AI models compared: Meta Muse, ChatGPT, Claude, Gemini and Mistral

In September and October 2026, almost all major providers introduced new language models: OpenAI the GPT-6 family, Anthropic Claude 5.5, Google Gemini 3.8 Flash and Mistral a preview of Large 4.[1][2][3][4] Meta introduced Muse Wonder, according to Meta its most powerful model to date.[5] For companies, one question matters most: which model is suitable for what, and where does the data end up?

This overview describes all five providers using the same pattern: which models there are, where their strengths lie and where their limits are. After that come rankings, prices and the questions you should clarify before making a decision. The order of the sections is not a ranking. We deliberately do not give a recommendation here, because it depends on your tasks. The information is as of October 13, 2026. In this market, that can change within weeks.

Meta: Muse

Muse is Meta’s new model series. It comes from Meta Superintelligence Labs, the group’s AI research unit. The name covers a whole family. Muse Spark arrived on April 8, 2026, as the first model in the series and powers the Meta AI app.[6] Muse Spark 1.3 is now the recommended version in the Meta Model API.[7] According to Meta, Muse Wonder and Muse Wonder Full are its most powerful models to date, built for agents. Wonder has been available since October 12, 2026, via the Meta Model API and the coding tool Muse Code.[5] Muse Glimmer is the only Muse model with open weights.[8] In addition, a personal AI agent for consumers in the US is launching under the Muse name.[9]

Strengths: Muse Glimmer has 30 billion parameters, is released under the free Apache 2.0 license and, according to Meta, runs on a computer with a single off-the-shelf graphics card.[8] So it can be self-hosted. In the Meta Model API, Meta does not use your content for training in Standard mode.[10]

Limits: We found no information from Meta on processing in the EU. Meta only states that model availability depends on the country.[7] Besides Standard mode, there is a much cheaper Contributor mode. In it, you allow training on your content. In Muse Code, this mode is preselected for subscriptions.[10] Meta has not signed the EU’s voluntary code of practice for providers of general-purpose AI models.[11] For the multimodal Llama 4 models of the previous series, Meta grants companies headquartered in the EU no usage rights under the license.[12]

OpenAI: GPT-6 and ChatGPT

OpenAI offers three tiers. GPT-6 Astra is the most powerful model in the series, GPT-6.1 Sol is said to be almost as capable and much cheaper, and GPT-6 Luna is intended for large volumes of simple tasks.[13] Since October 7, 2026, the paid ChatGPT plans have been running GPT-6 based on Sol.[14]

Strengths: By default, OpenAI does not use data from ChatGPT Business, Enterprise and the API for training. A data processing agreement is available for these offerings.[15] In the API, eligible customers can have data stored and processed in Europe, in the European Economic Area and Switzerland.[16] With gpt-oss, OpenAI has released two open-weight models under Apache 2.0 that can be self-hosted.[19] OpenAI has signed the EU’s voluntary code of practice for providers of general-purpose AI models.[11]

Limits: Processing in Europe is only available to eligible API customers.[16] For newer models, it costs a 10 percent surcharge.[17] For ChatGPT Business, OpenAI is currently rolling out a choice of storage location step by step. However, it only covers storage, not processing. A copy of every input and response is kept temporarily in the US for abuse monitoring.[18]

Anthropic: Claude

In September and October 2026, Anthropic introduced Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5. Opus is intended for long, demanding tasks, Sonnet for a balance of speed and performance, and Haiku for large volumes such as sorting and extraction.[2][20]

Strengths: In business plans and the API, Anthropic does not use data for training by default.[21] The data processing agreement includes the EU standard contractual clauses.[22] In the EU, Claude can be run via Amazon Web Services, for example in Frankfurt, or via Google Cloud with an EU endpoint.[24][25] Anthropic has signed the EU’s voluntary code of practice for providers of general-purpose AI models.[11]

Limits: Via Anthropic’s own API, processing can only be set to global or to the US, not to the EU.[23] The route via AWS or Google Cloud costs a 10 percent surcharge.[24][25] In Microsoft’s Copilot offerings, Claude models are excluded from the EU Data Boundary and initially switched off in the EU.[26]

Google: Gemini

Google offers Gemini 3.8 Flash, available since September 2026, and the larger Gemini 3.1 Pro.[3][27] For companies that use Google Workspace, Gemini is built directly into Gmail, Docs and other apps.[28]

Strengths: For Workspace, Google commits not to use content for training outside your own domain without permission.[28] Via Google Cloud, there is an endpoint that restricts processing to EU member states.[29] Google has signed the EU’s voluntary code of practice for providers of general-purpose AI models.[11]

