Home / Method
Method and sources
Two kinds of number appear on this site: official or published figures, each with a source, and estimates, each labelled estimate. Nothing else.
The formulas
Person = minutes per task estimate ÷ 60 × hourly rate.
Agent = (input tokens × input price + output tokens × output price) ÷ 1,000,000 ÷ USD per EUR, plus minutes to check the output estimate ÷ 60 × the same hourly rate.
Tokens estimate: input = calls × (3,000 overhead + words read × 4/3); output = 3 × words written × 4/3.
Break-even = setup hours estimate × professional wage ÷ (person − agent), rounded up.
4/3 tokens per word follows OpenAI's rule of thumb that a token is about three-quarters of an English word (source). Other languages usually need more tokens. The 3,000-token overhead per call (system prompt and tool definitions) and the ×3 output allowance (reasoning and tool-call tokens, billed as output) are our estimates.
Hourly rates
Wage, by occupation. Mean gross hourly earnings, Structure of Earnings Survey 2022, enterprises with 10 or more employees, both sexes, all ages. Each task maps to one ISCO-08 major group. Excludes employer social contributions. Dataset updated 2026-02-09. Eurostat, Structure of Earnings Survey 2022: mean hourly earnings by occupation (earn_ses22_14) (doi:10.2908/EARN_SES22_14).
Full labour cost, all jobs. Hourly labour cost (wages plus non-wage costs), industry, construction and services except public administration, 2025. Not split by occupation. Belgium has no 2025 value yet, so we show 2024. Austria 2025 is flagged provisional. Dataset updated 2026-04-23. Eurostat, Labour cost levels by NACE Rev. 2 activity (lc_lci_lev) (doi:10.2908/LC_LCI_LEV).
The 2022 wages are the newest occupation-level figures. Labour costs have risen since: the EU average hourly labour cost moved from €31.9 in 2023 to €34.9 in 2025 (lc_lci_lev). We do not adjust the 2022 wages.
Task estimates
Minutes per task and check minutes are our estimates for a trained worker doing a routine case. They are not measured. Group = ISCO-08 major group used for the wage.
| Task | Group | Person min | Check min | Words in / out | Calls | Tokens in / out |
|---|---|---|---|---|---|---|
| Reply to a customer support email | OC4 | 8 | 1 | 600 / 150 | 3 | 11,400 / 600 |
| Answer a customer chat | OC5 | 5 | 0.5 | 800 / 120 | 3 | 12,200 / 480 |
| Enter a supplier invoice | OC4 | 5 | 0.5 | 400 / 120 | 2 | 7,067 / 480 |
| Check an expense claim | OC4 | 4 | 0.5 | 1,200 / 80 | 2 | 9,200 / 320 |
| Reconcile 50 bank lines | OC4 | 30 | 4 | 4,000 / 400 | 4 | 33,333 / 1,600 |
| Triage 20 shared-inbox emails | OC4 | 10 | 1 | 3,000 / 200 | 3 | 21,000 / 800 |
| Write minutes for a 30-minute meeting | OC4 | 25 | 4 | 4,500 / 450 | 2 | 18,000 / 1,800 |
| Clean a 200-row contact list | OC4 | 30 | 4 | 3,000 / 3,000 | 3 | 21,000 / 12,000 |
| Research one B2B lead | OC3 | 20 | 3 | 6,000 / 250 | 8 | 88,000 / 1,000 |
| Log a 15-minute sales call | OC3 | 10 | 1 | 2,300 / 200 | 2 | 12,133 / 800 |
| Screen one CV | OC3 | 6 | 1 | 1,500 / 120 | 1 | 5,000 / 480 |
| Draft a social media post | OC3 | 15 | 2 | 300 / 80 | 2 | 6,800 / 320 |
| Write a help-centre article | OC3 | 40 | 6 | 1,500 / 600 | 2 | 10,000 / 2,400 |
| Write a product description | OC2 | 20 | 3 | 300 / 200 | 2 | 6,800 / 800 |
| Translate one page | OC2 | 50 | 10 | 300 / 400 | 2 | 6,800 / 1,600 |
| Write a job ad | OC2 | 30 | 5 | 600 / 450 | 2 | 7,600 / 1,800 |
| Summarise a 10-page contract | OC2 | 60 | 10 | 5,000 / 500 | 2 | 19,333 / 2,000 |
| Triage a bug report | OC2 | 15 | 3 | 2,500 / 200 | 5 | 31,667 / 800 |
| Review a small pull request | OC2 | 30 | 5 | 4,000 / 400 | 4 | 33,333 / 1,600 |
| Write a SQL report query | OC2 | 30 | 5 | 2,000 / 300 | 4 | 22,667 / 1,200 |
AI model prices
USD per million tokens, standard tier, read 2026-10-28. Converted at $1 = €0.8969 (ECB reference rate, 2026-10-27: €1 = $1.115, source). Default model: Claude Sonnet 5.5.
| Model | Provider | $ in / 1M | $ out / 1M | Source |
|---|---|---|---|---|
| GPT-6 Luna | OpenAI | 0.10 | 0.50 | OpenAI pricing |
| Claude Haiku 5.5 | Anthropic | 0.10 | 0.50 | Anthropic pricing — Price for prompts up to 100k tokens. |
| Mistral Large | Mistral AI | 0.50 | 1.50 | Mistral AI pricing — Taken from the worked example in the pricing page FAQ; the per-model table did not render for us. Verify before relying on it. |
| Gemini 3.8 Flash | 0.75 | 3.75 | Google pricing — Launch price until 31 Dec 2026; then $1.50 in / $7.50 out per the same page. | |
| GPT-6.1 Terra | OpenAI | 1.00 | 5.00 | OpenAI pricing |
| Claude Sonnet 5.5 | Anthropic | 2.00 | 10.00 | Anthropic pricing |
| GPT-6.1 Sol | OpenAI | 2.00 | 10.00 | OpenAI pricing |
| Claude Opus 5.5 | Anthropic | 4.00 | 20.00 | Anthropic pricing |
Running your own model
For context, the cheapest EU on-demand GPU hour in the gpueuro.com table (checked 27 Oct 2026):
| GPU | € / hour | Cheapest of | Provider |
|---|---|---|---|
| H100 | €1.89 | 5 EU offers | Seeweb |
| L40S | €0.85 | 5 EU offers | Seeweb |
| L4 | €0.38 | 3 EU offers | Seeweb |
We do not turn these into a per-task cost: that needs a measured throughput for a given model, which we do not have.
What this does not cover
- Software, tooling, monitoring and integration costs beyond the setup-hours estimate.
- Error rates. An agent that is wrong more often needs longer checks; raise the check minutes.
- Batch or cached-input discounts, which lower agent cost.
- Data-residency surcharges, for example OpenAI's 10% uplift on newer models.
Sources
- Eurostat, Structure of Earnings Survey 2022: mean hourly earnings by occupation (earn_ses22_14) · doi:10.2908/EARN_SES22_14
- Eurostat, Labour cost levels by NACE Rev. 2 activity (lc_lci_lev) · doi:10.2908/LC_LCI_LEV
- Eurostat news, hourly labour costs 2025 (31 Mar 2026)
- OpenAI API pricing
- Anthropic Claude pricing
- Google Gemini API pricing (page updated 27 Oct 2026)
- Mistral AI pricing (FAQ example)
- European Central Bank, euro reference rate USD, 27 Oct 2026
- OpenAI Help, Understanding and counting tokens
- gpueuro.com EU GPU price table (checked 27 Oct 2026)