I’ve never seen a gemini model cache something. I just sent the same 1900 token message to Gemini 3 Flash, with several seconds between each request, but i just never see anything in usageMetadata indicating caching.
I’ve been running benchmarks with native google-genai and found that the 1024 token threshold mentioned in the documentation is inaccurate. Your 1900 token prompt isn’t triggering caching because the actual activation floor is significantly higher.
In my tests, gemini-3-flash-preview and gemini-3.1-flash-lite-preview only showed the first cache hit at roughly 4192 tokens. For gemini-3.1-pro-preview, the threshold jumped to approximately 8161 tokens.
This behavior is problematic since the industry standard for prompt caching is typically 1k (max 4k). Currently, prompts in the 1k to 4k range are missing out on caching, leading to unnecessary costs, latency, and energy waste.
Wow, that’s very unfortunate, thanks for sharing your findings.
The implicit cache situation just got worst these 2 weeks on release of gemini 3.7 flash which works really nicely with my app except it now cost more than 2x as much because implicit caching is not working. I confirmed it is not an issue with my code because a swap to gpt model gets me >70% cache hit rate. with every turn and tool calls. With gemini 3.7 flash, this week, implicit cache didn’t trigger untill I have >12k input tokens! What’s worst is that it didn’t get triggered on every tool call! This got worst since I make this post showing how the threshold went up to ~10k and only cached the first ~8k in June-July 2026. How are you guys being affected by this issue?
Hi @Mathew
Can you share more details about the issue you are facing with me via DM
It would be helpful if you have any comparision between what you are seeing for 3.7 vs older models
Also could you share your projectID/billing ID with me ?
It would be nice if this is support from Google. However, I am not sure if you are really from Google. Will you be able to create an official support ticket for this issue and proof that I am not just sharing my raw data with anyone? Thanks.