AI cost calculator.

Use this calculator for anything that calls a model through an API: a document extraction pipeline, a support assistant, an agent or an internal tool. For per-seat subscriptions such as ChatGPT Business or Microsoft Copilot, see AI tools pricing for New Zealand businesses instead. Everything runs in your browser.

What each input means

The default rate of 1.77 is rounded from the European Central Bank’s euro reference rates for 16 October 2026: US$1.1253 and NZ$1.9868 per euro, which gives NZ$1.7656 per US dollar (ECB). Your bank or card provider will add its own margin.

The prices behind the calculator

USD per million tokens, standard tier, as listed on 19 October 2026 by OpenAI, Anthropic and Google.

How the calculation works

In plain words: the input that is not cached is charged at the input price, the cached input at the cached price, and the output at the output price. Add them, divide by a million, halve it for batch, then convert to NZ dollars.

Cost per request (NZ$) = (input tokens × (1 − cache share) × input price + input tokens × cache share × cached price + output tokens × output price) ÷ 1,000,000 × 0.5 if batch × exchange rate.

Monthly cost is cost per request × requests per month; yearly is monthly × 12.

Worked example

Illustrative example, using the defaults (GPT-6.1 Sol, 5,000 requests, 6,000 input and 800 output tokens, 40% cached, rate 1.77, no batch):

1. Uncached input: 3,600 tokens × US$2.00 per million = US$0.00720
2. Cached input: 2,400 tokens × US$0.10 per million = US$0.00024
3. Output: 800 tokens × US$10.00 per million = US$0.00800
4. Total per request: US$0.01544 × 1.77 = NZ$0.027
5. Per month: × 5,000 = NZ$137; per year NZ$1,640

Output is 52% of the cost from 12% of the tokens. The same workload on every model:

Which inputs move the cost most

Starting from the GPT-6.1 Sol default of NZ$137 a month:

Model choice changes cost about a hundredfold; output length and volume come next. The cheapest model is only cheaper if it does the job: test candidates on your own cases first, as described in AI model selection.

Common mistakes and when the estimate is unreliable

• Cache writes are excluded. OpenAI charges 1.25 times the input price to write to the cache on GPT-5.6 and later models (OpenAI), and Anthropic charges 1.25 times for five-minute cache writes.
• Long prompts cost more. OpenAI bills prompts over 272K input tokens at higher rates, and Claude Haiku 5.5 has a higher price above 100K tokens.
• Agents make many calls. One task can take several requests, each resending the conversation so far. Count calls, not tasks.
• Tool fees are extra. Web search costs US$10 per 1,000 calls on both OpenAI and Anthropic, plus tokens.
• Token counts differ by provider. Anthropic says Claude’s newer tokenizer produces about 30% more tokens for the same text than older Claude models.
• Prices change. Gemini 4 Argon doubles on 1 December 2026 and Gemini 3.8 Flash on 1 January 2027.
• GST and card fees are excluded. From New Zealand, Anthropic’s pricing page shows “Prices include 15% GST” beside its USD figures (Anthropic); confirm on a real invoice.

How to get real token counts

1. Run 20 real requests. Use your actual prompts and documents in a test account.
2. Read the usage figures. Every provider returns input, cached and output token counts with each response and in its console.
3. Use the 90th percentile, not the average. Long documents and long answers drive the bill.
4. Count calls per task. Especially for agents and multi-step workflows.
5. Add headroom. Retries, testing and growth usually add to the estimate; put the total into the AI ROI calculator as part of monthly running cost.

Next step

Our engineering guide to managing LLM costs covers caching, routing to cheaper models and spending limits in production. If you are weighing per-token pricing against flat subscriptions, read choosing between an API and a subscription.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.