How much does Dex save?
It depends on the LLM you use today and on how much of your traffic is decisions. In our six scenarios at 50,000 events a month, Dex plus Claude Fable 5.1 costs 36% to 85% less than Claude Fable 5.1 alone, and Dex plus Claude Sonnet 5 costs 86% to 95% less. For 1,000 support tickets, Dex costs EUR 0.11 at the Fierce plan rate, against EUR 51.11 with Claude Fable 5.1.
Is Dex always cheaper than an LLM?
No. Next to GPT-4.1 mini, the cheapest LLM we raced, Dex saves money in 3 of 6 scenarios at 50,000 events a month and in 6 of 6 at 500,000. At pay as you go, 1,000 support tickets cost EUR 1.21 with Dex and EUR 0.66 with GPT-4.1 mini. The saving grows with volume, because a plan's weekly allowance makes each token cheaper.
Where do the numbers come from?
Decision tokens and times come from our recorded races of 2026-09-25: the same questions answered by Dex and by 8 LLMs, 10 runs each. LLM costs are those tokens times the providers' list prices, input and output priced separately. The tokens of text an LLM writes are estimates, marked [E], and you can change them.
Why do the Dex columns still include an LLM?
Dex answers typed questions: pick, rate and check. It does not write text. In every scenario an LLM still writes the replies, drafts, emails, agent steps or answers; Dex takes over the decisions around them, so the LLM runs less often or can be a smaller model.
Which Dex plan do the scenarios use?
The plan that fits each month: pay as you go, or the plan whose weekly allowance, resets and use beyond the allowance cost least for that volume, with the same rules as the pricing page. The headline numbers use the Fierce plan rate and show pay as you go next to it.