The current methodology is at /en/methodology. This page keeps the list of published versions, newest first, with what each one changed. English translation of the French original (docs/methodology-versions.md).
v4.35-methodology-2026-11: measured Microsoft 365 Copilot actions. When the organisation connects the Microsoft 365 Copilot usage report (Microsoft Graph), the number of actions per licence measured over the last 28 days replaces the action assumption of the licence estimate, with the same value in all three sets. Tokens per action take their plausible range: 1,500 to 25,000 input, 100 to 1,500 output. The high set goes from about 30 to about 16 times the low one on the default model. With no report connected, nothing changes. No usage row is written.v4.34-methodology-2026-11: estimate from licences. New subsection "Estimate from licences" in "Source coverage": Claude Code, Microsoft 365 Copilot and chat (claude.ai, ChatGPT) are estimated from the declared plan and licence count, with three sourced sets of assumptions (low, central, high), a result given as a range rounded to one significant figure, on a line of its own in the first review, outside the measured totals and the ESRS export. GitHub Copilot, Cursor and Mistral Vibe return the cost of their licences only. The coverage table moves claude.ai chat (Team, Pro, Max, Ultra) and ChatGPT chat (Business, Enterprise, Plus) from "Out of reach" to "Estimated", path "Declared licences"; Claude Cowork on a Team plan and OpenAI embeddings stay out of reach. No factor, price or calculation range changes.v4.33-methodology-2026-11:gpt-6.2-solenters the catalogue (2 $ input, 10 $ output per Mtok, cache read 0.1, cache write 2.5), released by OpenAI on 2026-11-19 and absent from EcoLogits 0.11.3. Energies by the hand-applied price-proxy rule: gpt-6.1-sol × 1.000 (253 / 1515 Wh/Mtok), confidencelow. No other figure moves.v4.32-methodology-2026-11: the price-proxy anchor depends on the day of the usage. A row dated day D only takes as anchor a sheet that is eligible on that day. The catalogue robot dates that eligibility (anchorFrom) to the day it reads the model, never before its release: an addition no longer moves the anchor of a usage dated before the reading. A usage dated between the reading and the day the sheet enters the catalogue can still change anchor at the recompute. When the anchor is missing from the OpenRouter price snapshot, its price is read on the day of the usage. Sheets added before the rule carry no date and stay eligible at any date. Without a usage date (calculator, sheets), nothing changes. No figure moves.v4.31-methodology-2026-11: Llopis is cited in its published version, Sustainable Production and Consumption 70 (2026), of which arXiv version 4 (2026-11-18) is the accepted manuscript. The values we take from it do not change; the training term is in Table 5 of that version. "Claims that circulate wrong" describes Tier 3 as a customer-specific carbon figure supplied by the provider, and takes the finding of the published paper: no major provider yet offers that figure with independent assurance, so Tier 3 is empty. No factor, price or calculation range changes.v4.30-methodology-2026-11:gpt-6.2-astraenters the catalogue (8 $ input, 40 $ output per Mtok, cache read 0.4, cache write 10), released by OpenAI on 2026-11-18 and absent from EcoLogits 0.11.3. Energies by the hand-applied price-proxy rule: gpt-6-astra × 0.800 (2080 / 12482 Wh/Mtok), confidencelow. No other figure moves.v4.29-methodology-2026-11:gemini-2.5-proandgemini-2.5-flash, deprecated by Google on 2026-11-02 and routed togemini-3.8-flashandgemini-3.5-flash-litefrom 2026-12-01, becomelegacywith no shutdown date: they leave the current range of /models, the calculator and the comparisons, their sheets stay for past usage. The reference Google model of the comparisons becomesgemini-4-argon. No factor or price moves.v4.28-methodology-2026-11:gemini-4-hydrogenenters the catalogue (1.5 $ input, 7.5 $ output per Mtok, cache read 0.075), released by Google on 2026-11-12 and absent from EcoLogits 0.11.3. Energies by the v4.27 sibling rule: the EcoLogits 0.11.3 entry ofgemini-3.8-flash, taken as is (681 / 4,084 Wh/Mtok), confidencelow.gemini-3.5-flashandgemini-3.7-flash, routed by Google since 2026-10-08, becomelegacy: their sheets stay for past usage, their factors do not move. No other figure moves.v4.27-methodology-2026-11: three sheets lose their price multiplier, Claude confidence revised. The three energies that drop had one thing in common: an estimate borrowed from another model, then multiplied by a price ratio. Each layer inflated the result without measuring anything. v4.27 removes the multiplier and replaces the borrowed estimate with data closer to the model. Output energy in Wh/Mtok, before → after. Gemini 4 Argon, 44,588 → 11,881 (input 7,431 → 1,980). Before: the energy EcoLogits assumes for Gemini 2.5 Pro (17,625), the largest model it ascribes to Google, multiplied by 2.53 because Argon costs more. After: the energy EcoLogits gives Gemini 3.1 Pro, with no multiplier. The multiplier accounts for 2.53 of the gap, the assumed size of the reference model for the rest (2,000 against 1,200 billion parameters). Two independent calculations land lower: 10,891 by the price anchored on Gemini 