Are you overpaying for AI? Find out if you are paying 80% more for LLM unnecessarily

Why is this topic relevant

Two years ago, the choice was simple. There was practically only ChatGPT from OpenAI on the market, and that was it.

Today? It’s an entirely different world. In addition to American players, intense Chinese competition has emerged, with the DeepSeek service, and European Mistral has entered the game with an aggressive pricing policy. And it’s not just about price.

The geopolitical context plays a greater role than ever before. The Trump administration is pushing for US technological dominance, and relations with Europe are going through a turbulent period. China is investing heavily in AI. Models such as DeepSeek offer performance comparable to the best at a fraction of the price, but with one catch: the data goes to Chinese servers. Europe is lagging in this race, but Mistral shows that this may not be a permanent state of affairs.

For companies in regulated sectors, there are increasing questions about data security, GDPR compliance, and the EU AI Act. The US Cloud Act theoretically allows US authorities to access data even outside the US. In the context of AI, fines of over €20 million have been imposed for GDPR violations in the last year alone. That’s no small amount.

If your company develops or operates software that uses large language models, you are faced with a decision with real consequences—price, technical, and compliance-wise.

Price pressure changes calculations

Let’s start with the numbers. Mistral Medium 3 costs $0.40 per million input tokens and $2.00 per output. GPT-5.2, the latest model from OpenAI, released on December 11, 2025, costs $1.75 and $14.00. DeepSeek V3.2 offers $0.28 and $0.42. (For newly released models such as GPT-5.2, expect API stabilization to take 2–4 weeks typically. For critical systems, it may be wiser to wait.)

What does that mean in practice? If your application processes 10 million input tokens and 10 million output tokens per month, you’ll pay about $24 with Mistral. With GPT-5.2, it’s $157. With DeepSeek, it’s $7. Annually, this amounts to thousands of dollars per use case. When scaling across products, the amounts are orders of magnitude higher.

Choosing the right model is therefore really important; a cheaper option does not necessarily mean compromising on quality. According to benchmarks, Mistral Medium 3 achieves approximately 90% of Claude Sonnet’s performance. DeepSeek V3.2 outperforms GPT-5.1 in mathematics and code.

Here is a comparison of the current prices of the main models:

Prices are changing rapidly and downward. Competitive pressure from Mistral and DeepSeek is forcing even established players to lower their prices. It pays to monitor developments and reevaluate your choices on an ongoing basis.

Model example: carrier customer chatbot

To make the numbers less abstract, let’s look at a specific calculation.

Imagine a medium-sized parcel carrier that operates a customer chatbot. Customers ask about the status of their shipments, change delivery addresses, and resolve complaints. The chatbot processes 500,000 queries per month. The average query has 500 input tokens (customer question plus database context) and 300 output tokens.

This amounts to 250 million input tokens and 150 million output tokens per month. How do the costs differ?

Difference: $16,950 per year. At an exchange rate of 24 CZK/USD, that’s over 400,000 CZK.

And we’re talking about a single chatbot. A company with multiple AI products or a higher volume of queries will save many times more. Mistral Medium 3 achieves approximately 90% of the response quality compared to GPT-5.1. For most customer queries, this is more than sufficient.

Of course, migration has its costs: testing, prompt editing, and possible changes to integration. But with savings of $17,000 per year, the investment will pay for itself within the first few months.

Three geographical paths

American models

This category includes OpenAI (GPT series), Anthropic (Claude), and Google (Gemini).

What do you get? A mature ecosystem with rich documentation and a large community of developers. GPT-5.2 and Gemini 3 Pro lead in some general intelligence benchmarks. Claude excels at agent tasks and coding. Its Claude Code tool is currently the most reliable choice for autonomous development in the terminal and intent-based programming (so-called vibe coding), where the developer describes the intent and the AI generates functional code.

Disadvantages? Price and compliance. The US Cloud Act allows US authorities to access data outside the US, which creates a potential conflict with the GDPR. Over the past year, fines totaling over €20 million have been imposed for GDPR violations involving AI, including €15 million against OpenAI in Italy.

American models make sense where you need absolute top-of-the-line reasoning capabilities. They are also suitable for applications primarily targeted at the US market, where EU compliance issues are not an issue.

Chinese models

DeepSeek and Qwen from Alibaba are an interesting alternative from a price perspective.

DeepSeek V3.2 is the cheapest model on the market. In math tasks, the “Speciale” version achieves 99.2% in the HMMT benchmark. Better than Gemini and GPT. The price-performance ratio is hard to beat.

However, the data travels to Chinese servers, and the content is subject to Chinese jurisdiction, including potential censorship. This is often unacceptable for regulated sectors in the EU. And then there is another catch: DeepSeek reasoning models “think out loud,” which increases token consumption and brings the real price closer to that of the competition. Add to that the variable API latency, which can be limiting for real-time applications (voice bots, live chat).

When to use Chinese models? For internal computing tasks without sensitive data. Mathematics, algorithmic problems, and code generation for internal tools. Situations where compliance does not play a role and price is the sole deciding factor.

European models

France’s Mistral AI and Germany’s Aleph Alpha, with its PhariaAI platform, form the core of the European offering.

