Saying “thank you” to an AI assistant costs little, but each message resends the entire conversation, and the energy cost increases with its length. According to an analysis published as an op-ed, it is therefore the duration of exchanges, more than politeness, that weighs on the bill, and the author proposes a method to limit it.
The starting point is a calculation shared on LinkedIn, which estimated the cost of “thank yous” addressed to ChatGPT and Claude based on 120 million users and 8 thanks per day, or about 350 billion “thank yous” per year. Using the calculator employed by the author, an isolated “thank you” addressed to a model the size of Claude Sonnet 5.5 represents about 23 GWh per year, versus less than 1 GWh for a small model. The choice of model therefore weighs about thirty times more than politeness.
The real factor, according to the author, lies in how these tools work. A language model has no memory from one message to the next: the application sends back, each turn, all or part of the exchange, i.e., the messages, responses, instructions, and attached documents. A “thank you” sent at the end of a long exchange thus comes with thousands of tokens, meaning the units of text the model processes. In the chosen scenario, with 5,000 tokens resent, the estimated energy is multiplied by ten, and the value at the API rate goes from about 100 million to 3.6 billion dollars per year.
A cost that grows faster than the conversation
Each new exchange replays all the previous ones: by the hundredth message, the model processes the other ninety-nine again. For a project of 60 exchanges, with messages of 200 tokens, responses of 300 tokens, and 1,500 tokens of instructions and files, the model must process nearly a million tokens in total.
Providers have a cache mechanism, which allows re-reading already processed history at a lower cost. At Anthropic, a cache read is billed 10 to 40 times cheaper than an input token, depending on the model: 20 times for Sonnet 5.5 and Opus 5.5, 40 times for Fable 5.1, 10 times for Haiku 5.5. The cache thus lightens the bill, but it does not solve the length issue.
Longer, and less reliable
A conversation that grows longer also performs less well. As early as 2023, researchers from Stanford led by Nelson Liu showed that models poorly exploit information placed in the middle of a long context. In July 2025, the company Chroma tested 18 models, including o3, Claude 4, and Gemini 2.5, on simple, controlled tasks: their reliability decreases as the input lengthens. Reasoning mode improves scores, without erasing the gap between short and long versions.
The NoLiMa benchmark, presented at the ICML 2025 conference by Adobe Research and LMU Munich University, points in the same direction. On a subset of 10 of the most difficult questions, OpenAI’s reasoning model o1 answers correctly in 92% of cases with 4,000 tokens of context, 78% at 8,000 tokens, and 31% at 32,000 tokens. A 2026 preprint, titled “Reasoning Shift,” further indicates that a conversation history reduces the reasoning duration of models and their ability to verify their own work.
The mother conversation method
The author proposes organizing the work before starting. A “mother” conversation sets the objective, breaks down the project into tasks, and does not produce deliverables. Each task is entrusted to a “daughter” conversation, opened by a brief specifying the objective, expected format, sources, and limits, then closed at the end. Reference documents remain in a folder and are only called when a daughter needs them. Each daughter only reports back to the mother a short summary, reviewed by the user, and the model is chosen at the opening of each conversation, according to the task.
According to the presented calculations, the project of 60 exchanges becomes six daughters of ten exchanges, managed by a mother. Compared to a single conversation conducted on the main model, the method would be 7 to 16% more energy-efficient and 22 to 28% cheaper. If the mother drafts responses twice as long, the energy gain drops to 20 to 39%. The author notes that these figures remain theoretical, particularly the rate of avoided rework, and that the EcoLogits method, used to estimate models from OpenAI, Google, or Mistral, only counts tokens produced. According to him, providers should publish the energy consumed per token read, written, and re-read in cache, because without this data, estimates rely on approximations.
