Tapix AI

Pay-as-you-go LLM API: what you are actually billed for

A seat subscription prices the average user. If you call an API twice a week, you subsidise the team calling it eight hours a day. Per-token billing is the alternative - but "pay per token" hides three details that decide the bill.

1. Input and output are priced separately

Every serious provider quotes two numbers. Output usually costs several times input, so the shape of your prompts matters as much as the model choice: a support bot that writes two paragraphs costs far more than a classifier that returns one word, on the same model.

2. Long context is billed even when it is repeated

If you send the same 20k-token system prompt on every call, you pay for it on every call unless the provider supports prompt caching and you use it. Ask any provider which caching mechanism they expose; if the answer is vague, assume you are paying full price for those tokens.

3. Failed requests should not be billed

A 400 from a malformed request costs you latency, not money. A 429 costs you a retry, which does cost money. Our rule: rejected requests are not charged, and a request that dies mid-stream is billed for what the upstream actually reported, not for the tokens we hoped to generate.

Our price for the models people actually use

ModelOfficial (in/out per 1M)Tapix AI (in/out per 1M)Below list
gpt-6.1-terra$1 / $5$0.4299 / $1.4557%
gpt-5.5$5 / $30$3.5 / $2130%
claude-sonnet-5$2 / $10$1.4 / $730%
gemini-3.6-flash$0.75 / $3.75$0.525 / $2.6330%
kimi-k3$3 / $15$2.4 / $1220%

Practical advice

Prices above come from our live catalogue. The pricing page is always the source of truth.

Get $1 free creditSee all models

Related reading

Last updated: 2026-11-21最后更新:2026-11-21