Ranking methodology and data
Score calculation, criteria and scales, sensitivity to weights, calculation rules and sources. The data can be downloaded under a CC BY 4.0 licence.
Composition of the score
Performance1/3 of the score
Ethics1/3 of the score
Environment1/3 of the scoreSynthesis of 5 public rankings
33.3 %of the Celsius score
LMArena, Artificial Analysis, Epoch AI, compar:IA, ARC-AGI-2; best model of each company; in each ranking, the best of the 10 receives 100, the last 0.
- Performance : synthesis of 5 public rankings, equally weighted.
- Ethics : 4 criteria scored from 0 to 3, weighted 30, 30, 20, 20.
- Environment : 100 minus the pollution score; 4 criteria scored from 0 to 3, weighted 40, 20, 20, 20.
- Not in the score : web traffic, share of tokens on OpenRouter, price, declared energy use.
Criteria
Ethics and environment: each criterion scored from 0 to 3, scaled to 100
Performance
- Synthesis of 5 public rankings100 %
LMArena, Artificial Analysis, Epoch AI, compar:IA, ARC-AGI-2; best model of each company; in each ranking, the best of the 10 receives 100, the last 0.
Ethics
- Content produced30 %
Content the assistant produces on request (sexual, misleading images, minors) and results of independent tests.
- Use of data30 %
Use of conversations for training, retention period, storage location, sanctions by the authorities.
- Model safety20 %
Independent safety evaluations (Future of Life Institute, SaferAI, Phare) and resistance to attacks.
- Company governance20 %
Control of the company and deliberate choices: military uses, shareholding, commitments kept or abandoned.
Environment
- Simple query40 %
Consumption of a typical text query, on the default model served: figure published by the company, third-party estimate, failing that the size of the model.
- Advanced queries: video, images, agents20 %
Weight of the resource-hungry uses the product promotes: video and image generation, agents, reasoning enabled by default.
- Decarbonised electricity20 %
Share of low-carbon electricity in the data centres and efforts to increase it: local grid, on-site gas, renewable or nuclear electricity contracts.
- Company's environmental approach20 %
What the company publishes and commits to: energy and emissions figures, environmental report, targets, model optimisation.
Sensitivity to weights
Recalculated score and gap to the published rank
- 1
Claude74= - 2
Gemini71= - 3
ChatGPT62= - 4
Copilot58= - 5
Meta AI48= - 6
Vibe40= - 7
Qwen39= - 8
Perplexity36= - 9
DeepSeek35= - 10
Grok26=
Calculation rules
- 1Normalisation of public rankingsIn each ranking, the best of the 10 receives 100 and the lowest placed 0; the others are placed in proportion to their gap, in points or in rank.
note = (valeur − minimum) ÷ (maximum − minimum) × 100Example. Claude, 3rd on LMArena out of 419 models: 98.2. - 2Synthesis of public performanceAverage of the rankings in which the assistant appears. Copilot and Perplexity, which have no frontier model of their own, receive an estimate: the average of their 2 neighbours in the synthesis (Copilot: Meta AI and Qwen; Perplexity: Grok and Vibe). Grok, absent from compar:IA, is scored on 4 rankings.
synthèse = moyenne des notes normalisées disponiblesExample. Claude: 98.2; 97.1; 96.2; 100.0; 98.2, giving 97.9. - 3Ethics and environment scoresEach criterion is scored from 0 to 3 on the basis of dated facts, scaled to 100, then weighted. The environment score starts from a pollution score: 100 minus that score.
critère = note ÷ 3 × 100 ; environnement = 100 − pollutionExample. Gemini, simple query: pollution 0.5 out of 3, giving 83 out of 100. - 4Celsius scoreEqual-weight average of the 3 unrounded scores.
Celsius = (performance + éthique + environnement) ÷ 3Example. Claude: (97.9 + 76.7 + 46.7) ÷ 3 = 73.7, displayed as 74. - 5Rounding and tiesScores are displayed as whole numbers, 0.5 being rounded up. 2 assistants whose rounded score is equal share the rank. In a comparison of 2 assistants, a gap of less than 3 points is not treated as significant: the 2 scores and the gap are given, without saying which one is ahead. This gap is calculated on the rounded scores, the ones that are displayed.
rang partagé si notes arrondies égalesExample. Qwen 49.4 and Grok 49.1 on performance: joint 6th.
