One sentence conclusion: In multiple rounds of dialogue, there is a piece of content that must be read again in each round. It doesn't get diluted as the conversation gets longer - you pay it a number of times equal to the number of turns you have in the conversation.
This is the indicator that Tunenomics believes is most worth looking at first. The reason is simple:Changing it will bring the most direct benefit.
Phenomenon
After a conversation is used, the bill changes often do not "increase linearly with the length of the conversation", but faster than linearly.
In terms of physical experience, the first few rounds are very cheap, and the later rounds become more expensive. Many people will think that "the conversation becomes longer and therefore more expensive" - the direction is right, but it is not accurate enough:What really keeps paying is not the entire conversation, but the fixed part of it.
Teardown: Which piece should be re-read in each round?
Think of a conversation as three paragraphs:
| Section | Each round changes unchanged | Paid features |
|---|---|---|
| Fixed instructions and rules (systematic part) | No change | It is not covered by the "remembered" offer,Read it again every round |
| Previous history (the more we talk, the longer it gets) | Every round is getting longer | will grow, but at least the "same content" is getting bigger |
| New content for this round | Every round is different | It is new in the first place, so it is unavoidable to pay the original price. |
The first paragraph is the most easily ignored: It looks like "it's always there and doesn't last long", but because it has to be paid once every round, its relationship with the number of rounds isMultiplication, not addition.
If a fixed description that is not long is placed in a position that needs to be reread every round, its cost is not "this paragraph", but "this paragraph × number of rounds". The longer the conversation runs, the more obvious this area becomes.
Conclusion
The reason why this section is worth looking at first is because it meets three conditions at the same time:
- Attributable: Can directly point to the specific content;
- Can be changed: Usually it is just a matter of location and writing, and does not affect the business logic;
- Can be recalculated: You can compare the changes before and after using the same bill.
Its corresponding actions are also very specific:Try to move the fixed part to a position where it can be "remembered"instead of having it pay from scratch every round.
Boundary (under what circumstances does this article not hold)
- Single round call and stateless call do not exist. This is only apparent with multiple rounds of conversations - if your business is primarily one-time calls, this article doesn't apply.
- What this article says about “rereading every round” isStructural location issues, it is not an implementation defect of a certain company;
Different services have different discounts for which segments, and any specific discounts must be accompanied byEffective Date。
- This article does not contain actual measured figures, nor does it predict province size.
How can you verify it yourself?
Get a usage breakdown of multi-turn conversations and do one thing:Group by round and see if the part that is being paid repeatedly in each round after the first round is getting smaller.
If it's the same size every round, then it's being recharged; if it's shrinking, it's been memorized. This step only looks at the files you exported yourself and does not need to be connected to any system.
Source and Description
- Source: Tunenomics Tunenomics unified external measurement basis (compare and review date: 2026-09-18).
- This article does not contain original customer data, actual measured figures and savings commitments.
This article is from Tunenomics’ “AI Cost Observation” series. The description of methods and the measurement basis of conclusions in the article can be checked; the numbers are marked with the source, measurement basis and date. Sample data does not represent actual results for any customer.