Which AI assistant should you choose? 7 tools compared by use

There is no best artificial intelligence in the absolute. There is a best assistant per use, and above all per tool estate already in place. If your teams live in Microsoft 365, the answer is Copilot. If they write and analyse long documents, it is Claude.

This comparison reviews seven consumer and business AI assistants, in September 2026. For each one you get what it actually brings, the model running underneath, the price checked at the vendor and what it cannot do. No single ranking, because a single ranking lies about this category.

Which AI assistant to choose, the short answer

Start with the one already inside your tools. Microsoft Copilot if your company runs on Microsoft 365, Gemini if it runs on Google Workspace. Then add Claude for the roles that write and analyse, and Perplexity for those that research sourced information.

ChatGPT remains the most complete assistant for varied individual use. Mistral is the only option if your data must not leave Europe. Grok mainly earns its place for following live news, and it raises governance questions the other six do not.

How this comparison was built

The criteria come before the names, otherwise a comparison is only a ranking of opinions. I kept five criteria, in this order.

• Integration with the existing estate. An assistant already living in the team inbox gains several months of adoption over a stronger but separate assistant.
• The underlying model and where it runs. Knowing which model processes the data, and under which jurisdiction, drives the legal green light.
• The real price per user. The headline price matters less than the licences you must already own to qualify for it.
• What the tool cannot do. A limit known upfront costs far less than a limit discovered after rollout.
• Reversibility. The ability to change your mind in a year without rebuilding every use case from scratch.

1. Claude, the strongest on long-form writing

Claude is Anthropic's assistant. Its speciality fits in one sentence, it holds the distance on a whole document without losing the thread or inventing to fill a gap. It is the tool I recommend to teams producing writing that carries stakes, briefing notes, tender responses, contract analysis.

Key features

• Artifacts display the result in a separate panel, which makes iterating on a document far faster.
• Claude Code lets you delegate technical tasks from the terminal, including for non-developer profiles.
• Claude Design, Slides and Docs produce deliverables you can use directly, without an office suite.
• Projects group documents and instructions by topic, which avoids re-explaining context in every conversation.
• The assistant plugs into Chrome and Microsoft 365 from the Pro plan up.

Underlying model and limits

Anthropic runs its own model family, and the pricing page names four of them in September 2026, Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5. The main limit is functional breadth. Claude generates images less well than its rivals, and its consumer extension ecosystem stays narrower. Servers are operated outside the European Union, which is handled by contract rather than by a setting.

Pricing

A free plan exists. Pro costs 15 euros per month on annual billing, 18 euros monthly. Max starts at 90 euros per month. On the team side, the standard seat is 18 euros per month annually and the premium seat 90 euros, according to Anthropic's pricing page checked in September 2026.

My take on Claude

Claude has been my writing tool for a long time, for its turn of phrase and for the interface. When I compared models across ten trials in my August 2024 edition on choosing a model, it came out ahead, for a reason that has not moved since. It would rather say it is unsure than produce a plausible and wrong answer. On a client deliverable, that difference in temperament is worth more than a benchmark point. If you are only hesitating between it and ChatGPT, the full case is in our Claude versus ChatGPT for business comparison.

2. ChatGPT, the most versatile on the market

ChatGPT still does the most. Text, image, voice, data analysis, deep research and scheduled tasks live in the same application. It is the best starting point for someone working alone who does not yet know what they will automate.

Key features

• Custom GPTs let you freeze a role, a tone and an output format, then share them internally.
• Voice and video modes make the assistant usable on the move, which few rivals manage as well.
• Scheduled tasks trigger a run without any human action, which brings the assistant close to an automation.
• Codex handles development tasks, with quotas that rise by plan.
• Excel, Word, PowerPoint and Google Sheets extensions bring the tool closer to the working files.

Underlying model and limits

OpenAI runs its own models. Its pricing page names GPT-6 Astra for advanced reasoning, then the GPT-5.6 family in Sol, Terra and Luna. Two limits deserve attention. The free plan shows an instant context window of 27,000 tokens, roughly twelve pages of text, which you feel quickly on a document. And the Go plan, the first paid tier, may include advertising, which OpenAI states on its own page.

