Meta's Muse Spark 1.1 Is the Cheap, Fast Coding Model Nobody Saw Coming
Quick verdict
Muse Spark 1.1 is Meta's answer to the "good enough, fast, cheap" tier that now runs most real coding and product work. It is not the smartest model on the board. It is priced at $1.25/$4.25 per million input/output tokens, runs a 1M-token context at around 114 tokens per second, and does strong UI and frontend generation. Practitioners called it the most surprising release of the week, and the reason is economics, not a benchmark crown. If you route work by task instead of by brand, this is another cheap lane worth having.
What actually shipped
Meta released Muse Spark 1.1 as an incremental follow-up to 1.0, but the jump was larger than the version number suggests. The details that matter:
- Pricing is $1.25 per million input tokens and $4.25 per million output tokens
- Context window is 1M tokens, with median output speed around 114 tok/s
- Artificial Analysis scored it 51 on its Intelligence Index, up 8 points from 1.0
- Reported token efficiency was strong, which lowers the real cost per task below the headline price
That 51 puts it roughly level with GLM-5.2, GPT-5.4, and GPT-5.6 Luna, and behind Grok 4.5, GPT-5.6 Sol, and Claude Fable 5. So it is a real step up, but not frontier leadership. The clearer signal came from the coding arena.
| Metric | Muse Spark 1.1 |
|---|---|
| Intelligence Index (Artificial Analysis) | 51 (+8 vs 1.0) |
| Context window | 1M tokens |
| Median speed | ~114 tok/s |
| Input / output price (per 1M) | $1.25 / $4.25 |
| Code Arena: Frontend | #9 |
Where it lands on the leaderboard
Arena placed Muse Spark 1.1 at #9 on Code Arena: Frontend, with its biggest gains in instruction-following and longer-query categories. That matches what people reported by hand: fast responses, clean UI generation, and quality that holds up across a large slice of everyday product tasks. Alexandr Wang framed it as near-frontier quality for a wide set of coding work, and Rowan Cheung and others echoed the "most surprising release of the week" read.
The interesting part is not the rank. It is that a model sitting at #9 on a frontend leaderboard, at roughly a quarter of the input price of the top tier, is often the right tool anyway. Most frontend and product-scaffolding work does not need a frontier reasoner. It needs speed, a big context, and a low enough price that you stop rationing calls.
Why it matters
Two things stand out. First, Meta's compute-heavy bet is finally showing up as a shipped, cost-effective product rather than a talent-acquisition headline. That raises real pressure on OpenAI and Anthropic on the exact axis where they are most exposed: the cheap, high-volume tier. Several commentators pushed for OpenRouter availability within hours, which tells you where the demand is.
Second, it deepens the routing problem that already defines this market. With Muse Spark 1.1, GLM-5.2, GPT-5.6 Luna, Grok 4.5, and Claude Fable 5 all sitting on different points of the cost curve, no single model is the right default for every request. Sending every prompt to one flagship is how you overpay. That is the case for cost-based model routing and for routing your coding agent instead of hard-coding one model. For the broader squeeze this puts on margins, see our read on where AI pricing is heading.
The catch is the usual one. Benchmark scores and launch demos run under ideal conditions. A 51 on an intelligence index does not tell you how it handles your codebase, your prompts, or your edge cases. Test it on your own tasks before you route real volume to it.
Video: first look at Muse Spark 1.1
An early hands-on take on what Meta shipped and whether the "Meta is back" framing holds up, which is the exact question the benchmarks leave open.
FAQ
How much does Muse Spark 1.1 cost?
It is priced at $1.25 per million input tokens and $4.25 per million output tokens. That undercuts most frontier models, and the reported token efficiency pushes the real per-task cost lower still, which is why the price drove most of the reaction rather than the raw score.
Is Muse Spark 1.1 as good as Claude or GPT-5.6?
Not on overall intelligence. Artificial Analysis scored it 51, behind Grok 4.5, GPT-5.6 Sol, and Claude Fable 5, and roughly tied with GPT-5.6 Luna and GLM-5.2. It is a value play, not an outright win. If you want the full field, our guide to the best AI models for coding covers where each one fits.
What is Muse Spark 1.1 best at?
UI and frontend generation, fast responses, and long-context work. Arena placed it #9 on Code Arena: Frontend with strong gains in instruction-following. It is a good fit for high-volume product scaffolding where speed and price matter more than a few points of reasoning benchmark.
Where can I use it?
Meta released it through its own API, and demand for wider distribution, including OpenRouter, showed up almost immediately. If you want to run several models like this without paying for each subscription separately, see our take on the best app for multiple AI models.
Sources
- @ArtificialAnlys - Intelligence Index 51, 1M context, ~114 tok/s, $1.25/$4.25 pricing
- @arena - Muse Spark 1.1 at #9 on Code Arena: Frontend
- @alexandr_wang - near-frontier quality for a wide set of coding tasks
- @rowancheung - most surprising release of the week
- @kimmonismus - strong UI generation and aggressive pricing
- @scaling01 - asking for OpenRouter availability
- @alexandr_wang - Meta's compute bet showing up as cost-effective inference
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix