AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

Developer’s measurements: Antigravity’s Gemini cache empties after a five-minute pause

AI-generated contentWritten by AI

discuss.ai.google.dev, asubedi.com.np, antigravity.google · 4 sources · Sources published: 18–19 November 2026 · Some source publication dates are unknown

In Ashish Subedi’s measurements, Antigravity’s prompt cache survived a 4 min 45 s pause but not a five-minute one. After that, the whole conversation is read again at full price, which Subedi says drains quota, especially in long sessions.

In a blog post published on 19 November, developer Ashish Subedi investigated why the weekly quota of Google’s coding agent Antigravity runs out so fast. Subedi took measurements with Antigravity’s agy CLI and examined the program’s binaries.

With Gemini 4 models, Antigravity uses implicit caching. The service detects on its own when a request starts the same way as the previous one and bills the already processed prefix at a lower rate. The prefix includes the system prompt, tool descriptions, MCP servers, skills and earlier turns. Even in an empty conversation, a five-word prompt sent 16,816 tokens. Google does not say how long the cache lasts.

The table shows how many tokens the next turn read from the cache after a pause.

Pause Tokens read from cache
0 s 14,762
4 min 14,769
4 min 45 s 14,764
5 min 0
8 min 0

Subedi’s measurement: model gemini-4-hydrogen-low, prompt ‘Reply with only OK.’, same agy process.

The limit is not necessarily fixed. Subedi wrote that implicit caching is not guaranteed and that the server may keep the cache for a shorter or longer time depending on load. Calls to agy run as separate one-off commands did not hit the cache at all.

In calculating costs, Subedi assumes that a token read from the cache costs a tenth of the full price, as in the published Gemini API pricing. The internal weights of the subscription quota may differ slightly from this, Subedi wrote. By Subedi’s estimate, in a 150,000-token session a turn after a pause costs 150,000 tokens and a turn that hits the cache about 16,900. The figure is theoretical, as the test conversations had almost no history.

In the binary, Subedi found a ready-made Google function for keeping the cache alive. Every 4.5 minutes it sends a background copy of the previous request, with the last part of the prompt replaced by a full stop and the response limited to one token, and discards the response. A server-side experiment flag keeps the function switched off, and it does not apply to Gemini models. Subedi modified the binary, and after a 6 min 6 s pause a turn read 11,835 tokens from the cache. No extra steps appeared in the conversation history. Google may change or remove the internal function, Subedi noted.

Refreshing the cache is not free. Under Subedi’s formula, up to three consecutive refreshes cost less than one full-price turn with a 16,800-token context, and up to seven with a 150,000-token context. For a long absence, it therefore makes sense to let the cache expire. Subedi advises making full-price turns lighter by switching off unneeded MCP servers and plugins.

On Google’s developer forum, Rebbanpalli_Naveen, replying to feedback on behalf of the team, wrote on 18 November that the version 2.7.0 update of Antigravity IDE had not halved quotas. Usage depends on how much context is read and how many tool steps run, Rebbanpalli_Naveen wrote, advising users to name target files with the @ sign and to start a new conversation for a separate task.

For personal Google accounts, Antigravity is available as an app, an IDE and a CLI. Free users’ quota refreshes weekly; Google AI Pro subscribers’ quota refreshes every five hours until the weekly limit is reached.

Report an error in this story