3.8 Flash, 8,030 to 8,950 by measured speed against Claude Opus 5.5; we keep the highest value. DeepSeek V4 Pro, 36,335 → 6,463 (6,056 → 1,077). Before: Jegham's estimate for R1, multiplied by 2.08. It was already flagged as the weakest of the catalogue: the line forced through the origin absorbs the fixed cost of each reasoning request and inflates the slope by about 2.7. After: the EcoLogits formula applied to the architecture DeepSeek publishes (1,600 billion parameters of which 49 active) and the throughput it publishes for its service (20 tokens/s). The multiplier and the fit account for the whole gap between them (2.08 × 2.70). DeepSeek V4 Flash, 9,705 → 2,630 (1,618 → 438). Before: Jegham's estimate for DeepSeek V3, multiplied by 1.40, while this name serves V4.1 Flash since September. After: the same formula applied to the V4.1 Flash architecture, published on Hugging Face (552 billion parameters, 16 active to generate). The multiplier accounts for 1.40 of the gap, the rest separates a batch-8 calculation on H800 from a batch-64 one. New rule in "Models outside the catalogue": a sheet EcoLogits does not describe applies its formula when the architecture and a throughput are published, otherwise takes the entry of a sibling model, and the price proxy comes last, never anchored on a retired or routed model. Confidence: Sonnet goes fromhightomedium(the estimate is for Claude 3.7 Sonnet, not the model served), Haiku frommediumtolow(bounds of 0.07 to 0.57 times Sonnet). No Claude family ishighany more. New limit "Comparing two providers": two sources (Jegham for Claude, EcoLogits for the others) and two input conventions (1/21 against 1/6). What moves: energy, CO₂e and water of these three sheets (recompute required) and the confidence levels shown for Sonnet and Haiku. No Claude factor and no price moves. Each new value replays from the repo (tests/unit/v427-replay.test.ts).v4.26-methodology-2026-11: literature review re-read at the source, training, levers, where inference runs. No factor, price or calculation range changes. New subsections "Training: excluded from the calculation" (what Watershed and Llopis count, and the gap it makes with our figures) and "Mapping to the Watershed and Llopis tiers". "USD cost" gives the gap with the billed cost measured by Weinberger and Hozez: 8.7% of the cost of 2,848 Claude Code runs attaches to none of the four token classes. The "Deck recommendations" section becomes "Levers" and describes the catalogue; new section "Where inference runs". Librarian enters "Cache energy" in support of 0.08 (measured ratios of 4.5 to 8.4% in the agentic regime). Cited figures corrected after re-reading: the Mooncake cache rate goes from 34 to 55% (30 to 50% before); the values of 153.85 and 4.17M attributed to Bai et al. are not in the text of the paper and are withdrawn; the 107 → 242 mJ is pinned to version 1 of its source; the 1.8 L/kWh of Grubert and Sanders is flagged as not verified in the article; the ADEME-Arcep split is about the carbon footprint. The host server of the embodied constant is replayed on BoaviztAPI 2.6.0 (p5.48xlargeinstance, 5,733 kg excluding GPUs): the constant stays 49.v4.25-methodology-2026-11: limit of the CLI/API reconciliation on cost. "Source coverage" states that the adjustment removing the overlap is priced at the family price, while both sources carry the raw model id: for Claude Fable 5.1 since 2026-11-09, 20% of the overlap's cost stays counted twice, and the gap is given for the other models with their own price. Source coverage: the table is generated from the matrix in the code (lib/sources/coverage.ts), with a Level column (measured, estimated, not plugged in yet on request, out of reach) and rows for the workstation CLI, imported exports and subscriptions read by the CLI. TraceLab's first author is corrected (Zhu K., not Zhu J.). No factor, price or equation changes.v4.24-methodology-2026-11:gpt-6.1-terraenters the catalogue (1 $ input, 5 $ output per Mtok, cache read 0.05, cache write 1.25), released by OpenAI on 2026-10-26 and absent from EcoLogits 0.11.3. Energies by the hand-applied price-proxy rule: gpt-5.4-mini × 1.217 (87 / 523 Wh/Mtok), confidencelow. No other figure moves.v4.23-methodology-2026-11: Claude Fable 5.2 becomes the Fable family price. Claude Fable 5.2 and Claude Mythos 5.2, released on 2026-11-09, are at $8 input and $40 output per Mtok, cache write $10 (5 minutes) and $16 (1 hour), cache read $0.20. Claude Code'sfablealias resolves to Fable 5.2 from its version 2.1.321. Anthropic's price list, release notes and model-config page were re-read on 2026-11-10. The Fable family bills at that price from 2026-11-09, 20% less than before, and earlier rows keep their price. Fable 5.1 and Mythos 5.1, still sold at $10 / $50 (read 0.25), keep that price when the row carries their identifier, like Fable 5 and Mythos 5 (read $1).claude-fable-5-2andclaude-mythos-5-2resolve to the family. Two limits are written in "USD cost": a row stored at family level after 9 November is 20% too low when the session still ran Fable 5.1 or Fable 5, and fast mode ($16 / $80) is not modelled. What moves: the cost of Fable family rows dated 2026-11-09 or later and of rows carrying theclaude-fable-5-2orclaude-mythos-5-2identifier falls by 20% (production recompute