What's new
The decode-context term is now captured. Every session stores output_context_sum, the sum over its assistant messages of output_tokens x (input + cache_write + cache_read): the KV-cache size each generated token re-reads. Measured on open models, decode energy per token rises x2.3 from 1k to 16k context, and the median Claude Code message here carries 160k; nothing is measured beyond 16k, so no slope can be defended yet and no factor changes, CO2 and cost are identical to 1.6.0. The column exists so a context-dependent term can be applied to history once one is calibrated: transcripts are purged after 30 days, the column is what survives. Both writers add it to an existing database; scripts/backfill.sh fills it on earlier rows whose transcript is still on disk.
METHODOLOGY gains "Where the estimate stands (September 2026)": each parameter against what was published between April and September 2026 (Oviedo et al. in Joule, the Watershed framework, four HotCarbon 2026 papers, the Amazon and Google 2026 reports), the swing it produces on 30 days of real transcripts, and the decision. Short version: the output factor sits 2-3x above production-conditions figures and the input factor 2.6x below Watershed's, the unmodelled context term outweighs both, and the displayed figure is more likely low than high. Next review March 2027.
Also
CITATION.cfffollows the release version from now on (it was stuck at 1.1.1);check-versions.shfails when it drifts.- Social preview image (
docs/social-preview.gif) rendered from the demo; the demo prompt is an ordinary task now, and the model shown is Opus 5.
Upgrade
claude plugin update claude-carbon (or re-run the installer), restart Claude Code, then run scripts/backfill.sh once to fill the new column for the last 30 days. No recompute needed.