Pricing

Free, then Go at 8 euros per month, Plus at 23 euros and Pro from 103 euros. On the business side, the standard seat costs 21 euros per month on annual billing, 26 euros monthly. The premium seat rises to 105 euros annually, according to OpenAI's pricing page checked in September 2026.

My take on ChatGPT

I keep it for general tasks, brainstorming and anything involving image or voice. Its flaw is old and stubborn, it answers even when it does not know. I compared it to a schoolchild sitting an exam in my video on the paid subscriptions, and the image still holds. In practice, that means checking factual claims before sending them to a client.

3. Gemini, the assistant that lives in Google Workspace

Gemini is Google's assistant, and its main argument is not its model. It is that it already reads your Docs, your Sheets and your Drive with no connector to install. For a company on Google Workspace, that proximity is worth several points of raw performance.

Key features

• Gemini appears directly inside Gmail, Docs, Vids and the other Google applications.
• Gemini Notebook acts as a research and synthesis assistant over a document corpus you supply.
• Google Flow produces video scenes, with access to Nano Banana Pro for images on paid plans.
• Deep Research explores a topic in depth and returns a structured report.
• Cloud storage reaches 400 GB from the first paid tier, which rivals do not offer.

Underlying model and limits

Google runs its own family, with Gemini 3.6 Flash on the free plan and Gemini 3.1 Pro above it. The limit is how readable the offer is. The tiers mix the assistant, storage, image generation and video credits, which makes comparison with the others hard. Processing happens at Google, under the commitments of the Workspace contract.

Pricing

The free plan gives access to the fast model. Google AI Plus costs 4.99 euros per month and Google AI Pro 21.99 euros. Google AI Ultra starts at 99.99 euros, with a higher tier at 219.99 euros, according to Google's French subscription page checked in September 2026.

My take on Gemini

Google is catching up and nobody really talks about it, that is what I wrote in my November 2025 edition on the arrival of Gemini 3. What convinced me was not a ranking, it was the breadth of the ecosystem, from search to video creation and developer tooling. I mostly use it to centralise, test an idea fast and explore. For fine writing, I go elsewhere.

4. Microsoft Copilot, the one already installed at your company

Microsoft Copilot is not the best assistant on the market, and that is not its point. It sits in Outlook, Excel, Teams and Word, where your teams spend their day. For an IT department, that argument often settles the matter, because it removes the adoption question.

Key features

• Copilot Chat is included at no extra cost for users holding an eligible Microsoft 365 subscription.
• The assistant reads documents and conversations the user can already access, respecting existing permissions.
• Copilot Studio lets you build business assistants without writing code.
• Central administration controls who accesses what, by user or by security group.
• Features land inside the office applications themselves, not in a separate window.

Underlying model and limits

This is where it gets interesting. Microsoft does not only run its own models, it onboarded Anthropic as a subprocessor. Its documentation, updated in September 2026, states that Anthropic models are excluded from the EU Data Boundary. For customers in the European Union, EFTA and the United Kingdom, they are disabled by default and require an explicit opt-in by an administrator. The same documentation records the rename, Microsoft 365 Copilot becomes Microsoft Copilot.

Pricing

Microsoft Copilot is listed at 26.00 euros excluding tax per user per month on annual billing, on Microsoft's enterprise pricing page checked in September 2026. The trap sits in one line of that page. A separate eligible Microsoft 365 licence is still required. The real cost per user is therefore the sum of both, never the 26 euros alone.

My take on Microsoft Copilot

Tandem has not deployed it on a client project, so I stick to what the documentation allows me to verify. What I observe at clients running Microsoft 365 is that Copilot is almost always the right first step. It asks for no change of habit. The second assistant is added afterwards on the role where value is proven, which is exactly the logic of our AI copilot rollout.

5. Mistral, the only European assistant you can host yourself

Mistral AI is the French vendor on this list, and its assistant is now called Vibe. Its argument is not raw performance, it is control. The models are published in open weights, which allows a deployment on your own infrastructure, and the vendor falls under European law.

Key features

• Vibe runs on the web, in a mobile app, in the command line and inside the development environment.
• The Enterprise plan allows private deployments, with custom models and a custom interface.
• Studio is used to build and test agents, Forge to train and align bespoke models.
• More than a hundred connectors link the assistant to business tools.
• Audit logs and single sign-on are available on the enterprise plan.