required). No energy, CO₂e or water moves: Fable's energy stays extrapolated at 2× Opus, a ratio the new prices keep (8 / 40 against 4 / 20).v4.22-methodology-2026-11: France electricity, Ember 2025 data. The France CIF of the Mistral sheets goes from 0.056 to 0.0412 kgCO₂e/kWh: the 2025 value of Ember's yearly electricity data (41.2 gCO₂e/kWh, file dated 2026-11-03, CC BY 4.0), life-cycle, grid losses excluded. The former 0.056 was attributed to "Ember 2024" and matched no year of the revised series (2023: 53.3 g, 2024: 40.3 g). The CO₂e of the Mistral sheets falls by 15% (per million generated tokens, before → after): Small 4 23.2 → 19.7 g, Medium 3.5 54.9 → 46.7 g, Large 3 273 → 232 g, Large 4 302 → 257 g, Large 2 (retired) 58.7 → 49.8 g. The comparison with Mistral's LCA is rewritten: our parameters give 46 mL of water per gram of CO₂e against 39 in the LCA, instead of 39 with the former CIF. Anthropic's $200 monthly plan is called Ultra (formerly Max 20x): the "A subscription would cost less than usage" lever names it so, threshold unchanged. What moves: the CO₂e of the Mistral sheets. Water, energy and prices do not move, nor does any other provider. No Mistral usage row in the database at this date: nothing to recompute.v4.21-methodology-2026-10: workflow agents counted. New paragraph "Transcripts of a session" in "Token counting and deduplication": a session adds up its main transcript, those of its subagents and those of the agents its workflows start, which the CLI counts since its version 0.4.1 (deployed on 2026-10-19). A workflow'sjournal.jsonlis not read: the transcripts examined carry no token counter in it. The "Workflow agents" limit says which sessions remain undercounted: those pushed by an older CLI and not re-sent (over 30 days old, transcript older than the machine's enrollment, machine not updated). "Deck recommendations": the seat card is called "Flat seats", and the "Options to cut back" block no longer shows once the organisation has a validated action plan. No factor, price or equation changes.v4.20-methodology-2026-10: EcoLogits 0.11.3, monthly ECB rate, Gemini 4 Argon. Every OpenAI, Google and Mistral sheet that EcoLogits 0.11.3 (released on 2026-10-14) describes takes its energy from it, at the midpoint of the active-parameter range, input at one sixth of output, confidence unchanged except for the four retired models, which move tolow. Output in Wh/Mtok, before → after. Sheets estimated until now by the price proxy, by copying a neighbouring model or with a borrowed throughput: gpt-5.6-sol 5,366 → 5,144, gpt-5.6-terra 1,766 → 1,676, gpt-5.6-luna 141 → 589, gpt-6-astra 9,797 → 15,602, gpt-6-sol 1,499 → 1,473, gpt-6.1-sol 1,499 → 1,515, gpt-6-luna 53 → 110; gemini-3.5-flash 6,444 → 4,589, gemini-3.6-flash 6,239 → 3,991, gemini-3.7-flash 6,239 → 4,096, gemini-3.8-flash 6,239 → 4,084, gemini-3.5-flash-lite 1,611 → 335, gemini-3.1-flash-lite 1,139 → 340; Mistral Large 3 2,546 → 2,396 (mistral-large-2512entry, 59.2 tokens/s), Large 4 2,746 → 2,653 (published throughput of 51 tokens/s), Medium 3.5 508 → 482 (a 0.10.2 sheet, replayed with both Large models). Sheets pinned on 0.10.2, or on the 0.11.1 tag with the samemodels.json: gpt-5.5 5,366 → 5,067, gpt-5.4 1,766 → 1,741, gpt-5.4-mini 441 → 430, gpt-5.4-nano 141 → 136, gpt-4o 1,676 → 1,591, gpt-4o-mini 71 → 59, gpt-4.1 1,536 → 1,419, gpt-4.1-mini 329 → 368, o3-mini 55 unchanged; retired models that EcoLogits describes and that still sat on Jegham v6 Table 4: o4-mini 5,167 → 58, gpt-4.1-nano 502 → 62, gpt-4-turbo 5,183 → 5,410, o1 3,742 → 1,544,lowconfidence for all four (o3, absent from EcoLogits, stays on Jegham); gemini-2.5-pro 17,636 → 17,625, gemini-2.5-flash 1,611 → 1,602; Mistral Small 4 207 → 204 (mistral-small-2603entry, 110.2 tokens/s, since themistral-small-latestentry no longer has a throughput). gpt-4.1-mini sat at the midpoint of its 0.10.2 energy range; it moves to the midpoint of the parameter range, like every other sheet, hence the rise. gemini-2.0-flash and mistral-large-2411, which left EcoLogits, keep their 0.10.2 value. New section "Cross-check against EcoLogits": the Claude factors stay on Jegham v6, against the EcoLogits 0.11.3 middle of the latest generation (Sonnet ×1.97, Opus +28%, Fable -23%, Haiku ×7). EcoLogits 0.11.3 also covers Haiku 5.5: ratio 0.071 to Sonnet 5.5, geometric middle 0.20 with the speed bound, the 0.25 factor stays.gemini-4-argonenters the catalogue (GA on 2026-10-13): $2 / $10 per Mtok, cache read 0.10, through 30 November 2026, then $4 / $20 / $0.20. Absent from EcoLogits 0.11.3, it is estimated by the price proxy anchored on gemini-2.5-pro, computed on the 1 December price (×2.53): 7,431 / 44,588 Wh/Mtok,lowconfidence. The catalogue robot now looks the model up in EcoLogits before the price proxy. At ingestion too, a model without a sheet that EcoLogits describes takes its energy before the price proxy, through a table generated from the same release (pnpm ecologits:table,lowconfidence):gpt-5-nanomoves from 7 / 44 to 10 / 58 Wh/Mtok,gemini-2.5-flash-litefrom 269 / 1,611 to 9 / 56, andgpt-5,gpt-5-mini,gpt-5.1orgpt-5.2leave the proxy too. An id a later release no longer describes keeps its last value, with its version; a dated snapshot of a sheet keeps the sheet. The release read is now written in one place