Claude, ChatGPT or Gemini: in 2026, all three can write, code and read a long file. For a small or mid-sized business, this comparison is decided elsewhere: where your data goes, what a million words processed costs, and what you can keep in-house. Up to date as of November 15, 2026, refreshed every six months.
Key takeaways
- The flagship versions on November 15, 2026: Claude Opus 5.5 and Fable 5.2 (Anthropic), GPT-6 Astra (OpenAI), Gemini 4 Carbon (Google), Grok 4.7 (xAI) and Mistral Large 4 (Mistral AI).
- Price gaps are huge: from 0.10 to 10 dollars per million input tokens depending on the model. For many routine tasks, a small model is enough.
- On their paid APIs, Anthropic, OpenAI, Google and xAI commit not to train their models on your data by default. On free consumer versions, that is not the rule.
- Processing locations differ: Mistral hosts in Europe by default, OpenAI offers Europe as an option, Anthropic and xAI offer global or US processing.
- For files that must not leave the building, open models run on your own server: Qwen, DeepSeek, Gemma, Mistral or Meta's models.
Claude vs ChatGPT vs Gemini: versions and prices as of November 15, 2026
A "token" is a piece of a word: a million tokens is a few hundred thousand words. Public API prices, in US dollars, as published by each vendor.
| Vendor | Flagship and budget model | API price (input / output, per million tokens) | Where your data is processed | Training on your data |
|---|---|---|---|---|
| Anthropic (Claude) | Opus 5.5; Fable 5.2 for demanding reasoning; Haiku 5.5 | Opus 5.5: $4 / $20; Fable 5.2: $8 / $40; Haiku 5.5: $0.10 / $0.50 | Global by default, US option (+10%) | No by default on the API and the Team and Enterprise plans |
| OpenAI (ChatGPT, GPT) | GPT-6 Astra; GPT-6.1 Sol; GPT-6 Luna | Astra: $10 / $50; Sol: $2 / $10; Luna: $0.10 / $0.50 | Europe option (EEA and Switzerland), +10% | No by default on the API |
| Google (Gemini) | Gemini 4 Carbon; Gemini 3.1 Flash-Lite | Carbon: $2 / $10, then $4 / $20 from December 1, 2026; Flash-Lite: $0.25 / $1.50 | EU region available through Google Cloud (Switzerland excluded) | No on the paid tier; yes on the free tier |
| xAI (Grok) | Grok 4.7 | $2 / $6 | Global, or US (+10%) | No without your explicit permission |
| Mistral AI (Le Chat) | Mistral Large 4, open weights | $0.70 / $2.10 | European Union by default | Opt-out on Le Chat; nothing leaves if you host it |
| Open models in-house | Qwen 3.8, DeepSeek V4.1, Gemma 4, Mistral, Muse Spark Open (Meta) | No per-token price: the server and the electricity | On your machine | No: nothing leaves |
On subscriptions, Anthropic lists Claude Team at 25 dollars per seat per month (20 on annual billing), and Mistral lists Le Chat Team at 24.99 dollars per user per month, excluding taxes.
Which model for which job
No in-house ranking: benchmarks change every month. Here is what the vendors claim themselves.
- Writing and correspondence. All flagship models write well in French, English and Spanish. OpenAI presents GPT-6.1 Terra as suited to "clear, natural writing".
- Code and agents. Anthropic positions Opus 5.5 for coding and in-depth work handed to agents, and Fable 5.2 for long, demanding tasks. OpenAI presents GPT-6 Astra as its model for complex reasoning and coding.
- Long documents. Claude, GPT-6 and Gemini 4 accept about one million tokens at once, and so does Mistral Large 4; Grok 4.7 accepts 500,000.
- Images, audio, video. Gemini 4 is the broadest: text, image, audio, video and PDF input. Claude and GPT-6 read text and images.
- The European option. Mistral is the only one of the five that hosts in the European Union by default, and it publishes the weights of its flagship model.
Under the hood: for your IT person
Price sheets cannot be compared line by line. Before choosing, measure on your own documents.
- Different tokenizers: Anthropic states that its models since the 4.7 generation produce about 30% more tokens for the same text. At Google, the output price includes thinking tokens. Compare cost per task, not per token.
- Location: OpenAI serves Europe (EEA and Switzerland) from a regional endpoint, at a 10% uplift. At Anthropic, the inference geography parameter only offers "global" or "us", and storage at rest is in the US. The Vertex AI EU multi-region explicitly excludes Switzerland and the United Kingdom.
