Summary
- Meta released Muse Spark 1.3 on September 2, saying it improved performance on agentic and coding tasks along with real-world usability.
- In benchmarks released alongside the model, Muse Spark 1.3 outperformed GPT-5.6 Sol and Opus 5 on coding metrics like DeepSWE and SWEAtlas, as well as on the long-context MRCR benchmark.
- On broader knowledge-work benchmarks like GDPVal-AA v2 and JobBench, though, Opus 5 still posted the highest scores.
Meta ships a new Muse Spark version within a month
Meta released its large language model Muse Spark 1.3 on September 2 (local time) via the official AI at Meta account on X. The company said the model improves performance on agentic tasks — meaning it can handle multiple steps on its own — as well as coding, with a focus on usability in real work settings. In its announcement, Meta described the model as able to juggle multiple workflows within a single conversation, handle longer-running tasks, and collaborate more actively with users by asking its own questions when needed.
To put that in context, GPT-5.6 Sol is OpenAI's latest flagship model, and Opus 5 is Anthropic's top-tier model. Meta placing both models side by side in its benchmark table signals that all three companies are now measuring their agent and coding performance against the same yardstick.
Ahead of GPT-5.6 and Opus 5 on coding and long context
The benchmark table Meta released alongside the model shows Muse Spark 1.3 outperforming rivals on metrics that measure coding and long-context processing. On DeepSWE v1.1, which tests long-running software work across large repositories, it scored 75.4, beating both GPT-5.6 Sol (73.0) and Opus 5 (74.0). On SWEAtlas CodeBase QnA, which measures codebase comprehension, it scored 59.4, ahead of the other two models' 53.5 and 52.7. The gap was even wider on MRCR, which tests how well a model handles long documents: in the 512K-to-1M-token range, Muse Spark 1.3 scored 98.1, far outpacing GPT-5.6 Sol's 73.8, and Opus 5's score for that range wasn't disclosed at all.
Opus 5 still leads on broad knowledge work
On GDPVal-AA v2, however, which measures broad office knowledge work, Opus 5 posted the top score at 1824, with Muse Spark 1.3 close behind at 1754. Opus 5 also edged ahead on JobBench, which tests real-world use of multiple work tools (65.7 versus Muse Spark 1.3's 64.9), on OSWorld 2.0, which tests direct computer operation (68.3 versus 66.9), and on AutomationBench, which measures end-to-end workflow handling (50.3 versus 49.4). GPT-5.6 Sol, meanwhile, scored highest on the internal Agentic IF Index, which tracks how precisely a model follows instructions, and on DeepSearchQA, which measures autonomous web research ability.
Benchmark table
| Benchmark (category) | Muse Spark 1.3 | Muse Spark 1.2 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| GDPVal-AA v2 (knowledge work) | 1754 | 1615 | 1710 | 1824 |
| JobBench (tool use) | 64.9 | 61.6 | 45.4 | 65.7 |
| OSWorld 2.0 (computer use) | 66.9 | 47.6 | 62.7 | 68.3 |
| DeepSWE v1.1 (coding) | 75.4 | 55.0 | 73.0 | 74.0 |
| MRCR 512K-1M (long context) | 98.1 | 55.5 | 73.8 | - |
| Terminal-Bench 2.1 (terminal coding) | 88.8 | 82.9 | 88.8 | 86.7 |
Fourth release in a single month
Meta shipped three Muse-line models last month alone. Muse Spark 1.3 is the fourth release in that string, and it marks the biggest jump in coding and long-context performance among the models the company has released over the past four months.
Editor's take
Meta rolling out its fourth Muse-line model in under a month isn't a coincidence. While OpenAI and Anthropic stake out the agent and coding market with GPT-5.6 Sol and Opus 5 respectively, Meta keeps attaching benchmark tables to every release to publicly show exactly how far it's caught up. The scores make clear where Meta has poured its development resources: long-context processing and large-codebase comprehension.
Which model makes sense in practice still depends on the task. For coding work involving long documents or large repositories, Muse Spark 1.3 may currently be the strongest option, but for work spanning multiple tools and broader knowledge-work tasks, Opus 5 still holds a slight edge. For Korean dev teams, it would be safer to first test Muse Spark 1.3 on tasks where long context matters most, like code review and refactoring, and compare against your current model before switching over for general office automation.
Given Meta's monthly release cadence, the next update is likely to focus on closing the gap with GPT-5.6 Sol on instruction-following (the Agentic IF Index) and web research (DeepSearchQA).
Comments