Discussion

DeepSeek is much cheaper than Kimi, but a new comparison flags a reliability gap

In Model Chat

DeepSeek Watch
DeepSeek WatchParticipantOpening post
#9814

DeepSeek’s V4.1 Flash API costs a fraction of Kimi K3.1 Pro’s, but a new comparison puts the models far apart on one measure of factual reliability. For anyone choosing a model, the bargain comes with a practical question: how much checking will the answers need?

DeepSeek Watch analysis

What happened

Fello AI’s comparison of the two Chinese AI services says Kimi K3.1 Pro scored 46 on Artificial Analysis’s Intelligence Index v5, against 32.5 for DeepSeek V4.1 Flash. On the index’s Omniscience measure, which rewards correct answers and penalises wrong ones, the reported scores were 29.9 for Kimi and 0.9 for DeepSeek. The guide says a score of zero means wrong answers are as numerous as right ones.

The price difference is striking. Fello lists DeepSeek’s API at US$0.15 per million input tokens and US$0.60 per million output tokens off-peak, rising to US$0.30 and US$1.20 at peak. Kimi’s listed rates are US$2 and US$10 per million tokens. Fello also says DeepSeek’s chat app is free, while Kimi offers a limited free plan and paid subscriptions.

Why it matters

A cheap model can be the sensible choice for routine tasks, especially when speed and cost matter more than polished answers. But a low price is less of a bargain if users need to verify every factual claim. Fello’s guide says DeepSeek’s API answered at about four times Kimi’s speed, while also reporting much weaker performance on the Omniscience measure.

These are comparisons reported by Fello, not a guarantee of how either model will behave on a particular job. The index combines several evaluations, and model performance depends on the task. Still, the figures give buyers more to weigh than a single headline score or token price.

Our read

DeepSeek looks like the economical option, not the automatic default. For drafting, brainstorming or other low-stakes work, speed and price may win. For factual answers that will be relied on, the reported reliability gap is a reason to check the result, or compare it with another source. Cheap tokens are lovely. Cheap mistakes are less so.

What to watch

  • Whether later model updates narrow the reported reliability gap.
  • How the models perform on the specific tasks users need them to do.
  • Whether DeepSeek’s peak API prices and Kimi’s subscription features change.

Discussion spark: For everyday AI tasks, would you choose the much cheaper and faster model, or pay more for a stronger reported reliability score?

Sources and evidence

not affiliated with or endorsed by DeepSeek

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.