What are the best AI assistants of 2026?
Our multi-criteria ranking

The most widely cited AI rankings only score performance. Projet Celsius, a French consultancy in environmental strategy and AI ethics, adds ethics and environment, with equal weight, for the 10 best-known consumer assistants, from ChatGPT to Qwen.

1What are the best AI assistants in 2026?

Claude leads, ahead of Gemini and ChatGPT

An AI can be brilliant and behave badly, or frugal and not very capable: the best is the one that holds up on all 3 fronts at once. On that basis, Claude comes first with 74/100, ahead of Gemini (68) and ChatGPT (60).

Copilot, 4th with 56, is the most consistent: its 3 scores sit between 53 and 58. Vibe, the French assistant from Mistral AI long known as Le Chat, comes last on performance but climbs to 6th place thanks to its frugality. Grok brings up the rear at 24/100, with the worst score on both ethics and environment.

The planet weighs as much as performance

Each AI receives 3 scores out of 100, then a final score, their average with equal weights. In all 3, the higher the score, the better. The free version and audience size stay out of the score, as notes below the table:

• The performance score combines 5 public rankings (75%) and our own experience of use (25%).
• The ethics score is the one from our ranking of the most ethical AI assistants, covering what they agree to produce, your data, their safety and the company behind them.
• The environment score is the one from our ranking of the most polluting AI assistants, covering the model that answers, images and video, the electricity powering the servers and the developer's efforts.

2Which AI performs best?

Performance is what people compare first, and it is also what changes fastest. Our score combines 5 public rankings: 2 where users pick the better of 2 answers blind, LMArena and the French compar:IA, and 3 where every AI sits the same exams, those of Artificial Analysis, Epoch AI and ARC Prize.

On that basis, Claude is the best-performing AI, with 98/100, ahead of ChatGPT (90) and Gemini (86). Vibe, from Mistral AI, comes last at 15/100.

Progress is fast. The ARC-AGI-3 test is made of small logic games that are easy for a human. In March 2026, no AI solved more than 1% of them. By the end of October, the best, from OpenAI, solved nearly 2 in 3.

No AI wins everywhere

Claude wins only one of the 5 rankings, the French users' vote on compar:IA, but it is 2nd in all the others: its consistency gives it the best average. Gemini wins the worldwide vote on LMArena and the Artificial Analysis index, ChatGPT the hardest exams.

Claude excels at code, ChatGPT at the hardest problems

Each has its strength. Claude is the best at writing code, building websites and running agents, AIs that carry out a series of tasks on their own. ChatGPT is the only AI to answer more than half correctly on Humanity's Last Exam, specialist questions from every discipline chosen because AI assistants failed them: 54.8% in September 2026.

Vibe, the French AI from Mistral AI, is clearly behind: it ranks only 113th out of 419 in the users' vote. Yet it is the best European AI. Our comparison of ChatGPT, Claude and Gemini compares the models one by one.

3Which AI is the most ethical?

An AI's ethics are judged by what it agrees to produce, what it does with your data and how its developer holds its limits. On these points, Claude is the most ethical AI, with 77/100, ahead of Copilot (58) and ChatGPT (52): it refuses sexual content even in fiction and said no to the Pentagon in February 2026.

Grok comes last, at 17/100, just behind DeepSeek (18). According to the British NGO Center for Countering Digital Hate (CCDH), it produced around 3 million sexualised images in 11 days in early 2026, and DeepSeek stores your data in China.

With 7 of the 8 major assistants, your conversations train the AI unless you opt out, and the option is switched off in the settings. Our ranking of the most ethical AI assistants details the 10 scores, and should you boycott AI? gathers the arguments of those who give it up.

4Which AI is the least polluting?

Not all AI assistants use the same energy to answer the same question. The gap depends less on the brand than on 2 choices made by the developer: the size of the model that answers by default, and the electricity that powers its servers.

On these 2 points, Gemini makes the best choice: it answers with a lightweight model that Google measured at 0.24 Wh (watt-hour) per query, as much as an LED bulb lit for a minute and a half. Vibe, from Mistral AI, follows closely: it answers in fast mode by default, on European servers, where electricity emits about half as much carbon as in the United States.

At the other end, Grok combines everything that weighs: unlimited video, servers partly powered by gas, and not a single published figure. Our ranking of the most polluting AI assistants details the 10 scores.

Claude scores only 47/100 on environment, even though outside estimates judge it rather economical. The reason lies with its developer: Anthropic publishes no energy figures. Without a measurement, the energy used by a query is scored at the midpoint, and Claude loses the points reserved for developers that publish. That silence is the rule, and Google remains the only developer to have measured a query under real conditions.

Our experience of use adds a caveat to this score, without being able to put a figure on it: to reach a given result, Claude goes so directly to the point that it needs fewer back-and-forths, and therefore less computation. Used for real work, it is very efficient. Anthropic points the same way for its latest large model, Opus 5.5, which it says needs less computation than its predecessor, without giving an energy figure. Claude's carbon footprint has its own article.

