Meta's multimodal AI model climbed eight spots on the Vals Index after its recent launch, establishing a new efficiency tier for coding tasks. The Muse Spark 1.3 system currently ranks 8th overall, which is a jump from its previous position at 16th. While it trails top models like Claude Fable 5.1 by 6 to 9 points in accuracy, it is the most powerful option available for under $3 per task, while higher-ranked peers cost 2 to 10 times more.
The model incorporates a 1 million token context window and 131,000 output tokens when run with xhigh reasoning effort. Because Spark 1.3 costs roughly 20 times less than rivals like Sol or Fable, developers are considering routing execution and refinement stages to this cheaper model rather than using a single system for both ideation and final output.
Key sources
- SOURCE@valsai“every model above it costs 2–10x more per task”x.com
- SUPPORT@valsai“still 6–9 pts behind on the Index”x.com
- SUPPORT@valsai“1M context window, 131k output tokens, and is multimodal”x.com
- SUPPORT@arankomatsuzaki“Ideation still looks expensive, but execution/refinement might get cheap pretty fast”x.com
- SOURCE@alexandr_wang“significantly stronger coding and agentic performance on muse spark 1.3 max”x.com
- SOURCE@aiatmeta“available on Muse Code and Meta Model API”x.com
- SUPPORT@alexandr_wang“set reasoning level to max”x.com
- SUPPORT@finkd“recommend trying it out even if you've already tried muse spark 1.3 high”x.com