Blog · Published · Updated
Polish LLM or API model: choose the deployment route
Short answer
You choose a local Polish model, for example Bielik or PLLuM, when documents must stay in the company and the volume is steady. You choose an API when the volume varies and there is no team to maintain a server. In our count a GPU server at €214 a month breaks even with Claude Sonnet 5.5 at about 27,000 cases a month.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- analysis with an engineer and a free prototype
- €0
- team in ChatGPT or Claude, up to 10 people
- from €1,150
- automation, agent or application for one process
- from €3,000
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
Polish models and APIs: statistics 2026
27,000
cases a month at which a €214 GPU server costs as much as Claude Sonnet 5.5 through the API, at 2,000 input and 500 output tokens per case. Scenario without the cost of upkeep.
$2 and $10
According to Anthropic's price list (2026), Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, and Haiku 5.5 costs $0.10 and $0.50.
Source: Anthropic, API pricing (2026) ↗
146
According to Bartanowicz and Jassem (LREC 2026), in a RAG system on Polish technical documentation assessed on 146 reference questions, Bielik-11B-v2.3-Instruct consistently beat PLLuM-12B-nc-chat.
83%
According to Smits and Walkowiak (EACL 2026), in RAG on the Polish Wikipedia assessed on 1,000 PolQA questions, PLLuM-12B matched the quality of 70B models with 83% less energy.
230,085
According to a Syntalith measurement (2026), Qwen3.8-27B on an RTX 3090 card recalled 3 of 3 checked facts from an input of 230,085 tokens, and the run took 563.4 seconds.
Source: Syntalith, long-context measurement of Qwen3.8-27B (2026) ↗
Figures checked: 22 October 2026
One decision to start with
| The company's situation | First option | Condition for the decision |
|---|---|---|
| Documents must stay in a defined environment | a local model | the company provides access, backups, updates and an owner |
| The number of cases changes from month to month | an API | the contract sets where and how long data is kept |
| A repetitive task with a large volume | compare a local model with an API | the fixed server cost wins once the team's work is added |
| Every case needs broad reasoning | a stronger model through an API | the trial shows the quality advantage |
| No team to maintain a server | an API or a simpler automation | the cost of upkeep does not exceed the saving |
Do not choose a model for the whole company. Choose where one process runs: invoices, the inbox, employee documents or replies to customers. One process can use rules, a local model and an API at different stages.
The cost of the whole solution
Comparing the price of one request with the price of a server gives the wrong picture. Anthropic's price list and OpenAI's price list separate input tokens, output tokens and tools, and on the server side backups, updates and upkeep come on top.
- API, Claude Sonnet 5.5
- 10,000 × €0.008 = €80
- Hetzner GEX45 GPU server
- €214
- Break-even point
- €214 ÷ €0.008 = about 26,600 cases
Below the threshold the API is cheaper, above it the server, before the cost of upkeep
The number of tokens per case is the scenario's assumption. Anthropic's rates for Claude Sonnet 5.5, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026. Price of the GEX45 with 24 GB of GPU memory from Hetzner's website on 22 October 2026, without the one-off setup fee of €209. Server upkeep, backups and updates raise the threshold.
At Syntalith, AI automations, AI apps and AI agents start at €3,000 net, and typical full rollouts fall between €3,000 and €17,500 net. We price maintenance and infrastructure separately.
Data, the contract and responsibility
- The place of processing and any transfer outside the EEA
- The data processing agreement and the list of subprocessors
- How long questions, files and records are kept
- The rules for using data to improve the service
- Data export and ending the relationship
- The cost limit and how the system behaves once it is exceeded
Your own server limits the flow of data to a chosen environment, and permissions, encryption, backups, deleting documents and activity records remain to be handled. Running locally is no certificate of GDPR compliance, and an outside service can be the right choice when the company has a contract and controls the flow of data.
A quality test on your own cases
Prepare 30 to 100 real, anonymised examples from one process, with the expected result, the sources and cases of missing data. Run the same set on a local model, one or two models through an API and, if possible, on a rules-based solution. According to Bartanowicz and Jassem (LREC 2026), on 146 questions about technical documentation Bielik-11B beat PLLuM-12B. According to Smits and Walkowiak (EACL 2026), PLLuM-12B matched 70B models on the Polish Wikipedia. The results concern their datasets, so a test on yours decides.
Varying volume, data may go to a provider
an API with cost and data limits
from €3,000
A large steady volume, documents stay in the company
a local model with upkeep
from €3,000 plus the server
Some cases sensitive, some difficult
a mixed option with a written data flow
priced after the review
Rules are enough
a form or an automation without a model
from €3,000
A mixed option works when the routing rule is written down before launch: the app recognises how sensitive a document is, personal data stays with the local model, and the API receives only the part needed for the difficult analysis. With Bielik, first check the terms of use: version 1.1 of 31 July 2025 forbids using the model to process any personal data, so for such data compare a PLLuM option with a commercial licence or clarify the matter with the SpeakLeash foundation. We roll out local models as part of private LLMs. For the free process review, prepare the number of cases, the type of data, the cost of today's work and the effect of an error.
Worked example · Scenario
How we count the payback: a local model
We run the same count for your process in the free process review.
€4,363
a month for the work on 500 questions
EU cloud version payback: 0.8 mo.
Our assumptions are 15 minutes to find a passage in case files and templates, and checking in 15% of the old time. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer in the EU cloud version, at the ECB rate of 1.1205 USD per euro on 22 October 2026. We count the payback for an app with a model in the EU cloud (from €3,000 net). We price the model on your server after a call, and the cost of people's work stays the same.
See the full calculation →
