Anthropic releases Claude Fable 5.2 at $8/$40, topping AA index at 65

Anthropic released Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens, with the model topping the Artificial Analysis index at 65. Per-token price fell 20% from Fable 5.1, but because the model writes more tests and spawns checking subagents, GDPx cost per task rose from $67.16 to $88.15, and Artificial Analysis measured $26.54 and 46 minutes per index task, among the slowest and costliest. The release matches Claude Haiku 5.5's index score at about a third of the cost per task.

Per-token price fell 20% from Fable 5.1, but the model writes more tests and spawns checking subagents, so GDPx cost per task rose from $67.16 to $88.15. Artificial Analysis measured $26.54 and 46 minutes per index task, among the slowest and costliest. Read: Per-token price fell 20% from Fable 5.1, but the model writes more tests and spawns checking subagents, so GDPx cost per task rose from $67.16 to $88.15. Artificial Analysis measured $26.54 and 46 minutes per index task, among the slowest and costliest. Read: That matches Claude Haiku 5.5's index score at about a third of the cost per task. Tool calls, routing and extraction run at high volume inside agents, so the cheap tier of a multi-model stack now costs less without losing measured capability. Read: Small classifiers and routers now take a quarter of production traffic for about half a cent of every dollar. Amazon engineers report saving millions by replacing Sonnet 4.6 with one in a large ticketing system, so routine agent decisions are leaving frontier models. Try: The project reports up to 33 points higher benchmark scores at 2.7 to 5.2x lower cost after removing 136,866 lines. New defaults send runs to its cloud-hosted model and browser, so self-hosted users who upgrade without changing config move off their own setup. Try: Bash commands could leave the sandbox without a prompt, and MCP OAuth secrets leaked on redirects. Unattended agents rely on the sandbox and on credential handling to contain mistakes, so older versions expose both the host and connected services. Read: Raw FLOP/s comparisons understate newer chips for serving by a factor of about ten. Memory and interconnect set how many concurrent requests a chip handles at a given speed, so inference capacity is growing far faster than compute headline numbers show.