only,lib/providers/ecologits-release.json, and the methodology names it. A table at the top of the methodology gives the energy source and version of each model, and the frozen assessment states its own, read from its snapshot. Euros: the dashboard and the calculator convert at the latest published ECB monthly average (September 2026, €0.8686 per $1 instead of 0.8729 as of 22 June, i.e. -0.5% on euro amounts); the frozen assessment takes the average of its period's month, or the latest published one while the ECB has not released that month. What moves: energy, CO₂e and water of the thirty-one sheets whose value changes and of the models without a sheet that leave the proxy (recompute required), and the euro amounts shown. No dollar cost moves, no Claude factor either.v4.19-methodology-2026-10: imported data. New section "Imported data": an export in the TokenClimate format enters with the level of its measure (measuredfor tokens,dollars,counts), dollars and counts converted with the published calculator profiles, 30,000 tokens per agent request written as a derivation, activity uncertainty factors (×1.1, ×1.6, ×3, ×5) written but not yet propagated, teams under 5 people and tools already measured refused, imported rows removed when a connector later measures the same tool, replacement by tool and team. No factor, price or equation changes.v4.18-methodology-2026-10: history backfilled at connection. "Source coverage" states how far back the Anthropic admin key goes: up to 90 days requested on the usage report and on Claude Code Analytics, depending on what the API serves, with the same deduplication key as the regular sync and the reconciliation with the CLI extended to the backfilled days. No factor, price or equation changes.v4.17-methodology-2026-10: frozen assessment. New section "Frozen assessment": what a freeze holds, cost recomputed at the rate of each usage's day and at the recorded ECB rate, range of the three published ranges applied to every provider (the ±30 % CIF spread also covers non-Anthropic models), four sensitivities at 10 %, replaced versions. No factor, price or equation moves.v4.16-methodology-2026-10: 1-hour cache writes billed at 2× input, sessions split by model. A row now carries the part written to the 1-hour cache (cache_creation_1h_tokens) when its source measures it, billed at its model's 1-hour price: Admin API usage report for the hours pulled after the deployment, CLI from 0.4.0. On those rows the 1-hour part costs 1.6 times more than before (2 / 1.25). Nothing is recomputed: an unknown part (null, earlier rows and Claude Code Analytics) stays at the 5-minute rate, and the "billed versus estimated" card mixes the two conventions for 30 days. The server accepts the CLI's payload v2, one row per session and per model under the raw identifier, and counts sessions bysession_weight(section "Session counting"). Session counts do not move, except adjustment rows, which no longer count any session. In the ESRS export,events_countsums this weight anddata_qualitybecomes a share of rows. With CLI 0.4.0, a subagent under a dated identifier (Haiku) comes out atlowconfidence, like Admin rows since v4.6: thelowshare ofdata_qualitywill rise. No energy, CO₂e or water moves.v4.15-methodology-2026-10: Haiku goes from 0.5 to 0.25 times Sonnet, like the claude-carbon plugin 1.12.0 (10 / 207 gCO₂e/Mtok instead of 20 / 413). Haiku energy 61 / 1,262 → 30.6 / 633 Wh/Mtok. The value sits between two public bounds: EcoLogits 0.11.2, which puts Haiku 4.5 at about 0.076 times Sonnet 5.5 per output token, and the output speed read by Artificial Analysis on Anthropic's API (Haiku 5.5 237.3 tokens/s, Sonnet 5.5 134.9), which caps the ratio at about 0.57. Geometric middle 0.21, rounded up. The price ratio (0.05) is not used. Details in "The Haiku / Sonnet ratio". What moves: the energy, CO₂e and water of every Haiku row drop by 49.8%; every other row and every cost stay identical. For one million Haiku output tokens: 238 gCO₂e instead of 475. Production recompute required; the AnthropicfactorVersionbecomestokenclimate-v3-2026-10.v4.14-methodology-2026-10: six Google models not covered by EcoLogits enter the catalogue:gemini-3.8-flash,gemini-3.7-flashandgemini-3.6-flash(introductory price through December 31, 2026: $0.75 input, $3.75 output per Mtok, cache read 0.075, then double),gemini-3.5-flash(1.50 / 9, cache read 0.15),gemini-3.5-flash-lite(0.30 / 2.50, cache read 0.03) andgemini-3.1-flash-lite(0.25 / 1.50, cache read 0.025). Energies by the hand-applied price-proxy rule, with two rules now written for every hand-applied proxy: anchor in the same line (every Flash and Flash-Lite on gemini-2.5-flash, never on gemini-2.5-pro) and ratio on the list price, never on a promotional price. gemini-2.5-flash × 4 for 3.5 Flash (ratio 4.243 bounded to 4, 1,074 / 6,444 Wh/Mtok), × 3.87 for 3.6, 3.7 and 3.8 Flash on their list price of 1.50 / 7.50 (1,040 / 6,239), the energy of gemini-2.5-flash, sold at the same price, for 3.5 Flash-Lite (269 / 1,611), × 0.71 for 3.1 Flash-Lite (190 / 1,139), confidencelow. No other figure moves. Note of 2026-10-17 (v4.20): "not covered by EcoLogits" was inaccurate at that date. gemini-3.5-flash and gemini-3.1-flash-lite were already in EcoLogits 0.10.2, the other four in 0.11.2, released on 