Meta's new AI model "Muse Spark 1.1" has reached the same overall score as GLM-5.2 in an independent evaluation environment. According to results published by Artificial Analysis on July 10, 2026, both models scored 51 on the Intelligence Index. Muse Spark 1.1 recorded 71.3 in the coding category, surpassing GLM-5.2's 68.8. Its estimated cost per evaluation task also came in at $0.26, roughly 30% lower than GLM-5.2's $0.37. However, this gap should not simply be read as "Muse winning across the board." Meta's model is a closed API currently in public preview in the United States, while GLM-5.2's weights are available for organizations to run on their own infrastructure. Developers will need to weigh these scores alongside how much control they want over the execution environment.
Tied at 51 overall, but a 2.5-point gap in coding
Artificial Analysis's Intelligence Index v4.1 combines nine types of evaluations into a single index. The composition is 34% agentic tasks, 24% each for coding and scientific reasoning, and 18% general capability. Rather than limiting itself to competitive code-generation benchmarks, it also measures real on-device tasks, tool use in banking operations, long-context comprehension, and factual reliability. The organization estimates the index's 95% confidence interval at under 1 point.
Under this framework, Muse Spark 1.1's overall score rose from 43 for the first generation to 51. This ties it with GLM-5.2's max setting, GPT-5.4's xhigh setting, and GPT-5.6 Luna's max setting. The largest gain came in coding-related metrics: the Coding Index jumped 12 points, from 59 to 71. SciCode, which measures scientific code generation, also rose 6 points, from 52% to 58%, placing Muse Spark 1.1 third among the models Artificial Analysis has evaluated.
The difference with GLM-5.2 shows up not in the overall score but in the breakdown. On the Coding Index, Muse Spark 1.1 scored 71.3 versus GLM-5.2's 68.8—a gap of 2.5 points. Meanwhile, both models tie at 51 overall. Muse leading in the coding evaluation is not the same as it outperforming across every capability.
Meta itself, in its July 9 announcement, said 1.1's focus was on coding and long-running agents. The model manages a 1-million-token context, retrieving past actions and compressing them to retain only necessary steps. It was also trained so that a lead agent formulates plans and delegates work to sub-agents in parallel. Behind the 8-point gain over three months lies not just improved model knowledge, but strengthened ability to keep working while using tools.
The 94 million tokens behind the $0.26 figure
Looking at the price sheet alone doesn't explain the cost difference. Artificial Analysis tallies the input and cache costs used in each evaluation, adding in costs from reasoning and response generation. It then computes a weighted average based on task count and the index's composition ratios. Muse Spark 1.1 used 94 million output tokens across the evaluation suite, while GLM-5.2 used 141 million. Muse kept its output roughly one-third lower.
On top of that comes Meta's lower unit pricing. Meta Model API pricing is $1.25 per million input tokens and $4.25 per million output tokens. Since tokens generated during reasoning are billed as output, models that "think" longer incur higher costs. The per-task cost Artificial Analysis calculated came to $0.26 for Muse versus $0.37 for GLM. The $0.11 difference amounts to about 29.7%.
| Artificial Analysis Measurement | Muse Spark 1.1 xhigh | GLM-5.2 max |
|---|---|---|
| Intelligence Index | 51 | 51 |
| Coding Index | 71.3 | 68.8 |
| Output tokens across full evaluation | 94 million | 141 million |
| Estimated task cost | $0.26 | $0.37 |
| Output speed | 116.3 tokens/sec | 191.4 tokens/sec |
The lower cost has another side to it. Muse's output speed is 116.3 tokens per second, falling short of GLM-5.2's 191.4 tokens per second. If you want to push a large volume of processing through in a short time, the choice depends on whether you prioritize per-task cheapness or throughput. Muse kept output 33.3% lower than GLM-5.2, but that doesn't mean 94 million tokens is the lowest figure across the entire market.
How far the claim "better than GLM-5.2" actually extends
Benchmark rankings hold true when evaluation items and the execution environment are fixed. Artificial Analysis's 71.3 versus 68.8 is the result of setting each model to a specified reasoning intensity on their Coding Index. If a user's codebase, programming language, or tool design differs, success rates—and rankings—could shift.
Meta's 112-page evaluation report also reveals another limitation. The company's own overall capability table compares Muse Spark 1.1 against GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro—it does not include GLM-5.2. On coding evaluations, Muse recorded 80.0 on Terminal-Bench 2.1, 61.5 on SWE-Bench Pro, and 53.3 on DeepSWE 1.1. But on Terminal-Bench, GPT-5.5 scored 83.4 and Claude Opus 4.8 scored 82.7; on DeepSWE, GPT-5.5 scored 67.0. Muse did not take the top spot across every coding task.
The evaluation report also explicitly notes that self-reported figures from competing models are sometimes used, and that a common harness may not be optimized for the strengths of each company's model. Meta's published figures and Artificial Analysis's independent evaluation differ in both the models covered and the methodology. The two cannot simply be combined into a new unified ranking.
The same caution applies to factual reliability. On Artificial Analysis's AA-Omniscience, Muse Spark 1.1's hallucination rate dropped from 73% for the first generation to 38%. Meanwhile, the rate at which it attempted to answer fell from 95% to 82%, and the accuracy rate among attempted questions stayed roughly flat, from 45% to 41%. The main driver of the score improvement is the model's judgment to abstain when uncertain. While the reduction in wrong answers is a significant achievement, it doesn't necessarily mean the model's underlying knowledge accuracy has improved.
Closed API and open weights carry different pricing implications
Muse Spark 1.1 and GLM-5.2 stand in contrast in terms of how they are offered. Muse is a proprietary model accessed through Meta's API, with public preview currently limited to developers in the United States. It supports OpenAI-compatible and Anthropic-compatible formats, making it easy to connect to existing agents. But where it runs, how pricing changes, and which regions get access are all decided by Meta.
GLM-5.2 is a 753-billion-parameter model whose weights Z.ai has released under an MIT license. It supports a 1-million-token context and uses IndexShare, which shares an indexer across every four layers to reduce the computational load of long-context processing. Users can choose their API provider, or, given sufficient infrastructure, run it themselves. Artificial Analysis's $0.37 figure is based on the API pricing it tracks, not a total cost of ownership that includes one's own GPUs, power, and operational staff.
So while $0.26 versus $0.37 provides a useful point of comparison when running the same evaluation via API, it is not a number that settles the economics of closed models versus open weights. For short-term deployment, Muse's API pricing and its compatibility with existing tools may be the deciding factors. For organizations that prioritize data governance, regional requirements, or control over the inference infrastructure, GLM-5.2's open weights carry a different kind of value.
The numbers worth measuring before adoption are clear: fix your own use case, and track the total tokens—including retries needed to reach a successful outcome—the time required, and the resulting bill. Whether that result comes close to the $0.26 evaluation figure will determine whether Muse Spark 1.1's low price translates from a benchmark advantage into a practical one.