Underlying model and limits

Mistral runs its own models, including Mistral Medium 3.5, Mistral Small 4 and Mistral OCR 4. The limit is real and worth stating. On the most demanding reasoning tasks, the leading models remain American. In September 2026 the vendor announced a 3 billion euro round valuing the company above 21 billion, led by Samsung Electronics with the Scaleup Europe Fund and PSG Equity. It is the largest equity round ever closed by a European technology company.

Pricing

A free plan exists. The Pro plan costs 17.99 euros per month including tax and the Team plan 29.99 euros per user, on Mistral's pricing page checked in September 2026 for France. Mind the comparison, Mistral displays prices including tax while the others display them excluding tax.

My take on Mistral

Tandem has not deployed it on a client project, and I would rather write that than pad the section. What I see during scoping is that the sovereignty question arrives late in projects, often after the tool has been picked. When it arrives early, Mistral is the only serious candidate, because it is the only one allowing hosting you fully control.

6. Perplexity, the best for sourced research

Perplexity is not playing the same game as the others. It answers by citing its sources line by line, which makes it a research and monitoring tool rather than a writing assistant. For an analyst, a consultant or a marketing team that must source its claims, it is the clearest time gain on this list.

Key features

• Every answer carries inline citations pointing to the exact source of each claim.
• Paid plans open premium data sources, including PitchBook, Statista and S&P Capital IQ.
• Comet is a browser with the assistant built in, able to act on the pages you have open.
• Computer runs long tasks, from data analysis to document creation.
• Model Council has several models answer the same question so you can compare the results.

Underlying model and limits

Perplexity has no flagship model of its own, and it owns that. Its pricing page states that it orchestrates ChatGPT, Gemini, Claude and NVIDIA Nemotron, picking the best one per query. That is a strength for the reader and a dependency for the vendor. The other limit is long-form writing, where the sourced answer format becomes constraining. On compliance, the vendor states it does not train its models on enterprise customer data, with a SOC 2 Type II certification.

Pricing

The free plan covers simple search. The Pro plan comes to around 20 euros per month and the Max plan around 190 euros, the vendor billing in dollars. Enterprise pricing goes through a quote. Checked in September 2026.

My take on Perplexity

I use it to ask questions and get sources, not to write. That is exactly the slot I gave it in my AI tools pyramid published in January 2026. One point did change my mind though. I was sceptical about platforms aggregating several models, because the output fell short of what each model gave directly. On sourced research that objection no longer holds, because the value comes from verifiability rather than from the model.

7. Grok, the fastest on live news

Grok has one advantage the other six lack, direct real-time access to content on X. To follow a hot topic or gauge a public reaction, it is the fastest. For everything else, it arrives with a heavier governance file than its rivals.

Key features

• Search combines the web and live posts from X, with citations.
• Multi-agent mode puts several agents to work in parallel on the same problem.
• Imagine generates images and video inside the conversation, up to fifteen seconds.
• Voice mode allows a continuous exchange with low latency.
• An application building mode is reachable from the home page.

Underlying model and limits

Grok runs on in-house models, including grok-4.6 which opens a 500,000 token context window. The limit is not technical, it is contractual. The vendor no longer exists as an independent company. SpaceX acquired xAI on 2 February 2026 in an all-stock transaction, and the whole was rebranded SpaceXAI in July 2026. In a purchasing file, that means an entity under reorganisation and no consumer pricing grid reachable without an account.

Pricing

Free access exists. I found no public pricing page for the consumer subscriptions in September 2026, the plans being sold inside the application. On the programming interface side however, the grid is public and grok-4.6 sits at around 2 euros per million input tokens and 6 euros per million output tokens.

My take on Grok

Tandem has deployed it at no client, and no legal department has asked us for it so far. I keep it for personal monitoring on news topics. For professional use in France, the absence of public pricing and the vendor reorganisation are enough to rank it behind the other six.