2026-09-29.v4.13-methodology-2026-10:mistral-large-4-0(Mistral Large 4, version 26.10) enters the catalogue, released by Mistral on 2026-10-06 and not covered by EcoLogits. Energies by replaying the EcoLogits parametric model on the published architecture (52 billion active parameters, 1,050 billion total), with the Large 3 throughput: 458 / 2,746 Wh/Mtok, confidencemedium. Promotional price billed on 2026-10-09: $0.68 input, $2.09 output per Mtok, cache read 0.07.mistral-large-lateststays Large 3. Same day, Anthropic and DeepSeek price lists re-read. Claude Opus 5.5 (launched 2026-09-22), Claude Sonnet 5.5 (launched 2026-09-28) and Claude Haiku 5.5 (launched 2026-10-07) are what Claude Code'sopus,sonnetandhaikualiases resolve to: they become the family prices, Opus from 2026-09-22 ($4 / $20, read 0.20, about 20% lower), Sonnet and Haiku from 2026-10-07. The Haiku family moves to $0.10 / $0.50 (read 0.01, the price list for prompts up to 100,000 tokens), ten times lower; the Sonnet family cache read drops from 0.20 to 0.10, the day Anthropic cut it on Sonnet 5.5. Rows before that date keep their price. Claude Opus 5, 4.8, 4.7, 4.6 and 4.5 ($5 / $25, still sold), Claude Sonnet 5 (read 0.20) and Claude Haiku 4.5 ($1 / $5) keep their own price as exact-model entries. The Haiku 5.5 tier above 100,000 tokens (5x dearer) cannot be modelled on aggregated rows.deepseek-v4-flashserves V4.1 Flash since 2026-09-10 and moves to $0.30 / $1.20 (read 0.006) from that date, the old price staying valid before. Its energy is not recomputed. No energy moves.v4.12-methodology-2026-10: overprovisioning is named. The capacity kept in reserve for peaks and failures now appears among the exclusions in "System boundaries", with its order of magnitude (10% of the energy of a median prompt at Google), and the "Idle capacity" assumption carries that name. No figure moves.v4.11-methodology-2026-09:gpt-6.1-solenters the catalogue (2 $ input, 10 $ output per Mtok, cache read 0.1, cache write 2.5), released by OpenAI on 2026-09-29 and not covered by EcoLogits. Energies by the hand-applied price-proxy rule: gpt-6-sol × 1.000 (250 / 1499 Wh/Mtok), confidencelow. No other figure moves. Note of 2026-10-17 (v4.20): "not covered by EcoLogits" was inaccurate at that date, gpt-6.1-sol was in EcoLogits 0.11.2, released on 2026-09-29.v4.10-methodology-2026-09: limits of the main source and scope, published. No energy, CO₂e, water or cost figure moves, no calculation range changes. Additions, all re-read at source: the batch of 8 set by Jegham v6 and its effect in its own appendix A (a batch of 16 lowers energy by about 43%), the "Large" hardware class assigned for lack of a published size and the host counted at 0.5, in the sensitivity note; the idle capacity excluded by Jegham, about 10% at Google; the Haiku / Sonnet ratio, whose 0.5 has no primary source (sensitivity range 0.2 to 0.6); the eGRID2024 reference points for the AWS subregions and the TVA (0.429), which the ±30% range does not cover; ADEME's 3-year amortisation (82 gCO₂e/kWh); model development and datacenter construction, excluded, with their order of magnitude. Corrections: the 273 kgCO₂e of ADEME's LCA are manufacturing, distribution and end of life, with no use phase; Jegham's Table 1 cites Electricity Maps as reference [44], with no weighting note, and the quote "mix-weighted across AWS inference regions", absent from v6, is replaced by the text of §4.1. Source coverage: the Anthropic Admin key attributes a person only on subscription seats, API-billed Claude Code stays per key except through the CLI, Team-plan seats remain to be checked on a real account, and the missing OpenAI row is added.v4.9-methodology-2026-09: the Bilan IA is retired. This self-serve product, which read an Anthropic admin key and returned a one-off report, is no longer offered. The passages that cited it now speak of the admin key, the report or the dashboard, without changing what they describe. The published uncertainty ranges live inlib/methodology/ranges.ts. No energy, CO₂e or water figure changes.v4.8-methodology-2026-09: ADPe annex removed. It only counted electricity, with the US electricity-mix factor from ADEME's Base IMPACTS (2011 vintage), applied without PUE, whereas hardware manufacturing makes up most of this indicator (ADEME GPU LCA, March 2026). The indicator will come back once the manufacturing share can be allocated to usage from a source. The two Sb eq figures quoted from Mistral's LCA are removed with it. No energy, CO₂e or water figure changes.v4.7-methodology-2026-09: hardware manufacturing re-sourced. The GPU term of the embodied constant (EMB) moves from the manufacturer's declaration (NVIDIA PCF for the HGX H100 baseboard: 1,312 kgCO₂e for eight GPUs, cradle to gate) to the life-cycle assessment of GPUs published by ADEME in March 2026 (Lees-Perasso et al., Hubblo, TND, TIDE, CNRS and INRIA, Table 44: 273 kgCO₂e per H100 SXM 80 GB, cradle to grave, manufacturing, transport and end of life included), i.e. 2,184 kgCO₂e per node. The host server, the 5-year lifetime, the power and the utilisation rate do not change. EMB goes from 44 to 49 gCO₂e/kWh IT, the published 22-66 range stays the same. For Anthropic, K_co2 goes from 0.37118 to 0.37618 gCO₂e/Wh and the embodied share from 11.9% to 13.0% of total CO₂e. What moves: