integrity. Commenter GodelNumbering alleged Meta ran Terminal-Bench 2.1 with 6 CPU cores when the benchmark caps tasks at 4:
“This disqualifies the results. Each terminal bench task has a cpu upper limit and RAM upper limit.”
Meta hasn’t responded as of this writing.
The Batch has been tracking the open-weights abandonment since April and calls it a significant loss for the developer community. Startups raised money on the premise of free frontier weights. Internal tools got built on Llama. Every one of those bets now faces a widening gap between the last open Llama and a closed Muse Spark line, and Meta’s answer is a metered API.
And the trust problem showed up unprompted. From the same HN thread:
“I cannot think of a worse company to trust with additional personal data.”
You can dismiss that as reflexive Meta-bashing, but this model is about to sit inside WhatsApp and Instagram and is sold on reading your PDFs, your screen, and your tool outputs. The company asking for that access has the worst data-privacy track record of any frontier lab. Meta also published safety-framework results but nothing resembling the system cards Anthropic and OpenAI ship at launch.
Jake’s take
JobBench and MCP Atlas measure the boring multi-step tool work that makes up 80% of what my agents do, and Muse Spark 1.1 (supposedly) beats Opus 4.8 on both at a quarter of the price. The pricing shelf it landed on is exactly where Sonnet 5’s intro-rate land grab was pointed two weeks ago. The agent middle market suddenly has real competition, and I like everything about that.
What I don’t like is asking me to trust the scoreboard. I wrote last week about models gaming their evals, and here’s Meta launching its redemption model with charts instead of a system card, pricing delivered via press briefing, and an unanswered allegation that its Terminal-Bench run used more CPU than the benchmark allows. This is the same company that reorganized its entire AI division over Llama 4 benchmark contamination (and the same company that will run this model inside WhatsApp, reading whatever you feed it).
The capability story might be completely real. The JobBench spread is too big to be noise. But Meta has burned its benefit of the doubt twice now, once on evals and once on the open-weights community it orphaned mid-bet, and a cheap model doesn’t buy it back.