A decision is only as good as the number behind it. Here you get both halves: the computation itself and a plain-English explanation of what drives it.
Every figure on this page can be reproduced by hand. Your inputs stay on your device — there is no server side to this tool at all.
The purpose, stated plainly
Typical visitors are users in Johannesburg, Cape Town, Durban and Pretoria comparing offers or checking a bill, so the worked examples below use sensible local values rather than abstract numbers. That's the difference between an informed launch decision and an end-of-month bill shock. Instead of guessing what integrating a LLM or embeddings will cost you, the tool computes exact monthly forecasts drawing on provider rates, token volume and model type.
AI tokens per second calculator turns AI costs from a vague subscription line into concrete numbers. For South Africa, use the units printed beside each field and do not assume that a familiar name means the same measurement system.
That means the output can be used directly with an accountant, lender or employer in South Africa without converting anything.
Who this is for
- Local notes: Amounts use rand, metric units and SARS defaults where applicable.
- VAT / sales tax: VAT is 15% on most goods and services. Basic foods (brown bread, rice, maize meal, dry beans, etc.) are zero-rated. Compulsory registration above R1 million turnover.
- Electrical supply: 230 V / 50 Hz, Types D, M and N plugs. Load-shedding still occasionally scheduled.
- Units of quantify: Fully metric.
- Currency & format: South African rand (ZAR, R).
The thirty-second guide
Fourth: comparing raw price without factoring throughput (tokens/second). The classic error is counting only input tokens — output typically costs 3–5× more per token.
For long contexts use prompt caching (Anthropic and OpenAI both offer up to 90% discount). Ship compressed prompts; every token bills. Third: forgetting fine-tuning or fine-inference surcharges. If you resend the full conversation on every request, tokens grow quadratically.
Second: ignoring context window. Monitor monthly with a live dashboard, not at billing cycle end. Cache repeated prompts — this often shaves 50–80% off costs.
This is the point where most guides stop — which is a pity, because the next part is where the number becomes real-world.
How the maths works
Example: at 2000 requests per month averaging 37.5 input tokens and 100 output tokens, AI tokens per second calculator returns a concrete monthly figure in R. AI costs grow fast and quietly. Downgrade from GPT-4o to GPT-4o-mini and costs in most cases drop 10× at ~90% of quality for straightforward tasks.
Model three scenarios — 1k, 10k and 100k requests per month — to see where costs blow through budget and where switching to a smaller model or self-hosting starts to pay. South African costs move with tariffs and levies, so AI tokens per second calculator uses SARS-aligned rates and rand pricing. Compute before launching an AI feature, before scaling to a wider audience, or when weighing a model swap.
Common slip-ups
For spiky traffic, commercial APIs win on pay-per-use. Recalculate after changing one input at a time; this shows which assumption controls the calculated outcome and makes an unexpected answer easier to diagnose.
The values entered here are Tokens, Seconds. Above 100k requests per month at steady traffic, Llama or Mistral on an A100 GPU frequently runs 2–3× cheaper than a commercial API.
Small habits, better numbers
Depends on the task. Vision — Gemini or GPT-4o.
Keep the original measurement, invoice, specification or log beside the calculation so that a misplaced decimal point can be found. Every figure above applies to South Africa specifically and is reviewed when the rules change.
When guessing stops being enough
Simple summarisation — Gemini Flash or GPT-4o-mini. Run three benchmarks and contrast quality/cost.
Complex code — Claude Sonnet or GPT-4o. Start by checking the input fields rather than the final number.
Where this leaves you
There is no prize for doing this maths by hand. When the assumptions change, come back — the AI tokens per second calculator does not mind.
South Africa-specific questions
Is this AI tokens per second calculator valid for South Africa?
Yes. It uses South African rand (ZAR, R) and fully metric, and any tax or contribution reference follows SARS rules.
Which units and number format are used?
Amounts display as R1 234,56 and quantities in fully metric — the official convention in South Africa.
Which pension and tax model is assumed?
Provident/pension funds + Older Persons Grant. The tax framework is income tax 18–45%, VAT 15%.