Jev Daily

PostHog puts 444.6x cheaper on decision models

PostHog says decision models like Jev do some of its work up to 444.6x cheaper and 193.6x faster than constant LLM calls, without saying which work or against which model. Elsewhere, Gemini 4 Neon's prompt cache was clocked at a 9% hit rate after one minute.

Someone measured it

PostHog says decision models run parts of its product 444.6x cheaper

PostHog published how it uses decision models, saying Jev can do some of the work up to 444.6x cheaper and 193.6x faster than constant LLM calls. The post does not say which workload, which LLM it was measured against, or how the two were run. Diogo Almeida noted that half of his favourite software companies now use Jev.

PostHog
@posthog
X
Building a product that requires constant LLM calls gets real expensive, real fast.
Nov 19, 2026 · View on X

Read the full story

Gemini 4 Neon's prompt cache measured at 9% hit after one minute

kirill sh says Gemini 4 Neon's prompt caching is broken, hitting 9% after one minute against roughly 95% at an hour for Opus 5.5 and about 97% at 24 hours for GPT in Codex. He says that burns Antigravity limits around 3x faster, while still rating Neon strong for agent work. Theo agrees the platform needs fixing.

Theo - t3.gg
@theo
X
Neon is the first Gemini model that’s good enough to actually deserve a functioning platform.
Nov 20, 2026 · View on X

Read the full story

The case against

Theo says GPT-6.2 Astra stalls asking questions, then burns the task

Theo describes using GPT-6.2 Astra as a waiter who interrupts a steak order to ask what temperature you want the fries cooked at, holds up the kitchen waiting for the answer, and then returns with the steak burnt. It is an analogy, not a measurement, and he posts no run, model settings or task.

Theo - t3.gg
@theo
X
The waiter comes out with mostly passable fries, and an entirely burnt steak
Nov 19, 2026 · View on X

Read the full story

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.