If you stopped following AI news in spring, you have missed a lot. Every major lab has released new models since June, prices have moved in both directions, and the open versus closed debate has shifted. This is our plain-English map of where things stand at the end of September 2026, and what it means when you are choosing technology for your business.
Who are the main AI model providers in 2026?
For most European businesses, the practical choice is among five providers. Here is a short summary of each, based on their own announcements and reputable reporting.
Anthropic (Claude)
Anthropic released its Claude 5 generation in the summer and its 5.5 generation in September. Claude Opus 5.5 launched on 22 September at $4 per million input tokens and $20 per million output tokens, with Anthropic claiming it performs at the level of its top-tier Fable 5.1 on most work. Claude is widely used for coding, document-heavy work and agents, and is available directly and through AWS, Google Cloud and Microsoft Azure.
OpenAI (GPT and ChatGPT)
OpenAI’s GPT-5.6 went public on 9 July 2026 in three tiers, Sol, Terra and Luna, priced from $1 to $5 per million input tokens, according to Engadget. The release was delayed by a new US requirement for government review of the most powerful models before launch, a reminder that access to frontier AI is increasingly shaped by politics as well as technology.
Google (Gemini)
Google has concentrated on its Flash models, shipping Gemini 3.7 Flash in August and 3.8 Flash on 2 September, alongside specialist models for live voice and transcription. Its next flagship is on the way: Google DeepMind’s Koray Kavukcuoglu said Gemini 4 is in post-training and that Google wants to release it as soon as possible, without giving a date.
Meta (Muse Spark)
Meta’s April release, Muse Spark, was its first model from Meta Superintelligence Labs, and it launched as a proprietary model available in the Meta AI app with a private API preview. Meta has said it plans to open-source future versions, but for now its leading model is not something you can download and run yourself.
Mistral (Europe)
Paris-based Mistral remains the most prominent European model provider. Its OCR 4 document model, released in June 2026, supports 170 languages, costs $4 per 1,000 pages through the API ($2 with batch processing) and can run in a single container for fully self-hosted deployments. An updated OCR 4.1 became generally available on 31 August.
Are AI models getting cheaper or more expensive?
Both. The cost of a given level of capability keeps falling. Anthropic, for example, says Opus 5.5 is around 40% cheaper than Opus 5 on typical workloads. But providers are also repricing. Google’s blog states that Gemini 3.8 Flash’s introductory API price of $0.75 per million input tokens and $3.75 per million output tokens rises to $1.50 and $7.50 from 1 January 2027.
The lesson for budgeting: do not build a business case that only works at one provider’s current promotional price. Model the cost at double the current rate and check the project still makes sense.
Is open-source still an option for businesses?
Yes, but the picture has changed. Meta, once the flag-bearer for open models with Llama, launched its newest model closed. Openly available and self-hostable models still exist, and they matter for specific cases: sensitive documents that must not leave your infrastructure, predictable high-volume workloads, or offline environments.
For most small and mid-sized businesses, though, a managed service from a major provider, used under a proper business agreement with data processing terms, is simpler and safer than running models yourself. The right answer depends on the data involved, not on ideology.
How should a European business choose an AI model?
We use five questions with clients:
- What is the task? Reading documents, drafting text, answering customers and writing code favour different models and tiers.
- What data does it touch? Personal data, financial records and client files raise GDPR and contractual questions about where processing happens.
- What volume and speed do you need? A few hundred tasks a month and a few hundred thousand lead to very different choices.
- What does it cost per completed task? Not per token. A cheaper model that needs two retries may cost more.
- How easy is it to switch? If the answer is “hard”, fix that before you scale.
How do you avoid getting locked in to one AI provider?
With three new model generations from a single provider in one year, lock-in is a real risk. Three habits help:
- Separate the model from the process. Your workflow logic, prompts, checks and data connections should live in your system, with the model called through a thin layer you can repoint.
- Use open standards for tools. The Model Context Protocol, originally from Anthropic, is now a founding project of the Linux Foundation’s Agentic AI Foundation, whose platinum members include Anthropic, Google, Microsoft, OpenAI and AWS. Connecting your systems through a shared standard means more than one provider’s agents can use them.
- Keep your own test sets. A few dozen real examples with known answers let you evaluate a new model in hours rather than weeks.
This is how we approach connecting systems and building AI automation: the model is a component, not the foundation.
What should you do this quarter?
If you already use AI, review which models your tools depend on and whether newer, cheaper tiers would do the job. If you have not started, the choice of provider matters less than choosing the right first process. Pick a repetitive, measurable task, such as processing inbound documents, and prove the value there.
Frequently asked questions
Which AI model is best for business in 2026?
There is no single best model. The right choice depends on the task, the data involved, volume and cost per completed task. Testing two or three candidates on your own examples is the most reliable way to decide.
Is Gemini getting more expensive?
Google has announced that Gemini 3.8 Flash API pricing will rise from $0.75 to $1.50 per million input tokens and from $3.75 to $7.50 per million output tokens on 1 January 2027.
Is there a European alternative to American AI models?
Mistral, based in Paris, is the most prominent. Some of its models, including OCR 4 for documents, can be fully self-hosted, which helps with data sovereignty requirements.
What is the Model Context Protocol?
It is an open standard for connecting AI models to tools and data. Anthropic contributed it to the Linux Foundation’s Agentic AI Foundation in December 2025, whose members include Anthropic, Google, Microsoft and OpenAI.
Not sure which models fit your processes, or worried about lock-in? Book a free discovery call with Haystack and we will give you an independent view.