What I wrote in 2024, and what has not moved

In August 2024 I published an edition asking exactly the same question, which model to choose between GPT, Claude, Gemini and Mistral. I had compared two models across ten trials, split into writing, vision, data analysis, programming and reasoning. Two years later, rereading that edition, three observations still hold.

1. Rankings do not separate the top of the field. They sort the good from the weaker, and become useless as soon as you compare the top five.
2. Each model has a temperament. Some always answer, even when wrong. Others prefer to flag their uncertainty. That trait matters more than the score.
3. Running two assistants in parallel reduces the error risk. Comparing two answers on a question that carries stakes remains the cheapest verification there is.

What changed is the number of slots. In January 2026 I published a pyramid of AI tools by family of use that captures the situation better than any ranking. We moved from one magic tool to a series of tools each answering a precise need. Asking which AI is best in 2026 is asking the wrong question.

I had already run through the paid subscriptions on video, with a simple rule that has not aged. A subscription is justified when a precise feature is missing for you, not when a new model ships. The free tiers of these seven assistants cover discovery comfortably.

How to choose for your context

The right question is not which assistant is best, but which one your teams will open tomorrow morning without thinking about it. Here is the grid I apply during scoping.

• Company on Microsoft 365. Start with Copilot. The licence base is already paid and the assistant lands in Outlook and Teams with no rollout project.
• Company on Google Workspace. Start with Gemini, for the same reason. The assistant reads your Docs and your Drive with no connector.
• Roles that write and analyse. Add Claude on those specific teams. That is where the quality gap shows fastest.
• Roles that research information. Add Perplexity for monitoring, research and marketing teams, for source traceability.
• Sensitive data that must not leave Europe. Look at Mistral first, and ask the hosting question before choosing the tool.
• Individual, non-specialised use. ChatGPT remains the best entry point, because it covers the most cases with no setup.

The model behind your assistant is no longer stable

This is the point comparisons forget, and the one that changes the most for a European company. You no longer choose a model by choosing an assistant. You choose an interface that routes to models, and that routing evolves without you.

Two examples are enough. Microsoft onboarded Anthropic as a subprocessor in Copilot, and its own documentation states that these models fall outside the EU Data Boundary. Perplexity, for its part, orchestrates several vendors and picks the model per query. In both cases, the question of where your data is processed no longer has a single answer, it has an answer per setting.

The practical consequence is simple. Before signing, ask for the vendor's subprocessor list and check which settings are active by default on your tenant. For European customers, Microsoft disables Anthropic models by default, which means enabling them is a decision rather than a technical detail.

That instability has an upside. It makes the choice less final than it looks. What you build around the assistant, the use cases, the shared instructions and the team habits, travels from one tool to another better than you fear.

What regulation already asks of you

Many companies are waiting for future deadlines while an obligation has applied since 2 February 2025. Article 4 of the European AI Act requires providers and deployers alike to take measures to build AI literacy among their staff. Rolling out an assistant without training your teams is therefore not only ineffective, it is out of compliance.

On personal data, the French regulator has published concrete and readable guidance. It calls for an internal charter clearly defining permitted and forbidden uses. It also calls for disabling data reuse on public services, and for training users on the limits of these systems. The CNIL question and answer page on using a generative AI system fits on one page and belongs in any rollout file.

In practice these two requirements are handled together. They also come before the choice of tool. That is the purpose of our Claude and AI assistant training programme. It is also the underlying subject of our article on companies that see no return on their AI investment.

Which AI assistant to choose in 2026, what I recommend

Stop looking for the best assistant and take the one already inside your tools. Copilot on Microsoft 365, Gemini on Google Workspace. That default choice wins you the three months of adoption you would lose rolling out a separate tool, even a stronger one on paper.

Then add a second assistant on one precise role, and only one. Claude for the teams that write and analyse, Perplexity for those that research. Keep Mistral in mind if sovereignty is a subject at your company, and ask that question before choosing, never after.

And if you only have time to test one tool this quarter, take two and ask them the same question for a week. That is what I recommended in 2024 and it remains the cheapest way to form an opinion. The ranking of the month will tell you nothing about what your teams will actually do with it. To dig into a specific choice, our detailed analyses compare Claude and ChatGPT for business, break down the Claude Pro and Max subscriptions, and explain why teams switch to Claude.