Jev Daily

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

One builder's measurement says the cache expires faster than an agent loop can reuse it, and the bill shows it.

kirill sh posted that Gemini 4 Neon's prompt caching is broken, measuring a 9% cache hit rate after a one minute gap. For comparison he gives roughly 95% at an hour for Opus 5.5 and about 97% at 24 hours for GPT in Codex. The practical consequence he reports is that Neon is burning Antigravity limits around 3x faster than it should.

kirill sh
@shneural
X
Gemini 4 Neon is burning your Antigravity limits ~3x faster. prompt caching is broken (9% hit after 1m)
Nov 19, 2026 · View on X

That is one person's measurement on their own usage, not a published eval, and the post does not describe the harness. But the shape of the number is easy to reason about. Prompt caching, where the provider stores the already processed prefix of your context so a follow up call does not pay full freight on it, is what makes a long running agent loop affordable. An agent that re-sends a large system prompt plus accumulated tool output on every turn depends on that prefix staying warm between turns. A minute is less than many agents spend waiting on a single tool call.

What it changes

It changes your bill and your rate limit headroom, not your output quality. kirill sh is explicit that the model itself is not the problem, calling Neon strong for agent work in the same post and asking Google to fix the infrastructure rather than the weights. So the failure mode is the quiet kind. Nothing breaks, nothing looks wrong in the traces, and you discover it when the limits hit earlier than your estimate said they would.

Theo agreed that the platform around the model needs work, saying he would go back to pushing Google on it once the honeymoon phase wears off. That is the second voice on it, and it is an opinion about the platform rather than an independent measurement of the cache.

For people building on it

If you are costing a Neon agent loop, do the arithmetic on uncached input tokens until Google says something about cache lifetime. Any spreadsheet that assumes an hour of cache retention, which is what the Opus 5.5 figure in the post implies, will be wrong by a large multiple on a loop with slow tool calls in it. The other lever is turn frequency. If the cache really is cold after a minute, batching work into fewer, fatter turns is worth more here than it is on providers with long retention windows.

This is the second Neon platform complaint we have covered this week, after Max Blade's report that the model was earlier unusable in his agent CLI. Neither is about the model's reasoning.

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

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

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.