the CO₂e of every row, by +1.35% on Anthropic and Meta, +1.15% on OpenAI, +1.19% on Google, +0.62% on DeepSeek and +4.6% on Mistral, whose low French grid mix leaves more weight to manufacturing. Production recompute required; the AnthropicfactorVersionbecomestokenclimate-v3-2026-09. Energy, water and cost do not move. The cross-check against Mistral's LCA goes from 41 to 39 mL of water per gram of CO₂e, against 39 published by Mistral.v4.6-methodology-2026-09:gpt-6-sol($2 / $10) andgpt-6-luna($0.10 / $0.50), released by OpenAI on 2026-09-22 and not in EcoLogits, enter the catalogue. Energies by the price-proxy rule applied by hand: gpt-4o × 0.89 for Sol (250 / 1,499 Wh/Mtok), gpt-4o-mini × 0.75 for Luna (9 / 53),lowconfidence. Astra prices re-read, unchanged. No other figure moves. Price by exact model (same version, 2026-09-23): a row carrying the raw identifier of a model with its own price is billed at that price (Opus 5.5 at $4 / $20, reads $0.20; Fable 5.1 and Mythos 5.1 reads $0.25; Fable 5 and Mythos 5 reads $1; Sonnet 4.6, 4.5, 4, 3.7 and 3.5 at $3 / $15), families keep their dated prices, the energy of these models is their family's. The Admin API syncs now store the raw identifier. CLI events stay at the session's main model. What moves: the cost of the affected raw-identifier rows (recompute required), no energy, no CO₂e, no water. GPT-5.6 Luna correction (same version, 2026-09-23): its energy was extrapolated from gpt-5.4 on a $6 output price read on 2026-07-19 (6/15 = 0.4×, i.e. 118 / 706 Wh/Mtok). The 2026-09-07 price check corrected that price to $0.2 / $1.2 without re-deriving the energy. Luna sits at gpt-5.4-nano's price ($0.2 / $1.25): it takes nano's energy, 24 / 141 Wh/Mtok, as Sol and Terra take the energy of the model at their price. Confidence stayslow. GPT-5.6 Luna's CO₂e and water drop by 80%, no other model moves.v4.5-methodology-2026-09: published after the content audit of 2026-09-07 (743 claims checked, 142 corrections, every surface). Anthropic prices dated per family (pricingHistory,pricingAt): Sonnet at $2 / $10 since 2026-06-30 (Sonnet 5 introductory price made permanent), Fable cache read at $0.25 since 2026-09-01 (Fable 5.1), $3 / $15 and $1 before; the USD cost of each row is that of its day's rate, at every ingestion point and inrecompute:co2; the "tier 1" mention is removed (tiers only set rate limits); declared limit: a family is billed at the price of its current default model, Sonnet 4.6 stays at 3 / 15 at Anthropic and cannot be told apart in the database. The displayed cost is in euros at the dated ECB rate oflib/currency.ts, written in the "USD cost" section and on the calculator tile. Third-party catalogue rewritten provider by provider: Google, gemini-2.5-pro (2,939 / 17,636 Wh/Mtok) and gemini-2.5-flash (269 / 1,611) from EcoLogits 0.10.2,lowconfidence, gemini-2.0-flash (shut down 2026-06-01) and gemini-1.5-*legacy; DeepSeek, deepseek-chat and deepseek-reasonerlegacy(shut down 2026-07-24), deepseek-v4-flash (1,618 / 9,705) and deepseek-v4-pro (6,056 / 36,335) by hand-applied price proxy,lowconfidence, peak prices of 2026-09-07; Meta, PUE 1.14 (Meta/AWS row of Jegham v6 Table 1) instead of the 1.2 placeholder, Llama energies 19 / 133 and 31 / 640 Wh/Mtok (+5%), CO₂e and water per Wh aligned on the AWS row, both Llamalegacy(Groq retirement 2026-08-16); Mistral, prices of 2026-09-07 (Large 3 0.5 / 1.5, Medium 3.5 1.5 / 7.5, Small 4 0.15 / 0.6), mistral-large-latest = Large 3 replayed on the published architecture (424 / 2,546,medium), new mistral-medium-latest sheet (85 / 508), Large 2 kept aslegacyunder mistral-large-2411 (retired 2026-05-31), LCA paragraph attached to that sheet; OpenAI, gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini are current again (still sold), o1, o3-mini, o4-mini, gpt-4-turbo, gpt-4.1-nano (shutdown 2026-10-23) and o3 (2026-12-11)legacywith a date, cache reads added;legacynow reads "retired or being retired by its provider". Price-proxy example replayed on deepseek-v3.2. Usage profiles: TraceLab counts 37 developers on the Claude Code column (43 was the Claude plus Codex total), cadence 1.9 session per developer per week (8.0 per month) instead of 1.6; profile input/output ratio 294; 95.5% prefix, 91.5% re-read, 95.8% hit rate on the Claude column; the private corpus median is compared to a TraceLab mean; the cadence is no longer shown on screen; the 184-session corpus is declared non-reproducible; three questions in simplified mode. Equivalences: car 142 gCO₂e/km (Impact CO₂, ADEME, 2025 modelling, combustion car manufacturing included, same factor as the OSS plugin) instead of a 120 "ADEME well-to-wheel" that matched no publication, displayed kilometres drop by 15%; shower requalified as "5-minute shower, about 60 L" with the ADEME page cited; the 0.09 gCO₂e of the Gemini prompt is presented as a derivation of the published 0.03 market-based. ESRS export: water split on-site withdrawal / off-site consumption (E3-4), IT and facility energy distinguished, data quality weighted by CO₂e, label without a source list, columns documented here. Deck: new "Deck recommendations" section (thresholds and cache assumption), manufacturing attributed to NVIDIA PCF + BoaviztAPI. Bilan report: cache-read factor range injected from the code (0.05-0.20), "prefill