AI News DailyTips한국어
Sunday, October 18, 2026

Story 6 of 12

Theo: Gemini is '50% more expensive than it should be' because its cache misses 27% of the time

First seen on X 32 hours ago@theo ♥ 2,4351 views

Theo: Gemini is '50% more expensive than it should be' because its cache misses 27% of the time

Developer and YouTuber Theo claimed Google's Gemini API prompt cache works so poorly that real costs run 50% higher than they should.

"Gemini is 50% more expensive than it should be because their cache doesn't work. 27% cache miss rate," he wrote on X on Oct. 17, adding that it was a very basic cache test simulating normal agentic work, run identically against three APIs. The post drew 2,400 likes and more than 70,000 views.

Not everyone agreed. Windsurf co-founder Varun Mohan replied, "Not what we're seeing on our workloads," saying the miss rate is very low, and asked what Theo was running. Prompt caching stores repeated input so later requests are billed at a discount; a miss means the full input is charged at list price.

“Gemini is 50% more expensive than it should be because their cache doesn't work. 27% cache miss rate. Insane.”

Get notified when new stories land

Turn on notifications in the app. Free, no sign-up.

Get the app

More stories from the same day

  1. OpenCode breached in 'distillation attack'; users' OpenAI and Anthropic balances drained
  2. AI assistant Hark's $1 million free-pizza stunt backfires with surprise charges and inbox access
  3. Fearing AI clones, indie publisher cancels public demo of 'Drunken Goddess Reflux'
  4. OpenAI resets Codex usage limits again after users report quotas vanishing overnight
  5. Anthropic's junior share of technical hires falls from 27% to 9%; agents log six times the hours of humans
  6. Anthropic offers Blackstone a yield above 10% on $60 billion chip debt
  7. Claude Fable 5.5 rumored for next week: 'partners already have it,' and 'I built this with it' demos pile up
  8. Ben Horowitz calls Anthropic 'one of the most negligent companies' on security, says its reply to a hacked test site was 'you prompted it wrong'
  9. The web is shutting out AI agents: Delta, JPMorgan, Schwab and Yelp join Amazon
  10. Three mystery OpenAI models, 'aurora-red/green/blue-alpha,' pop up in the API — testers say they're image models
  11. Physicist says Claude Opus 5.5 redid a year of PhD quantum-gravity math in 10 hours — and caught a sign error
See all of Sunday, October 18, 2026 →

* Summaries and publishing on this page are done by an automated agent. Check the facts against the original links.

Privacy PolicySupportClaude Code & CodexSlack botListed on Sell With boost