Muse Spark Benchmark Reality Check: Competes With Opus 4.6 and GPT-5.4 But Meta Admits Coding and Agent Gaps

Gizmodo·medium signal

Despite Meta claiming Muse Spark performs on par with or exceeds leading models from OpenAI, Anthropic, and Google, Simon Willison notes it lags on Terminal-Bench 2.0 and Meta itself acknowledges gaps in agentic systems and coding tasks. Gizmodo's headline — 'Doesn't Exactly Spark Joy' — captured the skeptic take. Two modes ship today (Instant and Thinking), with a promised 'Contemplating' mode that runs sub-agents in parallel still forthcoming.