Four language models, the same knowledge base
This page compares four chatbots directly. All four answer questions solely on the basis of the content of this website. They use the same software and the same process: the question is analysed, relevant content is searched for in the index, and then processed by the language model into an answer. This method is called Retrieval-Augmented Generation – RAG for short.
What differs at the core is the language model and – provided the respective provider offers its own embedding model – the embedding model as well. Mistral, Claude, GPT and Gemini – four providers, one and the same knowledge.
Why this can be compared at all
RAG separates knowledge and model. The knowledge resides in a dedicated vector database and comes from your own documents. The language model processes the retrieved information and formulates the answer from it. This makes the language model an interchangeable component: it can be swapped without having to rebuild the knowledge base. That is exactly what you see here, four times over.
The language model is interchangeable
The software behind all four chatbots is dAi Pro; how requests are processed and logged is described under GDPR-compliant AI and in the information window ("?") of each chatbot.
dAi Pro is not tied to a specific AI provider. Depending on the requirement, different language and embedding models can be used – for example with a focus on answer quality, speed, cost or European data processing.
The knowledge base and application remain unaffected by this. Companies can switch the model in use without having to rebuild their content or their RAG application.
What you can look out for when comparing
Ask all four chatbots the same question one after another. Compare less the underlying facts – these come from the same sources – and more the way each model generates an answer from them:
- Length and structure. Does the model answer concisely and directly, or in detail and structured form?
- Handling knowledge gaps. What happens when the knowledge base does not contain a sufficient answer? Is this communicated transparently, or does the model supplement information from its own knowledge?
- Source citations. What evidence is cited, and does it support the statement?
- Speed. The response times differ considerably.
How to read this comparison
This comparison is not a scientific benchmark, but a practical snapshot based on a manageable knowledge base.
There is an important technical difference between the four configurations: each of the four configurations has its own vector index. If a provider supplies its own embedding model, that model is also used for the respective index. As a result, even the retrieved text passages can differ slightly – before the language model even formulates the actual answer. Models, prices and response behaviour also change constantly. A comparison only becomes meaningful with your own documents and your own questions.
Costs
Alongside answer quality and data protection, a third factor determines the ongoing operation of a RAG application: the price per token. Between the four providers, the difference is not measured in percentages but in orders of magnitude.
Billing is based on tokens – roughly speaking, fragments of words. The input comprises everything that goes in: the question and the text passages found for it. The output is the generated answer. Prices apply per 1 million tokens, in euros, as of September 2026.
- Input – question and text passages
- Output – answer
Mistral Large 3
€0.44
€1.30
Gemini 3.6 FlashIntro price until 31 Dec.
€0.64
€3.21
Claude Opus 5
€4.28
€21.41
GPT-5.5
€4.28
€25.69
- The difference between the cheapest and the most expensive model is a factor of 10 to 20. An input token with GPT-5.5 costs roughly ten times as much as with Mistral Large 3, and an output token roughly twenty times as much.
- The cheapest model in the comparison is the European one. Mistral Large 3 is even below Gemini 3.6 Flash – and the latter is expressly designed for speed and low cost. Anyone running their knowledge base on Mistral therefore not only stays within the EU for data protection but also comes out ahead economically.
- Answers cost more than questions. Output tokens are three to six times more expensive than input tokens across all four providers. In a RAG application, however, considerably more input tokens are involved, because the text passages found accompany every question.
- A price with an expiry date. Gemini 3.6 Flash runs at introductory prices until 31 Dec 2026. After that, an answer regularly costs €6.41 instead of €3.21 per million tokens – roughly five times that of Mistral Large 3.
- The list price is not the final price. Batch processing halves the cost, and for recurring prompt components caching reduces the input price by up to 90 percent, depending on the provider. With RAG this carries particular weight, because the same system prompt accompanies every question.
Source: public API price lists from Mistral, Google, Anthropic and OpenAI, retrieved in August 2026. USD → EUR conversion at the ECB reference rate of 21 August 2026 (1 USD ≈ €0.856). Token prices change continuously – and they are only half the equation: what matters is the cost per answered question, since the models require different numbers of tokens for the same answer.
How fast are the four?
What is measured is the time from sending the request to the provider until the first character of the answer arrives. What this website does beforehand — searching the knowledge base — is not included; that part is identical for all four. And only questions that were put to all four chatbots word for word and under the same conditions are counted. Anything else would not be a comparison but a random sample.
The figures are averages from live operation. Ask all four the same question and your own measurement will be part of this within seconds.
- Time to first character
Mistral Large 3mistral-large-latest
2.0 s
GPT-5.5gpt-5.5
4.0 s
Gemini 3.6 Flashgemini-3.6-flash
7.5 s
Claude Opus 5claude-opus-5
12.0 s
How reliable is this? These are measurements from live operation, not a lab test — and they fluctuate considerably. Contributing factors include: load at the provider and queueing at peak times; the time of day, since the three US providers are also used from North America during the European afternoon; the network path, as Mistral answers from within the EU while the other three sit across the Atlantic; routing, peering and packet loss along the way; the length of the question together with the text passages sent with it; whether a prompt cache applies at the provider; rate limits and retries after errors; and simply how the data centres happen to be doing that day. A single outlier therefore shifts the average noticeably. What the figure does not measure: how long or how good the answer turns out — only how long it takes the model to start speaking.
Mistral Large 3
Mistral AI, ParisMistral Large 3 is the language model used by default on this website and the most powerful general-purpose model from the French provider. Both embeddings and the processing of responses can take place within the EU – which may be a decisive criterion, particularly for public authorities and regulated industries. Anyone who compares this with the other three chatbots will see what European models are now capable of.
Claude Opus 5
Anthropic, San FranciscoClaude Opus is Anthropic's high-performance model for demanding tasks. In practice it answers at greater length and with more structure than the other three, and it flags uncertainty more explicitly. A particular feature of this configuration: Anthropic does not offer its own embedding model. The index is therefore built with Mistral embeddings. The selection of matches thus comes from Mistral, while the formulation is done with Claude.
GPT-5.5
OpenAI, San FranciscoThe model that most people already know from ChatGPT – and therefore the benchmark that many already have in mind anyway. Here it is bound to an outside knowledge base and may draw only from that. Both the language model and the embedding model come from OpenAI: what is compared here is thus an entire provider's stack, not just the responding model.
Gemini 3.6 Flash
Google, Mountain ViewGoogle's variant from the Gemini series, tailored for speed and low costs. It allows you to examine the question that is decisive for RAG: How powerful does the language model really need to be once the retrieval step has already selected the relevant text passages? Ask the same question here and further up the page – the difference you see is the contribution of the model.