Independent Numbers Are In: Meta's Muse Spark 1.1 Is a Serious Value Pick
Benchmark firm Artificial Analysis puts Meta's Muse Spark 1.1 ahead of GLM 5.2 in coding at a lower price — and its hallucination rate dropped by nearly half.
When Meta launched Muse Spark 1.1 with aggressive API pricing this week, the natural question was: is the model actually any good, or just cheap? Now the independent benchmark firm Artificial Analysis has published its numbers — and they suggest the answer is: genuinely good for the price.
On the Intelligence Index, a combined score across many tasks, Muse Spark 1.1 lands at 51 — tied with Zhipu’s GLM 5.2, OpenAI’s GPT-5.4, and GPT-5.6 Luna. That’s a jump of eight points in just three months, driven mostly by coding and agent-style knowledge work. On the Coding Index specifically, it scores 71.3, edging past GLM 5.2 (68.8) and sitting just behind GPT-5.6 Luna (71.4). The top of the table remains out of reach: GPT-5.6 Sol (77.4), Terra (76.7), and Claude Fable 5 (76.5).
The economics are where it gets interesting. Artificial Analysis estimates Muse Spark 1.1 costs about $0.26 per task, versus $0.37 for GLM-5.2 and $0.89 for GPT-5.4 — partly because it’s less chatty, using 94 million output tokens across the test suite where GLM-5.2 needed 141 million. Two more notable improvements: the hallucination rate (how often the model confidently makes things up instead of admitting it doesn’t know) dropped from 73 to 38 percent, mostly because the model now declines to answer more often. And the context window — roughly, how much text the model can keep in mind at once — quadrupled to one million tokens.
What’s behind this? A price war, in plain terms. Chinese open-source models like GLM 5.2 have been undercutting the big US labs all year, and Meta is now competing on exactly that turf: not the crown, but the best intelligence per dollar. For Meta, whose AI reboot has drawn skepticism, an independent firm confirming real three-month progress matters almost as much as the scores themselves. The usual caveat applies, and Artificial Analysis says so itself: benchmark scores don’t always match real-world performance. And for now, the model is only available through Meta’s own API — no open weights, no running it yourself.
What this means for you: If you just chat with an AI assistant now and then, nothing changes. But if you build with AI or pay per API call, the mid-range is suddenly crowded with good, cheap options — and switching costs between them keep falling. The smart move is running your own real tasks against two or three of these models rather than trusting any single leaderboard: at $0.26 per task, testing is cheaper than guessing.
Sources
Source: https://x.com/ArtificialAnlys/status/2075677416295739660
OpenAI Admits Its Big Work Launch Stumbled — Here's What's Being Fixed
After a wave of criticism over ChatGPT Work and GPT-5.6 Sol — burned-through usage limits, a confusing redesign, and reports of deleted data — OpenAI promises fixes within a week.
