Muse Spark 1.1 Review: The End of the Frontier Labs' Massive Lead?
Wait... Meta Can Make AI Models, and They're Good?
Meta has released its new Superintelligence team models, Muse Spark and Muse Spark 1.1, and they're actually really good. Shockingly good. I describe it as a "just do it" model. It's not Fable or GPT-5.6 Sol Ultra, but what this represents is something much bigger: the end of the frontier labs having such a massive lead.
What the benchmarks say
Muse Spark 1.1 benchmarks really well. It scores 53.3% on DeepSWE, 62.1% on Humanity's Last Exam, and 92.2% on SvelteBench, which is not bad at all. Overall, the model is fast, token-efficient, and just a solid model. Here are some of the other scores.
| Category | Benchmark | Muse Spark 1.1 | Muse Spark | Gemini 3.1 Pro (high) | Opus 4.8 (max) | GPT 5.5 (xhigh) |
|---|---|---|---|---|---|---|
| Agent | MCP Atlas Scaled tool use | 88.1 | 82.2 | 78.2 | 82.2 | 75.3 |
| JobBench Professional tool use | 54.7 | 17.0 | 15.9 | 48.4 | 38.3 | |
| Toolathlon-Verified Personal tool use | 75.6 | 49.4 | 61.1 | 76.2 | 73.5 | |
| OSWorld-Verified Agentic computer use | 80.8 | 53.3 | 76.2 | 83.4 | 78.7 | |
| Humanity's Last Exam Multidisciplinary reasoning (w/ tools) | 62.1 | 50.4 | 51.4 | 57.9 | 52.2 | |
| Finance Agent v2 Agentic financial analysis | 57.2 | - | 43.0 | 53.9 | 51.8 | |
| Coding | Terminal-Bench 2.1 Agentic terminal coding | 80.0 | 67.3 | 70.3 | 82.7 | 83.4 |
| SWE-Bench Pro Diverse software engineering | 61.5 | 55.0 | 54.2 | 69.2 | 58.6 | |
| DeepSWE 1.1 Long-horizon agentic coding | 53.3 | 10.0 | 12.0 | 59.0 | 67.0 | |
| Multimodal | CharXiv Reasoning Chart QA | 88.4 | 88.9 | 81.6 | 89.9 | 84.8 |
| BabyVision Visual reasoning | 76.3 | 39.9 | 51.5 | 81.2 | 83.6 |
As we can see, it is not state-of-the-art, but it is not bad by any means.
My usage
From my usage, this is an incredible model. I used it to build my shitty HTML plan distribution service because I use HTML plans a lot. The model one-shot a FastAPI project with a CLI and skills, so it worked on the backend, frontend, and skills flawlessly. It did an amazing job writing Python code. I'd honestly give it a 10/10, and that shows in its DeepSWE benchmark.
I've tested it not just against Muse Spark but also against Mimo 2.5 Pro and GPT-5.6, and this model uses the CLI flawlessly.
Then I had to add features to my church's internal fork of Lyrics Display. I added an API endpoint system so I could build a lightweight wrapper around it and control the app from my mobile phone if something went wrong while I was away from my computer.
I also added the ability to delete custom views so that if I had an external view I no longer needed, I could remove it.
I also used /goal to make the model run in a loop. I'll probably write a blog post about loops later, but I want to test them some more first. Theo has also been using loops, so if you want to know more about them, check out his videos.
The loops I was running is to make the model analyze the main repository to look for features I could use. I tried combining two approaches, but I didn't like the result, so I ended that experiment.
I also used it to remove some light bugs from the app before release. This was a React Electron app.
Then on another project, I built an Agent Communication Hub, which lets agents like Hermes, OpenClaw, OpenCode, Codex, and Claude Code communicate with each other using either scripts or MCPs, depending on what you want. The scripts also come with skills. That project is mostly finished now. I just have to fine-tune the code. It's already been useful. I used it while working on the internal Lyrics Display fork, the HTML plan project, and the communication hub. The first project was Python, while the second was SvelteKit, TypeScript, and Electron.
Then I used it on my latest project, which I don't have a name for yet. I'm trying to build an open-source ProPresenter killer that anyone can use for free, but more on that soon. I also used it on another project, a personal mobile app for reading and annotating documents. I didn't like how the annotation engine worked, so I had to redo it. I haven't fully tested it yet, but I had another model review it, and the code looked good. I also used another loop to analyze church sermons and generate titles for me. It did a really good job and was surprisingly accurate. That workflow used the YouTube MCP. If you're interested in the YouTube MCP, you can find it on my GitHub. If you want me to talk about my setup across Hermes Agent, OpenCode, and Pi, DM me. Maybe I'll make a Twitter thread, or I'll do a full write-up later this year.
Pricing
You're probably thinking this model is really expensive. It actually isn't. It's a very efficient model. Across all the projects I talked about, I only spent $28. Meta also gives you $20 in free credits, which is nice, so I really only spent around $ 8 of my own money. Pricing is:
- Input: $1.25 per million tokens
- Output: $4.25 per million tokens That's cheaper than Claude Sonnet 5 and Gemini 3.5 Flash.
Token Pricing Comparison (July 2026)
| Model | Input / 1M | Output / 1M | Cached / 1M | Context | AA Index |
|---|---|---|---|---|---|
| Muse Spark 1.1 | $1.25 | $4.25 | $0.15 | 1M | 51 |
| GPT-5.6 Sol | $5.00 | $30.00 | $0.50 | 1.05M | 59 |
| GPT-5.6 Terra | $2.50 | $15.00 | $0.25 | 1.05M | 55.0 |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.10 | 1.05M | 51 |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | $0.20 | 1M | 53 |
| Claude Sonnet 5 (Sep+) | $3.00 | $15.00 | $0.30 | 1M | 53 |
| Gemini 3.5 Flash | $1.50 | $9.00 | — | 1M | 50.2 |
| Grok 4.5 | $2.00 | $6.00 | $0.50 | 500K | 54 |
| Claude Haiku 4.5 | $1.00 | $5.00 | $0.10 | 200K | 30 |
How do other people see the model?
Most people support my take that the model is good and cheap, but there's one thing I want to talk about: Meta's new strategy. Who remembers the Llama series? A family of free, open-weight models that people loved. Now they've effectively killed that direction with both Muse Spark models. A random Hacker News commenter summed it up pretty well:
"Meta's open-source goodwill is a depreciating asset. Every month without open weights draws down the balance."
If you don't care whether the model is open-weight, then honestly, it's just a really good model.
The model is good. The pricing is disruptive. The API surface is clean enough for same-day third-party tooling. But the benchmarks still need independent verification, and the closed-weights pivot is a trust deficit Meta has not yet addressed.
- X: @JemoLife0213 https://x.com/JemoLife0213
- GitHub: Jemo69 https://github.com/Jemo69/
- BlueSky: @jemolife https://bsky.app/profile/jemolife.bsky.social
You can read more of my writing here: https://new-portfolio-ten-amber.vercel.app/blog Remember, your model matters, but you also need a good harness. I recommend Pi or OpenCode. Both are really good. And as always... Stay coding.