A 20% price cut that doubled the cost per task

Anthropic cut Claude Fable 5.2's token price by 20%. A day later Google shipped Gemini 4 Carbon at a quarter of Fable's price.

But per-token price is the wrong number. In Artificial Analysis's testing, one Fable 5.2 task at max effort costs $26.54, versus $12.99 for Fable 5.1, because the new model writes about 1.7 times as many output tokens.

The lever that matters is the effort setting. At max effort, Carbon wins on cost by a mile. Change that one setting and most of the gap disappears.

Takeaway: measure cost per completed task, and test your workload at a few effort levels before you switch models on a pricing headline.

Full breakdown (free friend link): https://pub.towardsai.net/gemini-4-carbon-just-made-claude-fable-5-2-look-overpriced-e13e111430c9?sk=4cfc9e36bbc1291656663955dcadf0ab