residual", "estimates from measured volumes", bibliography completed (Jegham five authors, AWS 2025, WRI 2020, Irminsul, Impact CO₂). Tokenizer caveat published under "Validity date of the energies": 4.7 and later models produce about 30% more tokens for the same text, a per-token factor calibrated before may overestimate by as much, no correction applied. Sourcing corrected: embodied 44 is attributed to Jegham nowhere and its derivation is published (7,012 kg over 43,800 h at 5.6 kW and 65%); Jegham et al. cited with its five authors; "measurements" becomes "estimates" wherever the source is Table 4; Artificial Analysis provides latency and throughput; Microsoft "estimates" 0.31 Wh (v2), the 111 Wh/Mtok is ML.ENERGY v3.0 on Qwen3-235B on B200, EcoLogits corrected its online calculator; Irminsul, MLA row averaged and MHA measured at 2,048 tokens; 1.3-51.8% from Solovyeva and Castor, 2K→10K tripling on Llama 3 70B; OpenAI bills cache reads at 0.1x on its current range; Mooncake 30 to 50%, 75-95% = CharacterAI and Kimi; 2.80 MtCO₂e = Amazon as a whole, not AWS; Google 0.03 market-based published, 0.09 location-based derived; a single EcoLogits repo name; Google PUE sourced to Elsworth et al.; 50/46/4 split sourced to the ADEME-Arcep update of January 2025; Colossus deal with SpaceX (formerly xAI), 2029 horizon reported by TechCrunch; AFNOR Spec 2314: min-best-max and versioned method requalified as own practices, location-based in a note to §4.3; the MiNumEco factsheet cites neither EcoLogits nor Article 53; gpt-6-astra "announced" on 2026-09-03; hardcoded "v3" removed from the body; Haiku 61 explained (rounded OSS factor); booking link present in both languages. What moves in figures: the cost of Sonnet rows after 30 June 2026 drops by about a third, that of Fable cache reads after 1 September by three quarters (production recompute required); Meta energies rise by 5% (no Meta row in the database); displayed kilometres are divided by 1.18. No Anthropic energy, CO₂e or water figure moves.v4.4-methodology-2026-09:gpt-6-astra(announced on 2026-09-03, limited preview then API opened in the following days) enters the catalogue, energies of gpt-5.5 × 1.83 by the price-proxy rule applied by hand,lowconfidence. gpt-5.6 prices re-read on 2026-09-07 (Sol $4 / $20 promotional, Terra $2 / $12, Luna $0.2 / $1.2, cache writes billed): only the USD costs of those three models change, and no OpenAI row exists in the database. "On water" rewritten: the gap between the 5.11 L/kWh EWIF and the literature's 1.8 L/kWh is an allocation convention (hydropower reservoir evaporation counted by WRI, excluded by Grubert and Sanders), not an error to fix, and 5.11 is indeed a consumption, declared as such by Jegham; the WRI guidance is dated 2020. Mistral's water sources corrected in the seed and on the sheet: the WUE of 0.18 is the AWS 2023 value reused for want of a disclosure (not Jegham's Table 1, which has no Mistral row), and the EWIF moves from 3.13 (US average) to 3.67 L/kWh, the France factor of Appendix 2 of the WRI 2020 guidance: water for Mistral models rises by 16%, and no Mistral row exists in the database. AWS's 5.11 is verified as reproducible on those same appendices (about 40% Northwest at 9.48 and 60% Virginia at 2.37). The calculator shows under water the share of on-site cooling, computed on the mix. New paragraph on Mistral's LCA (July 2025) under "Third-party providers": it provides none of our terms, its water/carbon ratio matches our France parameters within 3%, its per-response level is 44× ours, a perimeter gap published rather than smoothed over. No change to any equation or per-model energy; outside Mistral, not one CO₂e, water or energy figure moves.v4.3-methodology-2026-09: new section "Models outside the catalogue: default estimation". Step 1,pnpm catalog:proposeproposes a seed entry from EcoLogits 0.10.2 (human validation). Step 2, price proxy at ingestion: energies of the catalogued model of the same provider nearest in price, × the geometric mean of the input and output price ratios, clamped [0.1; 4], prices from the same list (public OpenRouter snapshot, stale at 120 days),co2_confidence = low,factor_version = price-proxy-openrouter-<month>. The family-regex fallback is removed for non-Anthropic providers. New T3 row "E_out price proxy" in the parameters table. No change for catalogued models; uncatalogued rows move from "not covered" to a T3 estimate or stay not covered (unseeded provider, missing or zero price), recompute traced.v4.2-methodology-2026-09: added OpenRouter to "Source coverage": management key, one row per day, person and model, attached to the real provider and priced with its factors on a strict resolution of the identifier. Declared limits: cache writes counted as input, BYOK rows skipped when the provider is already measured, no reconciliation with the CLI or the SDK. No change to equations, factors or prices: no existing figure moves.v4.1-methodology-2026-09: added the "Estimation per usage unit (simplified mode)" section: output tokens per session (96,900,000 / 2,676) and per exchange (150), plus TraceLab's measured cadence (1.6 session per developer per week) shown as a reference point. No cadence is published, unlike the per-task volume table removed in v3.7: the value is entered by the visitor and