AI Disclosure

AI disclosure

Most of ha-taggle was made by AI coding agents, with a human in charge. That includes the radio reverse engineering, the receiver, the add-on, the card, the tests, the diagrams and the docs (this wiki too).

Who did what

Who	Role
Anthropic Claude (Claude Code, Opus)	Planning, live probing, review and documentation
OpenAI Codex	Most of the code and signal analysis
A human	Chose the goals, captured and confirmed the signal from their own meter, checked results against the physical dial, and approved every change

Warning

ha-taggle remains experimental. Compare usage with changes on your physical meter's dial over several days. If they disagree, check calibration and reception before relying on the result.

What it cost (as of 2 November 2026)

Agent	Model calls	Tokens processed	of which cached re-reads	Output tokens
Claude (incl. sub-agents)	~380	~57.5 M	~56.5 M	~0.21 M
Codex	~685	~61.8 M	~59.7 M	~0.55 M (incl. ~0.18 M reasoning)
Total	~1,065	~119 M	~116 M (97%)	~0.76 M

These were counted from the local agent session logs for the Taggle research and the repository. Earlier, unrelated Home Assistant work isn't included. A little unrelated chat in the same sessions may be.

What's a "token"?

AI models read and write text in small chunks called tokens. A token is often a short word or part of a longer word. "M" in the table means million.

• A model call is one request to the AI: "here's everything so far, what next?"
• Tokens processed is everything the model read across all those calls.
• Output tokens is what it wrote: code, text and (for Codex) hidden "reasoning" steps.

What does "cached re-reads" mean?

An agent works in many small steps. At each step it sends the whole conversation so far (instructions, files it has read, earlier steps) back to the model. Most of that text is identical to last time.

The AI service keeps a cache (a short-term memory) of text it has just seen. When the same text arrives again, it can reuse its earlier work instead of processing it all from scratch. That's a cached re-read, or prompt cache hit.

That's why 97% of the tokens are cached re-reads. They count in the total, but they take far less computing than writing new text.

Estimated water use

Data centres use water, mostly for cooling. Nobody publishes a water figure for long AI coding sessions, so the project gives a range (a floor and a ceiling), not a measurement.

Floor: about 0.3 L

~1,065 model calls × 0.26–0.32 mL per call.

• 0.26 mL is Google's figure for a median Gemini text prompt. Google describes it as "the equivalent of five drops of water".
• 0.32 mL comes from OpenAI's CEO, who wrote that an average ChatGPT query uses "about 0.000085 gallons of water". That's roughly 0.32 mL.

Note

Why it's a floor: those figures are for a typical chat question. Each agent call here carried roughly 100,000 tokens of context, far more than a chat prompt. So this certainly undercounts.

Ceiling: about 85 L

~0.76 M output tokens, using Mistral's figure of 45 mL of water per 400-token response. Mistral took that figure from its life-cycle analysis and calls it the "marginal impact of inference".

The sum: 760,000 ÷ 400 × 45 mL ≈ 85 L.

Why a range and not one number?

• Companies measure differently. Some count only the data centre, others include more of the life cycle.
• The published figures are per chat question or per response. None is for an agent re-reading 100,000 tokens of context.
• Cached re-reads are cheap, but nobody publishes exactly how cheap in water terms.

So the honest answer is a bracket.

In everyday terms

Somewhere between a water bottle and roughly ten minutes under a 9 L/min shower, most plausibly in the tens of litres.

Tip

Fittingly, once calibrated, this receiver would show a household using that much water within a single hour's reading.

Estimated energy use

Electricity works the same way: a bracket built from published per-question figures, not a meter reading.

Floor: about 0.3 kWh

~1,065 model calls × 0.24–0.34 Wh. Those are Google's figure for a median Gemini prompt and OpenAI's figure for an average ChatGPT query. It's too low, for the same reason as the water floor: our calls carried far more context than a chat question.

Ceiling: about 40 kWh

~1,065 calls × ~40 Wh. Epoch AI estimates almost 40 Wh for one query with about 100,000 input tokens. It's too high: it assumes the whole context is worked through from scratch every time, but ~97% of our tokens were cached re-reads that skip most of that work.

In everyday terms

Somewhere between boiling a 2 kW kettle for ten minutes and running a 2 kW heater for about 20 hours, most plausibly a few kilowatt-hours.

For carbon, Mistral's figure of 1.14 g CO₂e per 400-token response gives roughly 2 kg CO₂e for ~0.76 M output tokens. The real number depends a lot on how clean each data centre's electricity is.

Further reading

• ha-taggle README: AI disclosure
• Epoch AI: how much energy does ChatGPT use?
• Google: measuring the environmental impact of AI inference (arXiv)
• Sam Altman: "The Gentle Singularity"
• Mistral AI: our contribution to a global environmental standard for AI
• Contributing