Anyone who uses AI in software development is no longer working with just a chat window. Agentic tools – that is, AI systems that independently plan multiple work steps, read files, write code and check results – generate a multiple of the computing load per job. A current self-measurement now provides concrete figures for the first time.
What was measured
The climate scientist Zeke Hausfather logged his own consumption while using Claude Code over eight weeks. The result: around 3.2 billion processed tokens and about 170 kilowatt hours of data centre electricity. Converted to the individual prompt, according to his analysis this corresponds to roughly 600 times that of an ordinary AI chat prompt.
Tokens are the processing units of a language model, roughly word components. The fact that billions of them accumulate here is inherent to agentic work: the agent reads in context, repeats steps, corrects itself and writes intermediate results back. A single job in the chat window can thereby trigger dozens or hundreds of model calls.
Why the communicated low values are misleading
Major providers such as Google and OpenAI have published very low energy values per request in the past. These figures refer to simple, short chat interactions. For assessing agentic use, according to Hausfather's measurement they are of little informative value, because they describe a usage pattern that hardly dominates everyday development any more.
For project managers this means above all: metrics from provider communications cannot simply be transferred to one's own use case. Anyone planning sustainability goals or budgets needs figures from their own operations.
Three levels on which this becomes relevant
- Budget: Token consumption is the billing basis for many AI services. If agentic tools increase the volume by orders of magnitude, cost forecasts shift accordingly – regardless of which pricing model currently applies.
- Sustainability: Companies that account for environmental reports or internal climate targets should no longer treat AI-supported development as a negligible item. A single person can cause consumption over eight weeks that can be measured.
- Architecture: The question is not only whether an agent is used, but how much context it has to process at each step. Task scoping, project structure and documentation have a direct effect on token volume.
What can practically be derived from this
The figures are a single measurement and not an industry average. They are therefore not suitable as a benchmark, but they are a reason to make one's own consumption visible in the first place. Three steps make sense.
- Measure consumption instead of estimating it. Most providers offer token statistics. Anyone who evaluates these per project or team quickly recognises which types of task are expensive.
- Choose the tool to match the task. Not every request needs an agent with full project context. For wording assistance, code explanations or short refactorings, a simple model call is often enough.
- Limit context deliberately. Clearly delineated modules, clean interfaces and concise, precise project documentation reduce the amount of material an agent has to read in on each run.
The benefit remains – the calculation becomes more honest
Nothing about these figures argues against agentic development. When an agent prepares a migration, adds tests or documents legacy code, it saves working time that would otherwise also tie up resources. What matters is that the effort is estimated realistically and not on the basis of values collected for a completely different usage pattern.
That is precisely the real message of the measurement: agentic AI is not a chat with additional functions, but a different mode of operation with its own cost structure. It should also be planned as such.
Classification for web and software projects
In TYPO3 and web projects, it is worth treating AI use in future like any other infrastructure decision: with measurement, allocation to cost centres and a deliberate choice between a simple model call and a full agent run. Anyone who tidies up project structure and documentation thereby lowers not only the barrier to entry for new team members, but also the context requirements of the tools. This way the productivity gain is retained without budget and sustainability figures causing surprises later.