Meta's Muse Ember matches Claude Haiku 5.5 on Artificial Analysis Intelligence Index at one-third the cost

Meta announced Muse Ember on November 9, 2026, and benchmark results from Artificial Analysis, Vals AI, and Arena.ai were published the same day. The model is positioned as an efficient-tier offering, and the evaluations show a trade-off: it achieves overall intelligence comparable to Claude Haiku 5.5 while costing less per task, but its knowledge accuracy is notably lower.

Artificial Analysis attributes the lower processing cost to two factors: cached input tokens are 60% cheaper, and Muse Ember completes tasks using about one-third the output tokens of Claude Haiku 5.5, even though their API input and output token prices are identical. On Vals AI's Vals Index v2.2, Muse Ember ranked 12th out of 35 models with a score of 60.48%, at about $0.15 per test, the cheapest among evaluated models, and finished tests in about 15 minutes, the fastest among the top 12. Individual Vals AI evaluations placed it 4th of 36 on the Legal Index and tied for 2nd of 36 on HLAB-Vals-1.1, while it ranked 10th of 38 on the Finance Index, 15th of 62 on Terminal-Bench 4.0, and 14th of 83 on Tax Agent Bench.

Arena.ai reported that Muse Ember (max) scored 1608 points on Code Arena: WebDev, the highest among small models, ahead of Claude Haiku 5.5 (1595) and GPT-6 Luna (1578), though it ranks 34th overall. A personal evaluation by X user tmuxvim on the Revise agent harness found Muse Ember failed slightly more cases than GPT-6 Luna but cost about 25% less on the standard service tier and finished tasks about 40% faster than Luna on the fast service tier, suggesting promise as a free-tier workhorse model.