Neutrality
Performance comes from 5 public rankings. Ethics and environment are scored by us on a published grid, each score with its dated source.
Right of reply
A company may ask for a piece of data about it to be corrected, with supporting evidence. Every correction is dated in the changelog and applied at the next update.
Limits
Energy declared by companies cannot be compared from one company to another; 7 companies out of 10 do not publish it. Web traffic ignores mobile apps and embedded uses, and its series comes from the press. The "Simple query" criterion compares figures of different kinds: Google's measurement, CLEER's estimates, the emissions published by Mistral AI, failing that the size of the model; it alone decides 1st place on environment.
Sources
| Score | Source | Data collected | Collected |
|---|---|---|---|
| Performance | LMArena ↗ | Rank of the best model of each company, text ranking with style control, out of 419 models | 28 October 2026 |
| Performance | Artificial Analysis ↗ | Intelligence Index v5.0.1, in points | 4 November 2026 |
| Performance | Epoch AI ↗ | Epoch Capabilities Index, in points | 4 November 2026 |
| Performance | compar:IA ↗ | Rank out of 121 models, French Ministry of Culture arena | 4 November 2026 |
| Performance | ARC Prize ↗ | Percentage passed on ARC-AGI-2 | 28 October 2026 |
| Ethics | ADL, Future of Life Institute, SaferAI, Phare, NewsGuard | Independent evaluations, decisions by authorities, press: one dated source per criterion | 4 November 2026 |
| Environment | Company publications, CLEER, Stanford transparency index | Energy per query, electricity use of data centres, reports: one dated source per criterion | 4 November 2026 |
| Use | Similarweb, publication of 12/10/2026 relayed by OfficeChai ↗ | Share of web traffic to generative AI sites, worldwide, excluding apps | September 2026 |
| Use | OpenRouter ↗ | Share of tokens processed by the platform, by company | week of 26 October 2026 |
| Price | Company pricing pages, jechange.fr | Standard subscription in France; API of the flagship model, input and output | from 7 to 9 November 2026 |
Changelog
- 5 November 2026First publication: 10 assistants, methodology 1.0.
- 7 November 2026Prices checked on company pages, or on 2 concordant sources; web traffic series extended to 8 monthly readings.
- 9 November 2026Prices for Mistral AI, Perplexity, Grok, Microsoft 365 Premium and Claude rechecked on the official pages (new Grok range of 12/10/2026); founding and headquarters taken from primary sources where they exist; the operator of Grok named SpaceX Super Intelligence.
- 11 November 2026Methodology 1.1: the criterion "our use at the consultancy" (25% of performance) is removed; performance is the synthesis of the 5 public rankings. Up: Gemini from 68 to 71, ChatGPT from 60 to 62, Copilot from 56 to 58, Meta AI from 45 to 48, Qwen from 35 to 37 and from joint 8th place to 7th, Grok from 24 to 26. Down: Vibe from 43 to 40, Perplexity from 39 to 36 and from 7th to 8th place. DeepSeek stays at 35 and moves from joint 8th place to 9th; Claude stays 1st at 74. On performance, Gemini moves up to 2nd, ahead of ChatGPT.
- 12 November 2026Qwen, "content" criterion of ethics: score from 1 to 1.5 out of 3 (clear and strict rules, documented political censorship, almost no independent testing of the app); the score is no longer provisional. Qwen moves from 37 to 39 on the Celsius score; on ethics, from 29 to 34 and from 8th to 7th place, ahead of Meta AI.
Files
CC BY 4.0 licence, with the credit "Projet Celsius, AI Assistants Ranking".
Continue in the ranking
The full ranking →
Database entryreferenceAI Impact DatabaseThe figures on the impact of AI, with their methodology and sources: the electricity use of a question, an image, a task.Open the database →MethodologyAI Impact hub methodologyHow the hub chooses, checks and dates its figures, and how it corrects them.The hub methodology →Ranking10 AI assistantsRanking of AI assistantsThe Celsius score: one third performance, one third ethics, one third environment.The ranking →The hub newsletter
This month's ranking, news on the impact of AI and one figure from the AI Impact Database. One newsletter a month at most, free of charge. Past issues