- Retention: 30 days by default at OpenAI (abuse-monitoring logs) and xAI; "zero data retention" is possible at both.
- Local: Qwen3.8-27B (dense, 27 billion parameters, Apache 2.0, native context of 262,144 tokens) and Gemma 4 (from E2B to 31B, Apache 2.0) fit on a reasonable machine. DeepSeek V4.1 Flash (MIT), Mistral Large 4 (about 1,000 billion parameters, 52 billion active, under a license that requires an agreement to resell it as a service) and Meta's Muse Spark Open are, by their size, aimed at multi-GPU servers. The latest Llama published on Hugging Face is still Llama 4 (April 2025). Serve them with vLLM or SGLang behind an OpenAI-compatible API: switching models becomes one line of configuration.
What I have learned
I use several models, chosen first by how sensitive the data is, then by price. My agents run around the clock on my own server, with Claude Code as the main engine and OpenAI's Codex CLI as a second engine, in the same portal: I described it in claudeVPS.
What must not leave does not leave. In my home automation setup, speech recognition and speech synthesis run locally; only the reasoning goes through an external API. It is the same rule I describe in my article on shadow AI: the question is not "which model is best", but "which data goes where".
Cost can be managed: about 46 dollars of API usage per month for my agents, with the model chosen to fit the task (see my article on LLM API costs).
Where to start
1. Sort your data into three levels
Public, confidential, secret. For secret data, a model that runs in-house: that is the topic of my article on self-hosted AI, and of the one on large models on small machines.
2. Choose a business plan, not a free account
Check three points in the terms: no training on your data, the processing location, the list of subprocessors. For personal data, Swiss law already applies: see the Swiss Data Protection Act and AI.
3. Test two models on your real tasks
One flagship model and one budget model, for a week, on ten real tasks: note quality and cost. For specialized tasks (transcription, reading invoices, translation), a specialized model often does better for less.
Frequently asked questions
What is the best AI model in 2026?
There is no single one: flagship models change every few months. Choose by data sensitivity and cost per task, and keep the option to switch.
Do ChatGPT or Claude use my data for training?
On the API and business plans, not by default, according to the OpenAI and Anthropic privacy pages. On consumer accounts, OpenAI uses content for training unless you opt out, and Anthropic leaves that choice in the settings.
Is there a European AI?
Yes: Mistral AI hosts your data in the European Union by default, and publishes the weights of Mistral Large 4.
Going further
To choose your models and keep sensitive data in-house, see my Open Source Server Stack guides and my infrastructure services. To discuss your case, book a 20-minute discovery call.
Sources
- Models overview, Anthropic, accessed November 15, 2026
- Pricing, Anthropic, accessed November 15, 2026
- Data residency, Anthropic, accessed November 15, 2026
- Is my data used for model training? (commercial products), Anthropic, August 18, 2026
- Plans and pricing, Anthropic, accessed November 15, 2026
- Models, OpenAI, accessed November 15, 2026
- API Pricing, OpenAI, accessed November 15, 2026
- Data controls in the OpenAI platform, OpenAI, accessed November 15, 2026
- Data residency, OpenAI, accessed November 15, 2026
- ChatGPT Pricing, OpenAI, accessed November 15, 2026
- Gemini models, Google, accessed November 15, 2026
- Gemini API pricing, Google, accessed November 15, 2026
- Data residency (Vertex AI), Google Cloud, accessed November 15, 2026
- Generative AI in Google Workspace Privacy Hub, Google, October 6, 2026
- Models and pricing, xAI, accessed November 15, 2026
- Regions, xAI, accessed November 15, 2026
- Security FAQ, xAI, accessed November 15, 2026
- Mistral Large 4, Mistral AI, accessed November 15, 2026
- Pricing, Mistral AI, accessed November 15, 2026
- Where do you store my data?, Mistral AI, accessed November 15, 2026
- Mistral-Large-4-1T-A52B, Mistral AI on Hugging Face, September 30, 2026
- Qwen3.8-27B, Qwen on Hugging Face, August 5, 2026
- DeepSeek-V4.1-Flash, DeepSeek on Hugging Face, September 10, 2026
- Gemma 4, Google on Hugging Face, accessed November 15, 2026
- Muse-Spark-Open-816B-A42B, Meta on Hugging Face, September 15, 2026
- meta-llama, Meta on Hugging Face, accessed November 15, 2026
Cover photo: Nikodem Nijaki / Wikimedia Commons, CC BY-SA 3.0 licence.
Tags: AI, Systems & infrastructure, LLM comparison, Claude, ChatGPT, Gemini, Mistral, sovereign AI