Limits: Gemini 3.1 Pro is still a preview.[3][27] According to the terms of service, anyone building applications for users in the European Economic Area may only use the paid services of the Gemini API.[30] The current price for Gemini 3.8 Flash only applies until the end of 2026.[27]

Mistral AI: Mistral Large and Vibe

Paris-based Mistral introduced Mistral Large 4 on October 6, 2026.[4] In addition, there are smaller models such as Mistral Small 4.[31] The chat service Le Chat is now called Vibe.[32]

Strengths: Mistral Large 4 was trained in Mistral’s own data centers in Europe.[4] Smaller models such as Mistral Small 4 are released under Apache 2.0 and can be self-hosted.[31] At Mistral, data is stored in the EU by default.[33] In the Enterprise plan, training on your data is off from the outset, and in the Team plan the administrator can switch it off.[34] Mistral has signed the EU’s voluntary code of practice for providers of general-purpose AI models.[11]

Limits: Mistral Large 4 is so far a public preview, and the open weights are expected to follow by the end of October.[4] Depending on the feature, data may temporarily go to subprocessors outside the EU.[33] Individual users are not automatically excluded from training.[34]

Models in German data centers

If you want to use models in German data centers, you will find offerings from German cloud providers. According to IONOS, its AI Model Hub processes requests only in Germany, without storage and without training.[35] STACKIT states that it neither stores data and requests nor uses them for training.[36] Open models such as gpt-oss run there.[37] The German provider Aleph Alpha plans to combine with the Canadian company Cohere, and the authorities still have to approve.[38]

What do rankings say?

Independent rankings such as the Artificial Analysis Intelligence Index compare models based on tests. The order changes frequently, often with every new model.[39] A placement is therefore only ever a snapshot.

For mid-sized companies, a position in a ranking is rarely decisive. What matters more is whether a model reliably handles your specific task, where it processes the data and what the contract covers.

What does it cost?

Via their APIs, all providers bill by tokens, separately for input and output. Here is a selection of list prices in US dollars per one million tokens in standard operation, as of October 13, 2026. The order follows the sections above:

Model	Input	Output
Muse Wonder	$4.00	$20.00
Muse Spark	$1.25	$4.25
GPT-6 Astra	$10.00	$50.00
GPT-6.1 Sol	$2.00	$10.00
Claude Opus 5.5	$4.00	$20.00
Claude Sonnet 5.5	$2.00	$10.00
Gemini 3.8 Flash	$0.75	$3.75

Sources: the providers’ pricing pages.[10][17][20][27] For Gemini 3.8 Flash, the price applies until the end of 2026 and doubles from 2027.[27] With several providers, processing in the EU costs 10 percent more. For your budget, what counts in the end is less the price per token than how many transactions you really need to send through a model.

Five questions before you decide

Before you decide on a model, clarify these points with the provider:

1. Is there a data processing agreement? Without one, no personal data should go to the model.
2. Is training on your data excluded? In the plan you have booked, not just in the enterprise offering.
3. Where is data processed and stored? Is there an EU endpoint, and what does it cost?
4. How long are inputs and results stored? This belongs in your deletion concept.
5. Can the model be swapped later? A switch should not require new software.

What matters when choosing

Which model fits depends mainly on four things:

• Task: Texts in everyday office work, processes via an API, many small tasks such as sorting or extraction, and multi-step tasks with agents place different demands. All five providers have larger and smaller models for this. Which software you already use often plays a role, too. Sometimes even a simple rule without AI is enough. What to consider with agents is covered in the article AI agents and GDPR.
• Data protection: Whether there is a data processing agreement, whether your data goes into training and how long it stays stored differs by provider and often by plan.
• Hosting: Processing in the EU is available at Anthropic via AWS or Google Cloud, at Google via Google Cloud, at Mistral by default and at OpenAI for eligible API customers. For Meta, we found no information on this. Open models such as Muse Glimmer, gpt-oss or the open Mistral models can also be run in your own data center or with a German cloud provider.
• Costs: The list price, the EU surcharge and above all the volume of transactions determine what a model costs you in operation.

Conclusion

There is no single best model. Every provider has strengths and limits, and rankings change often. Which model fits depends on your task, your data protection requirements, the hosting you want and the costs.

Do not commit yourself too firmly. Build your processes so that the model can be swapped. Then you can switch when offerings, prices or terms change. At the pace of this market, that is more a matter of months than of years.

We are happy to discuss in detail what is right for your company. In the intro call we clarify which model suits your tasks.