The advantages stem from the regulatory environment: a GDPR-native approach, EU data centers, and a clear audit trail. Mistral offers an open-source Apache-2.0 license, allowing operation on its own hardware without dependence on the cloud. Support for European languages, including smaller ones, is also strong.

The community is smaller than that of its American competitors, and the documentation is less extensive. In absolute benchmarks of general intelligence, Mistral currently lags behind Gemini 3 Pro.

European models are the logical choice for regulated sectors such as finance, healthcare, and public administration. For companies with data residency requirements in the EU. For multilingual applications targeting European markets.

Mistral in more detail

Mistral AI was founded in Paris in 2023 by former researchers from Google DeepMind and Meta AI. Within two years, it reached a valuation of over $14 billion and gained customers such as BNP Paribas, HSBC, Stellantis, Zalando, and the French government.

What is interesting about Mistral? The models are open-source under an Apache-2.0 license. This means that you can run them on your own hardware without relying on third-party cloud services. For companies dealing with data security and compliance, this is often a decisive factor.

For on-premise deployment, Mistral offers the Ministral 3 14B model. It is the most powerful model that can be run offline on standard hardware with 24 GB of RAM. In math and logic benchmarks, it outperforms comparable Chinese Qwen models. A typical use case is the processing of sensitive data that must not leave the company’s infrastructure: medical records, legal files, internal know-how, and due diligence materials. (An interesting development in this segment came from Apple, which recently released the CLaRa framework. It allows document compression up to 128× and RAG operation directly on the device. It is worth considering for companies considering on-premise solutions.)

For cloud deployment, there is Mistral Large 3 from December 2025. The model has 675 billion parameters (41 billion of which are active), a context window of 256,000 tokens, and multimodal capabilities. Its Elo score of 1418 ranks it on par with DeepSeek V3.2 and Qwen 3. It is the most powerful open-weight model on the market.

Decision matrix

For quick reference, we have prepared an overview of typical situations:

In practice, you rarely choose one model for everything. A multi-model strategy is standard today.

Services such as OpenRouter.ai act as a single access point to dozens of models. You pay for actual consumption and switch according to the task. The architecture may look like this: a cheaper model processes common high-volume queries, while a more expensive model handles analytical tasks.

Five steps to a decision

What to do? Try the following procedure.

First, you should audit the current state. Where in the company do you use LLM? What data do you process? How much do you pay? Most companies do not have an overview of the actual costs spread across teams and products. Without this knowledge, meaningful optimization is not possible.

Then define your requirements. Do you have regulatory restrictions? What languages do you need? Is on-premise a necessity or just a nice-to-have? What is the acceptable latency? These parameters significantly narrow down the selection.

A pilot on real data follows this. Compare the quality of responses on your actual use cases, not on synthetic benchmarks. Le Chat and most providers offer a free tier or credits for testing.

And then there’s the legal review. Compare the DPA (Data Processing Agreement) of individual providers. Prepare a roadmap for compliance with the EU AI Act, which will come into full effect in 2026.

Finally, gradual migration. Start with low-risk use cases. Build internal know-how and experience, and only then scale up to critical systems.

Why it matters

For companies operating in European markets, the choice of an AI model is also a question of strategic autonomy. Dependence on a single supplier, whether American or Chinese, is a risk. Diversification is not a buzzword; it is risk management.

The European ecosystem is growing. In addition to Mistral, Germany’s Aleph Alpha operates the PhariaAI platform for regulated sectors, with customers including BWI and Deutsche Bank. DeepL, valued at $2 billion, proves that a European AI company can be a global leader in its field.

Czech is not a priority for any major model. But Mistral has significantly better support for smaller European languages than its American competitors. For companies with products on the Czech, German, and other EU markets, this is a factor worth considering.

Price differences between models are likely to narrow. Quality will even out. What will remain are differences in access to data, the regulatory environment, and strategic position. And these matter more than they might seem at first glance.

At eMan, we focus on software development and integrating AI into business solutions. We will be happy to help you choose the right model or combination of models for your use case, deploy them into your existing infrastructure, or develop a customized solution. Contact us.

---

Frequently asked questions

What are OpenRouter and similar services?
OpenRouter is an aggregator that provides a unified API for accessing dozens of LLM models. You pay for actual consumption and can switch between models without changing the integration.

Is Mistral cheaper than ChatGPT?
Yes. Mistral Medium 3 costs $0.40 per input and $2 per output, which is approximately 3–5 times cheaper than comparable models from OpenAI.

Can Mistral be run on my own server?
Yes. Mistral models are open-source under the Apache-2.0 license and can be run on-premises. The Ministral 3 14B model runs on hardware with 24 GB of RAM.

Is DeepSeek safe for European companies?
DeepSeek processes data on Chinese servers. For companies in regulated sectors in the EU, this may pose a problem under the GDPR and data security requirements.

Which LLM is best for Czech texts?
No large model prioritizes Czech. Mistral has better support for smaller European languages than its American competitors.

What is vibe coding?
Vibe coding (or intent-based programming) is a development method where the developer describes the intent in natural language, and the AI generates functional code. It is currently best supported by Claude Code from Anthropic.

What are the GDPR fines for AI?
Fines of over €20 million have been imposed for GDPR violations involving AI, including €15 million against OpenAI in Italy.

When will the EU AI Act come into force?
The EU AI Act will come into full force in 2026.