Between 28 October and 17 November 2026, the four most closely watched AI companies each released a new model: Gemini 4 Carbon from Google, Mistral Large 4 from France’s Mistral AI, Claude Fable 5.2 from Anthropic, then GPT-6.2 from OpenAI. The top of the rankings changed hands several times in a month. For a business, the right question is therefore not “which one is best?” but “which one for which job, and on what terms?”.
The four families at a glance
Every provider now offers two tiers: a frontier model, slower and more expensive, for difficult tasks, and an everyday model, fast and much cheaper. Assistants often switch between the two without the user noticing.
| Provider | Assistant | Frontier model | Everyday model | Known for |
|---|---|---|---|---|
| OpenAI (USA) | ChatGPT | GPT-6.2 Astra | GPT-6.2 Sol | Versatility, built-in tools (images, voice, agents) |
| Anthropic (USA) | Claude | Claude Fable 5.2 | Claude Sonnet 5.5 | Writing, programming, long documents |
| Google (USA) | Gemini | Gemini 4 Carbon | Gemini 4 Argon | Web search, integration with Gmail, Docs and Android |
| Mistral AI (France) | Vibe (formerly Le Chat) | Mistral Large 4 | Mistral Medium 3.5 | European provider, model you can install yourself |
What the rankings say
Independent rankings compare models on reasoning, mathematics and programming tests. The most widely quoted, Artificial Analysis, put Gemini 4 Carbon at the top on release. According to several reports, Claude Fable 5.2 overtook it a few days later, and early results place GPT-6.2 Astra just behind. The top three are within a few points of each other.
Mistral remains clearly behind on these tests. Its appeal lies elsewhere: a much lower price, a French provider, and published weights, meaning a model you can download and run on your own servers.
Worth remembering: a few points on a ranking make no difference to a letter or a meeting summary. They show on long, difficult tasks: a complex program, the analysis of several hundred pages, multi-step reasoning.
What it costs
For individual use, subscriptions cost around €20 a month, and every provider has a limited free plan.
| Subscription | Indicative monthly price | Frontier model included |
|---|---|---|
| ChatGPT Plus | €23 incl. VAT | GPT-6.2 Astra, with usage limits |
| Claude Pro | €18 excl. VAT, about €22 incl. VAT | Claude Fable 5.2 |
| Google AI Pro | $19.99 in the USA | Gemini 4 Argon; Carbon only in the Ultra plan |
| Mistral Vibe Pro | €17.99 incl. VAT | Mistral Large 4 |
Team plans (ChatGPT Business, Claude Team, Vibe Team) add a business account and user administration, for about €18 to €26 excl. VAT per person per month depending on the provider and the commitment.
For developers and software that call a model directly, the price depends on the volume of text processed, counted in tokens (word fragments). The differences are considerable.
| Model | Text sent | Text generated |
|---|---|---|
| Claude Fable 5.2 | $8 | $40 |
| GPT-6.2 Astra | $8 | $40 |
| Gemini 4 Carbon | $2, then $4 after 30 November | $10, then $20 |
| Claude Sonnet 5.5, GPT-6.2 Sol | $2 | $10 |
| Mistral Large 4 | $0.70 | $2.10 |
For the same volume, Mistral Large 4 costs about ten times less than the American frontier models. For bulk processing (sorting emails, summarising records, translating a catalogue), an everyday model is often enough: the frontier model only pays off on difficult tasks.
Which one for which job
- Writing, rewording, summarising long documents. Claude and ChatGPT are the most often cited, and the difference is mostly a matter of style. A test on your own documents decides better than a ranking.
- Programming. Claude Fable 5.2, GPT-6.2 Astra and Gemini 4 Carbon compete for the top of the coding tests. The tools around them, such as Claude Code, Codex or Antigravity, matter as much as the model.
- Searching the web, working in Gmail and Google Docs. Gemini is built into Google’s tools, which saves copying and pasting.
- Working in Word, Excel and Outlook. Microsoft 365 Copilot relies on OpenAI’s models and, depending on the plan, also offers Anthropic’s. The model is then chosen inside the tool.
- Sensitive data. A model with published weights, installed on a server you control, sends nothing to a provider. Mistral Large 4 needs a powerful machine; smaller models, such as Alibaba’s Qwen3.8-27B, run on a single graphics card. The Chinese models DeepSeek and Qwen should only be used installed on your own premises, not through their online services.
- Small budget, large volume. The everyday models (Sonnet, Sol, Argon) or Mistral through the API.
The account matters more than the model
The same model does not offer the same guarantees with a personal subscription and with a business account: use of conversations for training, place of processing, access log. The article “AI at work: the same tool, not the same guarantees” explains these differences. Choosing a model without choosing the account is only half the job.
Key points
- The three American providers are within a few points of each other, and the ranking changes almost every month.
- Each provider has a frontier model and an everyday model: the latter is enough for most work.
- Individual subscriptions cost around €20 a month; through the API, prices vary by a factor of ten.
- Mistral trails in the rankings, but it is European, much cheaper and can be installed on your own servers.
- The type of account matters as much as the choice of model.
Situation on 20 November 2026. Names, prices and rankings change often: the provider’s own page is the reference before subscribing. Subscription prices taken from comparisons published in October and November 2026 (ChatGPT, Claude, Gemini, Mistral); API prices from announcements and press coverage. Sources: Gemini 4 Carbon and its availability (9to5Google), Mistral Large 4 (Lindy), Claude Fable 5.2 (CNBC), GPT-6.2 Astra and Sol (9to5Mac), Artificial Analysis ranking.