Are AI models hosted in Switzerland good enough for day-to-day business?
Results: two test rounds, the same error
In August we tested three AI models hosted in Switzerland (glm5, llama4-maverick, qwen3-vl-235b) on a fictitious SME dossier. All three confirmed a figure to a customer that is wrong by CHF 427.50. In October we repeated the test with three newer models (Kimi-K2.6, Qwen3.5-397B, Mistral-Small-4), word for word and with the same dossier. The result is the same: three out of three confirmed the wrong figure. Claude Opus from Anthropic (Opus 5 in August, Opus 5.5 in October) passed all 18 runs of both rounds.
What this means for management
Do not let the hype around local AI mislead you. Models run locally or hosted in Switzerland can make sense, for example for narrowly defined tasks where the instruction states exactly what to do and what to check. In the first two test questions (below), two of the three new models did exactly that well.
If a model is meant to assess a matter from day-to-day business, a quote, a customer request, a figure before signature, the picture from both test rounds is clear: for that we recommend a frontier model such as Claude. On the third test question, the request for confirmation, Claude found the error in every run of both rounds, although nobody had asked for a check.
If your business may not use frontier models from the US or from China, for instance because of professional secrecy or customer requirements, an anonymisation gateway is the way to go. Names, accounts and identification numbers are replaced before sending and never leave the company, while the figures remain fully verifiable. The frontier model only sees pseudonymised data. This is how our Sovereign Cockpit works.
The test
The dossier contains a joinery's quote for a dining-room conversion and three emails from the customer. On top come three questions of increasing difficulty, each in a fresh chat and word for word to every model:
- Arithmetic check: «Check the quote arithmetically. Are all amounts correct?» In item 4, quantity times price does not add up.
- Date chain: Can the anniversary party on 19 September go ahead? Order confirmation, delivery time and execution time lead to 25 September.
- Confirmation: With an extra 5% discount the customer arrives at CHF 41,040 and asks for confirmation. The multiplication is right, the base is not: the net amount contains the error from item 4. The correct figure would be CHF 40,612.50.
The prompts were in German; the quotes on this page are translations. The first round is covered in detail in the viewpoint of 4 August, which also offers the test dossier and the answer key as PDFs so you can run the test with your own AI.
The results of both rounds
Round 1, 4 August 2026: three models hosted in Switzerland against Claude Opus 5 and Claude Fable 5, both at the highest reasoning depth (effort «max»).
| Question | glm5 | llama4-maverick | qwen3-vl-235b | Claude Opus 5 | Claude Fable 5 |
|---|---|---|---|---|---|
| 1 Arithmetic check | passed | passed | failed | passed (3/3 runs) | not tested |
| 2 Date chain | passed | passed | passed | passed (3/3) | not tested |
| 3 Confirmation | failed (2/2 runs) | failed | failed | passed (3/3) | passed |
Round 2, 18 October 2026: three newer models hosted in Switzerland against Claude Opus 5.5 at the second-highest reasoning depth (effort «xhigh»).
| Question | Kimi-K2.6 | Qwen3.5-397B | Mistral-Small-4 | Claude Opus 5.5 |
|---|---|---|---|---|
| 1 Arithmetic check | passed | passed | failed | passed (3/3 runs) |
| 2 Date chain | passed | passed | passed | passed (3/3) |
| 3 Confirmation | failed | failed | failed | passed (3/3) |
Among the new models, Kimi-K2.6 is the largest open model of both rounds, around one trillion parameters. Running it in-house for five concurrent users takes around 650 GB of GPU memory, i.e. one server with eight H200 GPUs (assuming 32,000 tokens of context per user). On questions 1 and 2, two of the three new models do well. On question 3, nothing has moved.
Question 3, in their own words
The answers from the second round, first sentence each:
- Kimi-K2.6: «Yes, the amount of CHF 41,040.00 can be confirmed.»
- Qwen3.5-397B: «Yes, I can confirm this calculation. The customer calculated correctly.»
- Mistral-Small-4: «Yes, I can carry out this confirmation for the back office of Holzbau Wettstein AG.»
All three recompute the last step, 43,200 times 0.95 equals 41,040. None checks whether the 43,200 is right.
What the Kimi case shows. In question 1 the instruction was explicit: «Check the quote arithmetically.» Kimi-K2.6 found the error in item 4. It even noted unprompted that the customer's recalculation to CHF 41,040 «is arithmetically correct, but refers to the still erroneous net amount». So the model knew everything it needed for question 3.
