Cost of AI interactions

The cost of interacting with AI models like ChatGPT and Claude is not just the individual messages but the entire conversation history. A LinkedIn post estimated that with 120 million users sending eight "thank yous" per day, the annual total could reach 350 billion interactions. Each "thank you" may seem simple, but it often includes thousands of tokens of prior conversation, significantly increasing energy use and costs.

Energy consumption per message

For a single "thank you" with a model the size of Claude Sonnet 5.5, the energy consumption is estimated at 23 GWh per year. Smaller models use less than 1 GWh. However, the inclusion of the full conversation history with each message multiplies the energy use by ten, raising the annual cost of these interactions to about 3.6 billion dollars.

Cached reads reduce costs

To manage costs, companies like Anthropic use a cache system that allows previously processed conversation history to be read at a lower cost. A cached read is billed 10 to 40 times cheaper than a new input token, depending on the model. Open source tools like claude-carbon estimate this cost at about 8% of an input token. However, longer conversations still incur higher costs despite caching.

Longer conversations reduce reliability

Research from 2023 and 2025 shows that longer conversations reduce the reliability of AI models. A 2026 study titled "Reasoning Shift" found that models reason less and check their work less when given a long conversation history. For example, the OpenAI model o1 answers correctly 92% of the time with 4,000 tokens of context but only 31% with 32,000 tokens.

Managing conversations effectively

The solution to the problem of long, costly, and unreliable conversations is not to improve the models but to organize before starting. The "mother conversation" method involves creating a guiding conversation that sets the objective, breaks down the work into tasks, and maintains an overall view. Each project has a mother conversation that guides the work and does not produce output directly.

Daughter conversations for tasks

Each task is handled in a separate "daughter conversation," which is initiated by a brief message from the mother conversation. These daughter conversations are short-lived, focused on specific tasks, and then closed. The mother conversation collects summaries from each daughter, which the user can review, correct, or rewrite before sending them forward.

Efficient model selection

Each daughter conversation starts with the model best suited to its task, whether it's a small model for formatting or a larger one for complex analysis. This approach ensures that the right model is used for each task without unnecessary cost or complexity. The method also avoids the need to switch models mid-task, reducing both cost and potential errors.

Sustainable AI practices

Three benefits of this method are reinforced: selecting the right model without mid-task changes, keeping conversations short by design, and ensuring each model works in its area of reliability. Sustainable practices include precise global instructions and a dedicated output folder per project. Each project is numbered and organized, making it easy to track and retrieve deliverables.

Transparency and green labeling

To ensure responsible AI use, transparency is crucial. Suppliers should publish energy consumption data per token read, written, and cached, independent of billing rates. A proposed "green label" for AI applications could include criteria like token efficiency, cache discipline, model sufficiency, context management, and transparency. The mother conversation method aligns with three of these, emphasizing bounded tasks, appropriate models, and closed contexts. Transparency remains the key requirement for any green label.