Every model.
One chat.
One bill.
Compare the answer, price, and limits before sending. Continue in code through an OpenAI-compatible API with the same catalog.
Pay per token · 1 credit = $1 · no subscription
Publishing a model?Portal for AI companies
Using published prices, no discounts or rounding:
For RAG input cost dominates: Llama 3 8B is 1.5× cheaper than Mistral 7B.
A catalog that shows the price.
Models for text, embeddings, images, speech, and synthesis. Each one has a billing unit and a page with its licence, version, and limits.
Full catalogCatalog composition and prices change. The model page is authoritative.
The bill appears before the request.
Before sending, you see the price per 1M input and output tokens. After the answer, the charge is recalculated from actual usage. If a request fails, the reserve returns to your balance.
Try it in chatCalculation
Mistral 7B Instruct
- Input
- 1,000,000 × $0.15 / 1M
- $0.15
- Output
- 200,000 × $0.45 / 1M
- $0.09
- Reserve before the answer
- $0.24
- Charged after the answer
- $0.24
- Total
- $0.24
Switch models mid-thread
Start fast, continue on a reasoning model — context stays.
Answers stream in
No waiting for the whole generation.
Cost sits in the thread
Model, tokens, and charge under every answer.
The same catalog in your code.
OpenAI SDK, streamed responses, keys with scopes and limits. One endpoint for every model in the catalog.
- OpenAI-compatible
- SSE streaming
- per-key limits
- spend ledger
from openai import OpenAI
client = OpenAI(
base_url="https://novastep.tech/v1",
api_key="nh_live_...",
)
stream = client.chat.completions.create(
model="neuro-7b-instruct",
messages=[{"role": "user", "content": "Кратко: что такое RAG?"}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")Pay for usage, not for a seat.
No subscription. Top up a credit balance at $1 per credit. A reserve is placed before the request; the charge follows the answer.
- LLM
- 1M input / 1M output tokens
- Embeddings
- 1M tokens
- Images
- per image
- ASR
- per minute of audio
- TTS
- per 1,000 characters
Requests × (input × input price + output × output price) ÷ 1,000,000. Retries and cache excluded.
Monthly estimate
$5.70
≈ $0.000570 per request
Same workload
- Mistral 7B Instruct$5.70
- Llama 3 8B Instruct$3.80
1 credit = $1 · catalog prices; the model page is authoritative
The short version
Do I need a subscription?
No. The account is free and requests are charged to your balance at the selected model's price. No card is needed until you top up.
Is chat priced differently from the API?
No. It is one catalog and one tariff. Chat shows the charge under the answer; the API records it in usage.
What happens when a model fails?
The request returns an error code and the reserved amount goes back to your balance. The request_id finds it in the log.
Can I delete my history?
Yes. Conversations live in your account and can be deleted at any time. We do not train models on your conversations without explicit consent.
Where do I start?
Create an account, choose a model in the catalog, and send a first request in chat. For code, create a key in the console.
Start with one request.
The account is free. After that, you see the price, the answer, and the charge for every request.