The repository accompanying "So, What's the Carbon Cost of Using AI?" A Bounded Estimation Framework for Calculating the Per-Token Carbon Intensity of Large Language Model Inference Under Limited Disclosure, Using the Worked Example of OpenAI's GPT-5 Model Family (Manktelow, 2026): the collection engine, the parameter tables, and the scripts that reproduce the paper's tables.
- Paper: read it on Zenodo. Text v1.0.1 of 1 October 2026, DOI 10.5281/zenodo.23084616; it revises the wording of v1.0.0 and changes no table. Also submitted to arXiv on 25 September 2026; identifier to follow (see CHANGELOG.md). The paper is not in this repository.
- Archive: concept DOI 10.5281/zenodo.22962475 (always resolves to the latest version); this release, v1.0.0: version DOI 10.5281/zenodo.22962476
v1.0.0 — version DOI 10.5281/zenodo.22962476 — see CHANGELOG.md. Cite the version DOI to fix a specific calculation.
engine/ Collection engine (paper Appendix B): benchmark, drivers, scheduled-task prompts
manifests/ Tracked-model manifests, one per provider, with entry/retirement dates
calibration/ Tokenizer calibration (o200k_base vs native); the canonical basis statement
parameters/ Everything Appendix A tabulates, with per-row source and access date
data/ Frozen corpus and checksum manifest, Zenodo fetch script, derived workbooks
reproduce/ corpus_summary.py (corpus → workbook inputs) and paper_tables.py (workbook → tables, paper check); expected outputs committed
briefs/ Per-brief data folders and reproduce scripts (none published yet)
tools/api-harness/ ic_api.py, a standard-library command-line client for the public API (sign-up, key, every endpoint)
BASIS_OF_PREPARATION.md Boundary, method, conventions, data-quality indicators — for inventory preparers and assurance providers
VERIFICATION.md Chain of evidence, verification steps and record, known discrepancies, reproduction-statement template
The tables regenerate from committed parameters and workbooks — no corpus download and no API key required:
git clone https://github.com/InferenceCarbon/inferencecarbon-framework.git && cd inferencecarbon-framework
pip install -r reproduce/requirements.txt
python reproduce/paper_tables.py --workbook data/workbooks/OpenAIModelCalculations_v1.0.0.xlsx --out out/tables
diff -r out/tables reproduce/expected/tablesRuntime: about five seconds. To go one step further back — from the frozen corpus to the workbook's
measured inputs — run reproduce/corpus_summary.py (about ten seconds), which can also regenerate the
workbook's input sheets outright (--write-workbook); reproduce/formula_audit.py checks that no
numeric cell on a calculation sheet is an undeclared constant. See reproduce/README.md.
The full verification procedure, and what has and has not been independently verified, is in
VERIFICATION.md.
Not required to reproduce the tables. The engine reads each provider's API key from local key files that are never committed (names and per-provider budget caps in engine/owned_benchmark/providers.json); a full campaign run makes real API calls against every tracked provider up to those caps. See engine/README.md.
The raw per-request corpus ships in this repository as
data/InferenceCarbon_corpus_frozen_20260822.zip (8.9 MB; 14,733 request
records, 24 Jun – 22 Aug 2026, of which 52 are synthetic pipeline-test records
that the reproduction scripts exclude) and is mirrored in the Zenodo data deposit.
Both copies verify against the same SHA-256 in data/manifest.csv:
(cd data && shasum -a 256 -c InferenceCarbon_corpus_frozen_20260822.sha256)To fetch the Zenodo deposit and verify the corpus inside it instead:
python data/fetch_corpus.pyThird-party benchmark series used for the cross-checks in section 5.3 of the paper are not reproduced here. The paper states what was compared and over which window; a replicator needs their own access to those series.
This repository is the frozen, per-release snapshot from which the release's published tables regenerate; it exists to be recomputed, subsetted and built upon.
- Code — Apache 2.0: the same openness as MIT with an explicit patent grant and no implied trademark rights.
- Summary workbooks and parameter tables — CC BY-NC 4.0, with express permissions: non-commercial use (recomputation, subsetting, teaching, research, publication) needs no permission; use by regulators, standards bodies and public authorities, and verification or assurance of figures derived from the tables, including by commercial auditors, is expressly permitted. Commercial use requires a licence (info@inferencecarbon.ai).
- Frozen raw measurement corpus — CC BY-NC 4.0, with the same express permissions: regulators, standards bodies and public authorities, and verification or assurance of figures derived from it, including by commercial auditors. What the non-commercial term withholds is the assembly of deposited corpora into a commercial data product; commercial use requires a licence.
- No trademark rights are granted. The paper text is licensed separately (CC BY-NC-ND 4.0 in its preprint versions). Paper Section 8.2 and Appendix H.1 are canonical.
The live operational stack (multi-vantage scheduling, model manifests and quality-assurance layers) and the live measurement database served through the InferenceCarbon.ai API are not published here; the API's terms of service are the operative license for that service (methodology@inferencecarbon.ai).
The current, continuously updated figures (every published model and reasoning setting, with bounds and confidence; OpenAI text models at v1.0.0, other providers as each companion brief is published) are served by the InferenceCarbon.ai API: free for academic research, regulators and public bodies, personal use and a 30-day evaluation, with a key from https://www.inferencecarbon.ai/api. The same figures are shown at https://www.inferencecarbon.ai/models, and the calculator at https://www.inferencecarbon.ai/calculator applies them to a query. To challenge a figure or the method: https://www.inferencecarbon.ai/review-this-paper.
See CITATION.cff (GitHub renders it as "Cite this repository"). Version DOI to fix a calculation; concept DOI for the framework as a whole; the paper's own DOI, or its arXiv identifier once announced, for the text. A headline figure travels with its plausibility band and its confidence label — quoting the central value alone misrepresents the finding.
methodology@inferencecarbon.ai. If you are a provider: an audited per-model disclosure of energy per token would retire these estimates, and we would welcome that outcome.
Not a measurement system. The figures are bounded estimates, at Low confidence on the headline, and the Low/Central/High envelopes are part of the finding, not decoration around it.