In question 3 the same figure arrived as a friendly request from the customer, in a new chat and without an instruction to check. Kimi confirmed the CHF 41,040.
What the model lacks is initiative; the knowledge is there. It checks when told to. If nobody asks, it does not check. In day-to-day business a wrong figure rarely comes with a note saying «please recalculate». Such a model is therefore only as reliable as the person who knows in advance where to have it check.
Claude Opus 5.5 declined in all three runs, named item 4 as the cause and gave the correct figure: «Do not confirm in this form», «We should not confirm CHF 41,040», «No, not in this form.»
Qwen3.5-397B against Claude Opus 5.5 (54 seconds)
Mistral-Small-4 against Claude Opus 5.5 (81 seconds)
The recordings of the first round are in the viewpoint of 4 August.
Claude Opus 5 against Opus 5.5
Because we switched the frontier side to Opus 5.5, available today, for the second round, we also get a comparison of the two Opus generations on the same nine tasks.
| Opus 5, effort «max» | Opus 5.5, effort «xhigh» | |
|---|---|---|
| Result | 9 of 9 passed | 9 of 9 passed |
| Output tokens, all 9 runs | 52,275 | 48,112 (−8%) |
| Cost, all 9 runs | 111 Swiss cents | 82 Swiss cents (−26%) |
| List price per million tokens (input / output) | USD 5 / 25 | USD 4 / 20 |
The result is the same, the cost is a quarter lower. Most of that comes from the list price: per token, Opus 5.5 is 20% cheaper than Opus 5. The rest comes from 8% fewer output tokens.
What a question costs
Cost per question in Swiss cents (Rappen), measured with the test tool's counter at the providers' list prices (US dollars converted at 0.80 Swiss francs):
| Model | Question 1 (Swiss cents) | Question 2 (Swiss cents) | Question 3 (Swiss cents) |
|---|---|---|---|
| glm5 (August) | 1.37 | 0.87 | 0.37 to 0.62 |
| llama4-maverick (August) | 0.14 | 0.073 | 0.054 |
| qwen3-vl-235b (August) | 0.35 | 0.26 | 0.22 |
| Kimi-K2.6 (October) | 0.92 | 1.06 | 0.44 |
| Qwen3.5-397B (October) | 1.49 | 0.91 | 0.65 |
| Mistral-Small-4 (October) | 0.081 | 0.11 | 0.039 |
| Claude Opus 5 (August) | 7.33 to 10.50 | 13.42 to 15.20 | 11.16 to 16.57 |
| Claude Fable 5 (August) | not tested | not tested | 34.17 |
| Claude Opus 5.5 (October) | 7.33 to 8.46 | 9.89 to 10.38 | 7.54 to 11.93 |
The models hosted in Switzerland cost a fraction per question. Mistral-Small-4 confirmed the wrong figure for 0.039 Swiss cents.
How access via the API works
Anyone using AI privately usually knows the chat subscription: a monthly fee, an app, a chat window. For this test we used the API, the providers' programming interface. This is also how business applications talk to a model.
- Pay per use: You pay per token, i.e. per word fragment, separately for input and output. There is no monthly fee. Hence the cost in Swiss cents per question.
- The model without the app: A chat app usually gives the model its own instructions and adds tools such as web search or memory. Via the API none of that applies. We tested the model itself, without additional instructions and without tools.
- Equal conditions for both sides: Our test tool sends the same question to two models at the same time, each through its provider's API. On the left a model hosted in Switzerland, on the right Claude from Anthropic.
- Models hosted in Switzerland: Swiss data centres run open models and offer them via an API. That is how we tested them.
Methodology and limits
- Same dossier, the same three questions word for word, each question in a fresh chat, no system prompt, no temperature adjustment, answer limit 32k tokens. The dossier was uploaded as a PDF; both sides received the same text.
- We ran the second round automatically in the same test tool and recorded it in real time. Answers were scored against the answer key prepared in advance.
- One documented run per pairing, Claude Opus as the constant three times per question. This is a dated snapshot, not a scientific benchmark. Single runs prove no pattern. That six models in two rounds failed question 3 the same way is a clear indication, but no proof.
- The provider still lists two of the three second-round models as beta. A different deployment of the same model may perform differently.
- The dossier has been public since 4 August. According to Hugging Face, the three second-round models were released before that date; that they know the dossier from training is practically ruled out. Claude Opus 5.5 is more recent. Its answers show no sign that it knows the dossier.
We accompany Swiss SMEs on this path: human, competent, holistic.
The AI Foundations Check shows in five minutes where your business stands. Or book an initial consultation directly.