stays editable in the volume field, and only the tokens per unit are derived from already-sourced aggregates. The result is rounded to two significant digits and presented as T3, the dominant uncertainty coming from the declared cadence. No change to equations, factors, reference volumes or prices: at equal volume, no displayed figure moves.v4.0-methodology-2026-08: source over-qualification corrected. Jegham v6 is not a measurement campaign: its section 4.2, "Per-Query Energy Consumption Estimation", describes a Monte Carlo simulation over nameplate power, an assumed utilisation and a batch size fixed at 8, with Artificial Analysis latency as the only empirical input. Every row of Table 4 is a modelled estimate. The product wrote "direct measurement" on the/modelssheets, in thehightier gloss, in the seed notes published as JSON-LD and in the cockpit tooltips: all of that copy is reworded, across the five report locales. The confidence scale is redefined and now ranks estimates against each other. DeepSeek moves fromhightomedium, its fit being output-only under an NNLS constraint with an input convention, i.e. the case that already earnsmediumat OpenAI. Two newly published caveats: DeepSeek is the most dispersed row of the panel, and thedeepseek-reasonerfit is mis-specified (centred R² -2.76, slope inflated ~2.7x by forcing it through the origin), the figure being kept and tracked rather than silently patched. Embodied 44 gCO₂e/kWh is no longer attributed to Jegham indeepseek.jsonandmeta.json: Table 1 has no embodied column and the paper excludes scope 3, the constant comes from NVIDIA PCF plus BoaviztAPI. Removed the version hardcoded in the body, which announcedv3.7two versions later. No change to any equation, energy, infra parameter or price: not one CO₂e, water or energy figure moves.v3.9-methodology-2026-08: Anthropic confidence levels corrected. Opus and Haiku move fromhightomedium: their energy is extrapolated from Sonnet (×2 and ×0.5), only the Sonnet class is measured, and the registry published them ashighpurely through its default value, for want of an explicit declaration in the seed. All four families now declare their confidence explicitly (Sonnethigh, Opus and Haikumedium, Fablelowunchanged). Alongside it, the infra source attribution on the/modelssheets is corrected: Anthropic's and Meta's PUE and on-site WUE are credited to the first-party AWS disclosure rather than to Jegham, which v3.8 had already recorded here without propagating it to the public copy. No change to any equation, energy, infra parameter or price: not one CO₂e, water or energy figure moves, only the confidence labels and two sourcing sentences.v3.8-methodology-2026-08: on-site WUE 0.18 → 0.12 L/kWh, re-sourced on the first-party AWS disclosure (2025 vintage) instead of the Amazon 2023 report; PUE 1.14 unchanged but re-sourced the same way. The parameter block becomes explicitly composite. Tier reclassification: CIF and EWIF T1 → T2, cache_read_factor T3 → T2, E_in/E_out ratio T2 → T3. Published range of cache_read_factor widened from 0.05-0.15 to 0.05-0.20. "Cache energy" rewritten around the Irminsul measurement, including the finding that the term does not have the right shape. "On water" (withdrawal plus consumption) and "Assumed location" (Colossus 1, Memphis) rewritten. The 5-year amortisation is now sourced. New sections "Validity date of the energies", "Sensitivity note: the overestimation charge" and "Claims that circulate wrong". Three caveats added. No change to the equations, to per-model energies or to prices; CO₂e does not move for any model, water drops 1.0% on the Anthropic side.v3.7-methodology-2026-08: removal of the "Use cases (one occurrence)" section and its table on the public calculator: per-task volumes judged too imprecise for a public surface (mostly T3). Usage profiles remain; no change to equations, factors or prices.v3.6-methodology-2026-08: added the "Usage profiles of the public calculator" section (reference volumes per token class, token-based use cases, sources and tiers); no change to equations, factors or prices.v3.5-methodology-2026-07: added the "Source coverage" section (surfaces measured by the admin key, claude.ai chat and Cowork out of scope on Team plans, Enterprise case); no calculation change.v3.4-methodology-2026-07: current OpenAI range enters the catalogue (gpt-5.6, gpt-5.5, gpt-5.4, sourced from EcoLogits 0.11.1 or price-extrapolated withlowconfidence); the ten previous OpenAI models becomelegacy, factors kept to quantify past usage.v3.3-methodology-2026-07: added a paragraph explicitly linking EcoLogits (pinned release) to the MiNumEco factsheet of November 2025, cited by the AI usage report wording; no methodological change.v3.2-methodology-2026-07: DeepSeek recalibration (direct Jegham v6 DS-native measurements) and OpenAI (Table 1 verbatim), Meta provider added.v3.1-methodology-2026-07: Anthropic energies recalibrated on the Jegham v6 3-point OLS fit (Sonnet 119/2525 Wh/Mtok, previously 580/3480); the Anthropic input/output ratio becomes a measured value (~1/21) instead of the 1/6 convention; Opus moves from 3x to 2x Sonnet. Prices, infra parameters, cache_read_factor and formulas unchanged.v3.0-methodology-2026-06: energy-first switch (Wh/Mtok + infra parameters), water and embodied added.