Pause for five minutes in Google Antigravity and your next message may cost many times more than the last one. That is the finding from a developer who runs the blog asubedi.com.np, who measured the tool’s prompt cache and published the results on November 19, 2026. These are one person’s measurements on one machine (a Mac with an arm64 chip), not Google’s numbers. They are detailed enough to be worth a look.

First, the plain version of the vocabulary. A token is a small chunk of text, roughly a word fragment, and AI usage is metered in them. A prompt cache is a short-term memory on the provider’s servers: when your new message starts with the same text as your last one, the server reuses its earlier work instead of reading everything again. Antigravity uses what the blogger calls implicit caching on Gemini models, an automatic, best-effort match. The post says cached tokens get a 90 percent discount against regular input, and a miss means paying full price for the whole prefix.

The blogger found where the cliff sits by testing idle gaps. Waits of 4 minutes, 4 minutes 30 seconds, and 4 minutes 45 seconds all produced cache hits, with about 14,764 to 14,769 tokens read from cache. At 5 minutes the cache read dropped to zero, and the same happened at 8 minutes. The boundary sits between 4

and 5
. The post says that matches a 285 second staleness cutoff it found inside Google’s own code.

The second finding is about the starting weight of every conversation. Even a five word prompt (“Reply with only OK.”) opened with 16,816 tokens, which the blogger attributes to system instructions, tool declarations, MCP servers (the Model Context Protocol connectors that link agents to outside tools), and skills. Across tests the baseline ranged from 13,800 to 16,800 and more. In the same process, the second turn read 14,762 tokens from cache and billed only 2,152 fresh ones.

Why does this drain a weekly quota? Coding rarely moves at chat speed. You read a diff, run a test, think, and come back after five minutes. The blogger’s arithmetic, weighting cached tokens at one tenth, puts a cold turn near 16,900 effective tokens against about 3,640 for a warm one. In a hypothetical session with 150,000 tokens of context, a single expired cache costs about 133,065 extra effective tokens, by the post’s math.

The post also reports something stranger. Antigravity’s binary appears to contain an unused keep-alive routine that would send a tiny background request every 4 minutes 30 seconds to refresh the cache. The blogger says two internal checks switch it off, one tied to a server-side experiment flag and one that skips Gemini models. After patching those checks on a local copy, a 6 minute 6 second gap still returned 11,835 cached tokens. Google has not commented in the post, and the author warns the code could change or vanish in a future release.

Refreshing is not free, and the post says it stops paying off quickly. With the roughly 16,800 token prefix, the break-even point is 3 pings, about 13.5 minutes of pinging and 18 minutes of idle coverage. At 150,000 tokens it is 7 pings. The blogger also tested the /btw side-question command as a keep-alive. It worked, but it cost 17,697 uncached tokens, about 26 percent more than letting the cache lapse once.

The caveats are the author’s own. The discount figure is Google’s published one, internal subscription quota weights may differ, and implicit caching is best-effort, so expiry times could shift under load. Nobody else has reproduced the numbers.

The advice that needs no patched binary is the last item: disable MCP servers and plugins you are not using, because that overhead is paid again on every cold turn. If your weekly quota vanishes faster than your work suggests, audit your connectors before you blame your habits, and avoid walking away mid-task for exactly five minutes.

Reported from the personal blog asubedi.com.np, post “Why Antigravity Eats Weekly Quota Fast”, published November 19, 2026.