Skip to content

OpenAI, Anthropic, Google,
Mistral or open models?

For most EU businesses building an agent, the choice is between a hosted model with EU processing and an open-weight model you run yourself. OpenAI, Google and Mistral offer EU processing on their own APIs; Anthropic offers it through Amazon Bedrock or Google Cloud. Self-hosting gives full control but rarely costs less at small volume.

What does each model cost per million tokens?

Model providers charge per token. A token is a small piece of text, roughly three quarters of an English word. Input tokens are what you send (instructions, the conversation, documents). Output tokens are what the model writes. Agents send a lot of input, because each step resends the instructions and the conversation so far.

The table shows one large and one small current model from each provider, from each provider’s own pricing documentation on 19 November 2026. Prices are in US dollars as shown. They change often, and new models replace old ones every few months.

Standard API prices per 1 million tokens, 19 November 2026 (USD)
ProviderModelInputOutputContext window
OpenAIGPT-6.2 Astra (large)$8.00$40.00About 1 million tokens
OpenAIGPT-6 Luna (small)$0.10$0.50About 1 million tokens
AnthropicClaude Opus 5.5 (large)$4.00$20.001 million tokens
AnthropicClaude Haiku 5.5 (small, prompts up to 100,000 tokens)$0.10$0.501 million tokens
GoogleGemini 4 Carbon (large)$2.00 until 30 Nov 2026, then $4.00$10.00, then $20.00About 1 million tokens
GoogleGemini 3.1 Flash-Lite (small)$0.25$1.50About 1 million tokens
MistralMistral Large 4 (large, open weights)$0.70$2.101 million tokens
MistralMistral Small 4 (small, open weights)$0.15$0.60256,000 tokens

All four providers offer a batch mode at 50% off for work that can wait, and discounts for cached input that the model has seen before. For an agent that resends the same long instructions on every turn, caching can cut the input bill by a large share.

Can you keep data in the EU?

EU processing and data use, as stated by each provider on 19 November 2026
ProviderEU processing on own APIOther EU routesAPI data used for training?
OpenAIYes, for eligible organisations: an EU region (EEA and Switzerland) set per project. Contact sales. 10% price uplift for newer models.Not covered here.No, unless you opt in. Abuse monitoring logs kept up to 30 days by default.
AnthropicNo. Inference options are global or US only, and storage is in the US.Amazon Bedrock EU regions (including Ireland) and Google Cloud EU multi-region, at a 10% premium for regional endpoints.Not without your express permission.
GoogleThrough Vertex AI: an EU multi-region endpoint that covers EU member states only.Single regions such as Germany for some models.No, on paid services. Paid terms apply to all use in the EEA, Switzerland and UK.
MistralYes. Data hosted in the EEA by default. A US endpoint is optional.Open weights for some models, so you can host them.Default not stated clearly on the pages we checked. An opt-out toggle exists.
Open-weight, self-hostedWherever you host it.Any EU cloud or your own hardware.No third party sees the data.

EU processing reduces the transfer question under GDPR, but it does not remove it. Some account, billing and usage data can still be processed outside the region; OpenAI says so in its own data residency documentation. Our page on EU data residency for AI models goes into the detail. For Ireland specifically, note that Amazon Bedrock lists Ireland (eu-west-1) as an in-region option for Claude models.

Which models handle agent work well?

An agent needs a model that can call tools (functions in your code), return structured data in a fixed format, and keep track of a long conversation. All four hosted providers support function calling and structured outputs on the models in the table. Context windows of about one million tokens are now common on large models.

The real difference shows in your own tests. Models differ in how reliably they follow instructions, when they ask for confirmation, and how often they call the wrong tool. Build a test set of 50 to 100 real cases from your business and run each candidate model against it. Pick the cheapest model that passes. Our OpenAI and Anthropic integration pages describe what each API offers.

When does an open-weight model make sense?

Open-weight models are models whose weights you can download and run on your own servers. Examples include Mistral Small 4 (Apache 2.0 licence), Qwen3.8-27B (Apache 2.0) and Meta’s Llama 4 models. Licences differ, and some have conditions:

  • Mistral Large 4 uses a Mistral licence that requires a separate commercial licence if you sell the model as a service.
  • The largest Qwen3.8 model uses its own licence, with conditions above set user and revenue levels. The smaller Qwen3.8-27B model is Apache 2.0.
  • Llama 4 uses the Llama 4 Community License. Its acceptable use policy does not grant the licence rights for multimodal Llama 4 models to companies with a principal place of business in the EU. The policy says this does not apply to end users of a product that includes such a model. Read the licence with your solicitor before you use Llama 4 in an Irish company.

Self-hosting gives you full control of the data and no per-token bill. In return you pay for GPU servers whether the agent is busy or idle, and you must run, update and secure them. At the volumes of most small businesses, a hosted API with EU processing costs less in total. Self-hosting makes sense for strict data rules, steady high volume, or a model you have fine-tuned. See fine-tuning vs RAG for when that is worth it.

Which one suits you?

  • You want EU processing on a direct contract with the model maker: Mistral (EEA by default), OpenAI (EU region for eligible organisations), or Google through Vertex AI.
  • You want Claude models with EU processing: use them through Amazon Bedrock or Google Cloud in an EU region, not the Anthropic API directly.
  • You process sensitive data and must keep it on your own servers: an open-weight model with an Apache 2.0 licence, hosted in an EU data centre.
  • Your volume is low and the data is not sensitive: the cheapest model that passes your tests, on the standard API. Do not over-engineer it.

Common questions

Which AI model is cheapest for an agent?

On 19 November 2026 the smallest models from OpenAI and Anthropic were listed at $0.10 per million input tokens, with Mistral Small 4 at $0.15 and Gemini 3.1 Flash-Lite at $0.25. Output costs more. The cheapest model is the one that passes your tests at the lowest total cost, including retries and caching. Check current prices first.

Is the Anthropic API available with EU data residency?

Not on Anthropic’s own API, according to its documentation on 19 November 2026: the inference options are global or US only. You can use Claude models in EU regions through Amazon Bedrock, which lists Ireland, or through Google Cloud’s EU multi-region. Regional endpoints carry a 10% price premium.

Do OpenAI and Google train on our API data?

Both state that they do not, by default, on paid API use. OpenAI keeps abuse monitoring logs for up to 30 days unless you are approved for zero data retention. Google’s terms say paid service prompts and responses are not used to improve its products. Confirm the terms for your own account and keep a copy.

Is it cheaper to run an open-weight model ourselves?

Usually not at small volume. A GPU server costs money every hour, even when no one uses the agent, and someone must maintain it. Self-hosting starts to pay when volume is high and steady, or when data rules leave no other option. Run the numbers with real traffic before you decide.

Sources

Bring us the task.

Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.

Start a project.

Two people on a sofa with laptops, planning a project beside a whiteboard of sticky notes

Book a call or send us a message. We read everything.

tash@neu.ie
Working with teams in Ireland and worldwide.

Get in touch

Free 30-minute AI review.