# Token Use in Energy *April 2026* One of my favourite books, [Where Is My Flying Car?](https://press.stripe.com/where-is-my-flying-car), argues that watts are a better measure of living standards than dollars alone. So what if we applied that lens to tokens? This is a back-of-the-envelope sketch of my mid-April token use across Codex and Claude Code. ## The Back-of-the-Envelope ### Working Assumptions Three assumptions do most of the work: * Output tokens, including reasoning, cost about 5x an input token. That roughly matches public API pricing: `gpt-5.4` sits around 4.5-6x, and `opus-4.6` is 5x * Cache reads cost about 1/20th of a fresh input token. The same public API pricing prior is doing the work here too * An input token costs ~5J of direct and indirect energy That last number comes from [Jegham et al. (2025)](https://arxiv.org/abs/2505.09598). They estimate that `gpt-5` at `high` reasoning effort uses: | Input Tokens | Output Tokens | Wh Estimated | J per input-equivalent Estimated | | ------------ | ------------- | ------------ | -------------------------------- | | 0.1k | 0.3k | 10 | 22 | | 1k | 1k | 17 | 10 | | 10k | 1.5k | 33 | 7 | Because almost all my usage comes from coding at higher reasoning levels, 5 J per fresh input token feels like the right order of magnitude. ### April 1st-16th Across Codex and Claude Code, the first 16 days of April come out to: * 6.06B tracked tokens * 5.75B cache reads * 0.298B uncached input tokens, including cache writes * 0.0157B output tokens, including reasoning tokens ### Putting It Together The weighted sum: $ E \approx 5\,\mathrm{J}\times 0.298\times 10^9 + 0.25\,\mathrm{J}\times 5.746\times 10^9 + 25\,\mathrm{J}\times 0.0157\times 10^9 \approx 3.32\times 10^9\,\mathrm{J} \approx 922\,\mathrm{kWh} $ That is about **58 kWh/day** so far in April, or about **152 kWh per billion tokens**. This is equivalent to: * Driving about 380 km/day in my car at 150 Wh/km efficiency * Charging my 70 Wh MacBook about 820 times per day * Flying one-way from Sydney to Canberra every two days, using about [4 L/100 km/passenger from](https://www.iata.org/en/iata-repository/publications/economic-reports/global-outlook-for-air-transport-december-2025/)) and about [37.6 MJ/L for jet fuel](https://www.eia.gov/state/seds/sep_use/notes/use_petrol.pdf) * Roughly 2.5-5 Australian homes, using the government's indicative [11-23 kWh/day range](https://www.energy.gov.au/solar/solar-system-design/size-your-solar-system) ## Open Questions ### How Wrong Could This Be? Sensitivity analysis of kWh/B tokens, holding the April token mix fixed and moving one factor at a time: | Factor | x0.1 | x0.5 | x1 | x5 | x10 | | ----------------- | ---- | ---- | --- | --- | ---- | | J per input token | 15 | 76 | 152 | 761 | 1521 | | Output multiple | 136 | 143 | 152 | 224 | 314 | | Cache discount | 93 | 119 | 152 | 416 | 745 | We're comfortably in the 100s of kWh/B tokens range even if any of those factors half, double, or triple. ### So What? If you're spending 100-500M tokens per day, you're at roughly 15-76 kWh/day on the central estimate: from about one household-day of electricity up to several. If you're spending 1B+ tokens per day, you're at 152+ kWh/day: roughly 7-14 Australian households, or more than 1,000 km of driving per day at 150 Wh/km. Would I drive 380 km a day? Not even as a passenger. Would I charge my laptop 820 times? Those questions are still a bit silly, because transport, laptops, and household electricity all sit inside mature equilibria of price, friction, and utility. Marginal benefit already meets marginal cost. But the more interesting question is why token demand has not saturated already. A simple answer is subscriptions: with a fixed A$300/month plan, there is no visible marginal price signal between the first billion and the second. But that cannot be the whole story. Even when food or flights are prepaid, demand still tends to saturate on its own. So what is the natural saturation point for tokens, and have we found it already? We're yet to find out.