Pricing and benchmark ยท 2026-09-19
EVREN quota and credit system: free, but metered
The EVREN LLM API is free until 1 November; but your account carries a token cap over a sliding window, a credit balance and a per-request charge state. We cover the /v1/quota fields, the cooldown on consecutive requests and planning for after 1 November.
How free is free?
According to SAYZEK's announcement the EVREN LLM Inference Service is free until 1 November. No invoice, but a meter: even during the free period the API schema keeps a token cap, a credit balance and a per-request charge state for every account.
We found no public price list for what happens after 1 November. Before tying a workflow to this service, it is worth having a plan for after that date.
What /v1/quota tells you
The /v1/quota endpoint returns your account's state. The cap works over a sliding window: it does not reset at a fixed hour, and room frees up as the oldest consumption rolls off. The level field moves to warn at 75 percent and to critical at 90 percent.
| Field | Meaning |
|---|---|
| used_tokens | Tokens consumed in the sliding window |
| cap | Token cap for the window |
| window_minutes | Window length (minutes) |
| reset_at | When the oldest consumption bucket rolls off |
| usage_ratio | used_tokens / cap, between 0 and 1 |
| level | ok; warn (75 percent and above); critical (90 percent and above); unknown |
| held_cr | Credit reserved by in-flight requests, not yet settled |
| remaining_cr | Available credit balance |
Where the credit comes from
The EVREN platform is built on a credit model. According to university announcements, users earn credit by uploading or labelling datasets and spend it on access to high-capacity GPU resources.
We found no public text explaining how the LLM service's credit relates to that model. But the schema keeps credit explicitly: remaining_cr shows the available balance and held_cr the amount reserved by in-flight requests.
Charge state per request
/v1/requests/{request_id} returns, for each request, whether the charge is final (charge_terminal), whether billing is pending (is_billing_pending), the credit collected (collected_cr) and the credit outstanding (outstanding_cr).
Seeing an 'outstanding credit' field per request on a free service looks odd at first. What it says is simple: the billing machinery is in place, and the price is zero for now.
Consecutive requests and the cooldown
A community-written OpenCode integration notes that EVREN applies a cooldown of about 5 seconds on consecutive requests and returns 429 when it does. The same source recommends sequential work over parallel sub-tasks, and a small helper model to take load off the main one.
That figure is not a limit EVREN has published; it is one integration's observation. Tools that start several requests at once, such as coding agents, should expect 429 responses and handle them. The example below checks the quota before the request and backs off with growing waits on 429.
Quota check and back-off on 429
import os, time, requests
BASE = "https://evren-llmapi.ssyz.org.tr/v1"
HEADERS = {"X-API-Key": os.environ["EVREN_API_KEY"]} # evren_llm_...
def quota_ok() -> bool:
q = requests.get(f"{BASE}/quota", headers=HEADERS, timeout=10).json()
# level: ok / warn (>= 75%) / critical (>= 90%) / unknown
return q.get("level") in ("ok", "unknown")
def chat(message: str, attempts: int = 5) -> dict:
wait = 5.0
for _ in range(attempts):
r = requests.post(
f"{BASE}/chat/completions",
headers=HEADERS,
json={"model": "glm-5.3", "messages": [{"role": "user", "content": message}]},
timeout=120,
)
if r.status_code != 429:
r.raise_for_status()
return r.json()
time.sleep(wait) # back off on consecutive requests; double every attempt
wait *= 2
raise RuntimeError("Could not get past the quota or cooldown")
if quota_ok():
print(chat("Hello")["choices"][0]["message"]["content"])
Consumption in auto mode
When you write auto in the model field, EVREN picks the model that serves the request. In the documentation's example, the response to an auto request comes from zai/glm-5.3-fp8.
Because the quota is counted in tokens, the same prompt going to different models means different token consumption and different response behaviour. If you want repeatable results and predictable quota use, name the model explicitly.
Planning for production
The most expensive moment of putting a free API into production is the day the free period ends. A few precautions keep that day quiet.
- Retry 429 responses with growing waits; do not start many requests at once.
- Check remaining capacity with /v1/quota before batch jobs; reduce load when level reads warn.
- Keep model names in configuration separate from code, so switching services is a one-line change.
- Try a second provider that works with the same OpenAI SDK now, not later.
- Do not tie a critical workflow to the free period alone.
Frequently asked questions
Until when is EVREN free?
According to SAYZEK's announcement the EVREN LLM Inference Service is free until 1 November.
Will EVREN be paid after 1 November?
We found no public price list for that. The API schema carries a credit balance and per-request charge fields, so the billing machinery is in place.
How do I see my EVREN quota?
GET /v1/quota returns consumption in the sliding window, the cap, the window length, the reset time, the usage ratio and your credit balance.
I am getting 429 from EVREN. What should I do?
Queue your requests and retry with growing waits. Check your quota with /v1/quota. A community integration observed a cooldown of about 5 seconds on consecutive requests; that is not a figure EVREN has published.
How is EVREN credit earned?
According to university announcements, credit on the platform is earned by uploading or labelling datasets. We found no public text on how the LLM service's credit relates to that model.