Gemini 4 Argon introduced noticeable shifts in prompt-caching behavior that teams building AI agents should understand. At Induction AI, researchers explored how caching compares to earlier models and identified several quirks that affect cost optimization.
Basics of the prompt-cache system
The prompt cache operates by reusing tokens the provider has seen before. When a user only ever appends to a thread, a discount applies to those previously seen tokens. Like recent changes from OpenAI and Anthropic, Gemini 4 offers a 95 percent discount off the normal token price. Getting this right matters for inference-cost management.
Researchers tested caching by appending to a thread to locate the threshold. On Argon, caching activates at the documented 4096-token mark. Sending 4000 tokens on step one produces no cache hit on the next turn. Sending 4100 tokens triggers caching on the following turn. This stands in contrast to Gemini 3.1 Pro Preview, where caching proved dependable only after 9000 tokens in similar tests.
The 2049-token quirk
Caching behavior operates on explicit breakpoints — defined points on each request where reuse becomes possible. OpenAI's implicit breakpoint sits at the end of each request. Anthropic and other newer models follow a similar pattern. Gemini 4 sets its breakpoints at the end of the request, consistent with other providers, but also inserts another breakpoint 2049 tokens before the end of every request. This backward-looking approach is new behavior not observed in earlier models.
When 4100 tokens are sent and then 1000 more are appended, Gemini 4 caches the 4100 from the prior call and charges full price for the new 1000. On a subsequent turn of 1500 tokens, the same logic applies — only the new tokens incur full cost. What makes Gemini 4 distinct is the requirement to always pay for at least 2049 tokens. If the new content is only 1000 tokens, the model reaches backward into the previous request. In this scenario, it accesses only 2051 tokens from the prior message (4100 minus 2049) and the user pays for 3049 tokens (2049 plus 1000).
This backward-reach continues on later requests. Sending 1500 more tokens causes the model to go back to the -2049 breakpoint from the previous turn. Researchers observed that the original 4100 breakpoint does not persist into the third request, suggesting it is removed when not immediately used or that only the most recent turns' breakpoints remain active. However, if the second message exceeds 2049 tokens — for example, 2500 — the entire first message gets cached. The user then pays for 2500 tokens and 4100 from the first message becomes cached. If a subsequent message does not meet the 2049 threshold, the model again reaches backward into the prior request.
Cost implications for Gemini users
This breakpoint structure makes Gemini 4 Argon inefficient when messages are short by nature, because a minimum of 2049 tokens must be paid for regardless of actual content length. When messages consistently exceed that threshold, caching works as expected. The system creates a counter-intuitive situation where sending more tokens can be better — adding content may hit the breakpoint and enable reuse of prior tokens rather than forcing payment for the minimum 2049.
Induction AI is researching these patterns to improve context-management layers for customers. The team plans to update its benchmarks to account for this caching behavior and help agents become more cost-efficient when using Gemini 4.