Qwen3.8 27B on a prepaid API.
One endpoint, one price per model, prepaid by the token. Point any OpenAI-compatible client at it and spend what you use — no subscription, no seats, no minimum.
First 100 accounts. Email address only — a card is needed the first time you top up.
curl https://api.tiyuvta.ai/v1/chat/completions \
-H "Authorization: Bearer $TIYUVTA_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "qwen/qwen3.8-27b",
"messages": [{"role": "user", "content": "Hello"}]}'At an 8:1 input-to-output mix. All input: 26.3M tokens. All output: 3.8M. Billing is per token, never per request or per month.
Measured performance
measured 2026-08-15Protocol: 984-prompt agentic bench file, greedy, streaming, 128 max output tokens, 3 repetitions per cell, medians reported. No shed requests and no errors across all cells. The complete table, including the limits, is on the model card.
Price
per 1M tokens · checked 2026-08-15| Endpoint | Input | Output |
|---|---|---|
| tiyuvta | $0.38 | $2.60 |
| The only other public Qwen3.8 27B endpoint | $0.45 | $3.20 |
16% below on input, 19% below on output, with cached input priced. Read from the public routing table on 2026-08-15; prices there change without notice. Step-3.7-Flash is in bring-up and will be listed, with its own price, once its correctness gates pass.
Built for agent traffic
Cached continuations
When a request continues a conversation the endpoint already served, the shared prefix bills at $0.12 per million instead of $0.38. First token stayed at 0.13s on a 5,690-token second turn.
Tool calls that parse
Function calling, JSON mode and JSON schema are constrained during decoding, so the schema you send is the schema you get back. Streaming responses include token usage.
262,144-token context
The model's native window, served in full. Reasoning effort is a per-request parameter, not a plan tier.
Gated before it serves
Serving runs NVFP4 4-bit weights for speed. Every build is checked against an eager reference and the vendor's higher-precision release first, and speculative decoding is verified hash-identical to plain decoding. The gate list.
Four steps to the first token
- Sign in with an email address A one-time link. No password, no sales call, no waitlist.
- Claim $10 of trial credit Granted immediately to the first 100 accounts. No card.
- Create a key and change one line Set
base_urltohttps://api.tiyuvta.ai/v1and keep your existing client. - Top up when you need more Credit packs from $5. First top-up is 15% off the first $15. Auto top-up is available and off by default.
from openai import OpenAI
client = OpenAI(
base_url="https://api.tiyuvta.ai/v1",
api_key=os.environ["TIYUVTA_KEY"],
)
r = client.chat.completions.create(
model="qwen/qwen3.8-27b",
messages=[{"role": "user", "content": "hi"}],
)The same call works from the Node SDK, the Vercel AI SDK, LangChain, LlamaIndex, Cline, Aider or curl. It is the Chat Completions API.
Limits, stated up front
A short roster
Qwen3.8 27B is live today , with Step-3.7-Flash in bring-up.
Chat completions only — no embeddings, no batch API, no fine-tuning.
GET /v1/models is the authoritative list.
Caching covers continuations, not fan-out
A cached prefix is reused when a request continues a previous one — the same conversation, one turn later. Fan-out over one shared document with different questions does not hit the cache on this model architecture.
Failed requests are not billed
A request that fails with our error consumes no credit. Unused credit stays yours and is refundable within 30 days — billing terms.
Paid traffic is never used for training
Requests paid from purchased credit are never retained for training or evaluation. Requests paid from free trial credit may be. Both rules are in the privacy terms.
$10 of credit, one email, no card.
Limited to 100 accounts. This page states when the offer is closed.