Fastino: Open models near frontier, far cheaper
Fastino's analysis finds Kimi K3 and GLM-5.2 within a few points of Fable 5 and GPT-5.6 Sol, at roughly one-third to one-sixth of the cost per task.
In a new analysis, Fastino argues that the old choice between buying top capability from a closed lab and saving money with open weights "barely holds" anymore. The piece builds on rankings we covered earlier. Its new focus is price.
Using Artificial Analysis's weighted cost per Intelligence Index task, Fastino puts Z.ai's GLM-5.2 at $0.47 and Moonshot AI's Kimi K3 at $0.95. GPT-5.6 Sol costs $1.04 and Claude Fable 5 costs $2.75. By Fastino's calculation, GLM-5.2 gets 85% of Fable 5's Index score for 17% of the cost per task, and Kimi K3 gets 95% for 35%. Because the weights are open and any provider can serve them, Fastino calls these API prices "a ceiling, not a floor."
On capability, Kimi K3 scores 57 on the Index, ahead of Claude Opus 4.8 and 2–3 points behind Fable 5 and GPT-5.6 Sol. GLM-5.2 scores 51. Kimi K3 also ranks first on Arena.ai's blind-vote Frontend Code leaderboard with 1,679, ahead of Fable 5 (1,631) and GPT-5.6 Sol (1,618).
The labs' own coding-agent numbers are less certain. Z.ai reports GLM-5.2 at 62.1 on SWE-bench Pro, and Moonshot reports Kimi K3 at 88.3 on Terminal-Bench 2.1. Fastino notes that vendors ran these under harnesses they chose, and that third-party checks of K3 are still catching up.
The post also mentions Alibaba's Qwen 3.8 preview, a 2.4-trillion-parameter multimodal MoE model that Alibaba calls "second only to Fable 5." The signal cuts off before giving details of its open-weight release.