Cheapest GPT API: what actually drives the price
Every provider quotes the same two numbers - price per million input tokens and per million output tokens - but the bill you get at the end of the month depends on three things most comparison pages ignore: output length, retries, and whether you are paying for a seat you don't use.
The three numbers that decide your bill
- Output tokens cost 4-8x input tokens. A model quoted at $0.50 in / $3.00 out is not "a $0.50 model". Long answers are what hurt.
- Retries are invisible until they aren't. A 5% failure rate on 100k requests is 5,000 extra calls you pay for twice.
- Subscriptions price the average user. If you use AI twice a week, you subsidise the people using it eight hours a day.
Official list vs what we charge
| Model | Official (in/out) | Tapix AI (in/out) | Saving |
|---|---|---|---|
claude-haiku-5-5 | $0.10 / $0.50 | $0.07 / $0.35 | 30% |
gemini-3-flash | $0.50 / $3.00 | $0.35 / $2.10 | 30% |
gpt-6.1-terra | $1.00 / $5.00 | $0.43 / $1.45 | 57-71% |
claude-sonnet-5 | $2.00 / $10.00 | $1.40 / $7.00 | 30% |
If a provider quotes a price far below official list, ask which account tier the traffic is served from. We publish the tier for every model on the pricing page.
Live prices from our catalogue
These are read from our live catalogue, so they cannot go stale. Prices are USD per 1M tokens.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| — | — | — |
How to switch in one line
If your code already speaks the OpenAI format, change two values:
base_url = https://tapixai.com/v1 api_key = sk-fg-...
Nothing else changes: same SDK, same request shape, same streaming. Model IDs differ, so check the model list.
When paying per token is worse
If you run a steady 8-hour-a-day workload on one flagship model, a flat subscription can still be cheaper. Pay-per-token wins when your usage is bursty, when you route easy traffic to cheap models, or when you need several models side by side.
Try it free with $1 credit See full pricing