Artificial Analysis scores StepFun's Step 5 Preview at 44 on its index
The evaluator said the model matches Kimi K3 at about a third of the cost per task, but trails rivals on agentic work.
StepFun's Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, level with Kimi K3 at its maximum setting, the evaluator said on Monday. The model has 600 billion total parameters and 27 billion active, and succeeds Step 3.7 Flash, released in May 2026.
Artificial Analysis put the cost of running its index on Step 5 Preview at about $0.72 per task. It said Kimi K3 at maximum cost about $2.00 for the same score, and GLM-5.3 at maximum about $2.01 for one point more. The evaluator's page lists the model at $1 per million input tokens and $2.70 per million output tokens.
The gap is on agentic work. Artificial Analysis said Step 5 Preview sits behind GLM-5.3 at maximum and Qwen3.8 Max on GDPval-AA, at 1,566 Elo against 1,646 and 1,668, and on AA-Briefcase, at 1,432 against 1,525 and 1,640. On Terminal-Bench 4.0 it scored 33% against 42% and 39%.
On knowledge the ordering reverses. The evaluator reported 46% on Humanity's Last Exam against 42% and 43% for the same two models, and 42% on AA-Omniscience against 34% for GLM-5.3 at maximum and 48% for Kimi K3 at maximum. Step 5 Preview has 600 billion total parameters against 753 billion and 2.8 trillion.
Token use does not explain the price gap, Artificial Analysis said, putting Step 5 Preview at about 164 million output tokens to run the index against about 170 million for Kimi K3 at maximum. It said the model hallucinates on 43% of attempts rather than declining to answer. The index combines several evaluations, according to Artificial Analysis, and every figure here comes from that one group.