5Which AI should you use for what?

The scores do not tell the whole story: each AI has ground where it serves better than the others, depending on its strengths and where you come across it.

For documents and code: Claude

Claude is the assistant that comes closest to a colleague: you can hand it a whole task, from code to reviewing a report. Its large model is 1st in the users' votes on website building and agents. Its weak point is the environment, for lack of a single published figure.

For everyday life and images: ChatGPT

ChatGPT serves everyday life: a quick question, an email, an image, a search. It is the most used, with 56% of web traffic to AI sites in September 2026, and OpenAI also leads the votes on images. Its free tier lags far behind the paid one: its free model ranks only 97th out of 419 in the users' vote. The detail is in our article on the most used AI assistants.

The other assistants serve narrower uses

The other 8 each have narrower ground, shaped by their scores and their uses:

• Gemini is built into Google and Android phones: it is the most frugal in the ranking, and its free version is the best placed in the users' vote, 10th out of 419.
• Perplexity suits research, since every question launches a web search and cites the pages it finds.
• Copilot serves the office, installed by default in Windows, with more than 30 million paid licences for Microsoft 365 Copilot in July 2026.
• Vibe, from Mistral AI, keeps your conversations in the European Union by default, which makes it the natural choice for sensitive data.
• DeepSeek is free and its models can be downloaded freely, but its app stores your data in China.
• Meta AI lives in WhatsApp, Instagram and Messenger, and its Muse app is not available in Europe.
• Qwen, from Alibaba, remains little known, since almost no independent test covers its app.
• Grok sits one click from the X feed, but it comes last in the ranking, on ethics as on environment.

6Our verdict

The scores settle the extremes and, for the rest, leave the choice to your use.

Putting the most powerful AI first has a cost for the climate: Claude's large model uses about 20 times as much as the small one, according to CLEER, an independent database of estimates. The right setting is the small model by default, and the large one for the work that needs it.

Our 10 actions ranked by effectiveness show which setting changes the most, our article on the energy used by a ChatGPT query starts from a single question, and AI in a company's Bilan Carbone®, the French carbon accounting method, shows how to account for it.

7Key takeaways

• Claude is the best AI of 2026 in the Celsius ranking, with 74/100, ahead of Gemini (68) and ChatGPT (60).
• The final score gives equal weight to performance, ethics and environment.
• Claude wins on performance, at 98/100, and on ethics, at 77/100, as in our ranking of the most ethical AI assistants.
• Gemini is the most frugal AI, at 73/100 on environment, and Grok has the lowest score on both ethics and environment, according to our ranking of the most polluting AI assistants.
• ChatGPT is still by far the most used, with 56% of web traffic to AI sites, according to our ranking of the most used AI assistants.

For the overall figures, see the carbon footprint of AI and all the impacts of AI, and for the so-called green alternatives, Ecosia, Euria and GreenPT put to the test.

Further resources

Frequently asked questions

1What is the best AI in 2026?

Claude, from Anthropic, with 74/100 in the Celsius ranking of 5 November 2026, ahead of Gemini (68) and ChatGPT (60). The final score gives a third to performance, a third to ethics and a third to environment. Claude wins on performance and ethics, Gemini on environment, and Grok comes last.

2How is the Celsius ranking calculated?

Each AI receives 3 scores out of 100. Performance combines 5 public rankings (75%), such as LMArena, where users vote, or ARC, which sets exams, with Projet Celsius's own experience of use (25%). Ethics and environment come from its 2 detailed rankings. The final score is their average, with equal weights.

3Is Claude better than ChatGPT?

In the Celsius ranking, yes: 98/100 against 90 on performance, 77 against 52 on ethics and 47 against 40 on environment. Claude's large model leads on code and agents in November 2026, and ChatGPT wins the hardest reasoning exams, such as the ARC logic games and Humanity's Last Exam.

4What is the best free AI?

Gemini, from Google. Among the models each developer offers for free, Google's ranks highest in the LMArena users' vote, 10th out of 419 in November 2026, and Gemini is also the most frugal AI in the Celsius ranking. ChatGPT's free model ranks only 97th.

5Which AI is the least polluting?

Gemini, with 73/100 on environment in the Celsius ranking, just ahead of Vibe from Mistral AI (72). Gemini answers by default with a lightweight model, measured by Google at 0.24 Wh per median prompt. Grok is the most polluting, with 13/100 on environment.

6What is the most used AI in the world?

ChatGPT, with 56% of global web traffic to AI sites in September 2026 according to Similarweb, and more than 1.2 billion users a week according to OpenAI. Gemini is 2nd with 25% of traffic, Claude 3rd with 9.6%.

Ask Guillaume your questionA personal reply within 24 working hours

Send my questionor: guillaume@projetcelsius.com