Estimate the energy and emissions behind your AI use. No token counts? Start with an example.
Your usage #
K = thousand · M = million · B = billion · T = trillion
Examples replace all settings; energy estimates are modeled.
Advanced settings #
Your estimated footprint #
Enter token usage to calculate.
Energy breakdown
— Cached input
— Output
—
In everyday terms #
Select a comparison for its assumptions and source.
Comparisons are alternative yardsticks, not amounts to add together. Energy-only benchmarks are separate from emissions.
Methodology & benchmark assumptions #
Example energy estimates are modeled, not measured. The chat and agent assistant presets use illustrative token counts, not measured averages. “This page” uses a rounded estimate of development usage, not recorded token totals. “Major project” uses the reported token totals from Anthropic’s compiler experiment —2B input and 140M output tokens across nearly 2,000 sessions. That does not establish a typical feature size, cache rate, or input/output ratio. Claude Code’s usage documentation separates fresh input, cache reads, and cache writes; include all three in this calculator’s input total. Example cache rates and model classes are assumptions, not measured characteristics of those workloads.
“Agent team project” uses
[Bun’s reported pre-merge Zig-to-Rust rewrite usage](https://bun.com/blog/bun-in-rust): 5.9B uncached input + 72B cached input = 77.9B total input,
and 690M output. Its 92.4% cache share is rounded to 92% to
match the whole-percent control. Both sourced projects are
exceptional examples, not typical workloads; their reported
token totals do not validate this calculator’s energy
coefficients.
The estimator uses three contemporary public model proxies on modern NVIDIA Blackwell hardware. Its core equation is:
E = fresh_input × Jfresh + cached_input × Jcache + output ×
Joutput
Current per-token coefficients are modeled from public AgentX benchmarks on B300 hardware. The model assumes cache hits use ~56% of fresh-input energy and output costs ~10× per token.
Flash Qwen3.8-Flash-Next · 176B total / 6B active · B300-class AgentX proxy
Frontier DeepSeek V4 Pro · 1.6T total / 49B active · B300 AgentX proxy
SOTA Kimi K3 · 2.8T total / ~104B active · B300 AgentX proxy
Serving profile matters. Efficient serving batches more work and reduces energy per token; low-latency serving sacrifices utilization. The displayed range reflects different serving profiles, not a statistical confidence interval. Architecture, quantization, context length, cache placement, hardware generation, speculative decoding, and datacenter routing can move the true value outside it.
Carbon is calculated separately:
CO₂e = kWh × grid intensity. Changing the grid does
not change estimated inference energy.
Grid presets are scenarios, not routing claims. Presets use EPA eGRID2023 revision 2 annual total-output CO₂e rates, not marginal or consumption-based emissions. These generation averages do not adjust for electricity imports, transmission losses, or upstream fuel emissions. West Virginia is a coal-heavy benchmark; selecting it does not imply that a request was served there. Dedicated generation or contractual clean-energy purchases can differ from these averages.
Benchmark/source links: Qwen3.8 / InferenceX · DeepSeek V4 Pro / InferenceX · Kimi K3 / InferenceX · EPA eGRID2023 tables (XLSX).