openference.com: the pricing is reasonable. Then it burns 10 minutes 39 seconds and dies with an error.
Verdict
Pricing per request instead of per token is a fine idea and I have nothing against it. The problem is that the request does not come back — 10 minutes 39 seconds of thinking, then the stream was interrupted. Verdict: do not pay them.
That is the whole review. The models are real and the decode speed is normal. The money leaves, the time leaves, and nothing comes back.
The pricing itself is not the problem
It bills per request, not per token. One API call is one request whether it carries two thousand tokens or two hundred thousand, so it is cheap for huge contexts and expensive for a flood of tiny calls. Lite is $15 for 400 requests per 5 hours and 7,500 a week; Pro is $30 for 800 and 13,000; Max is $60 for 1,600 and 26,000.
- The weekly cap binds before the 5-hour one at every tier. 1,600 per 5 hours is 7,680 a day if you run continuously, but 26,000 a week is 3,714 a day, so the week is what you hit.
- The $15 Lite tier excludes the Kimi models.
- Max was sold out at check time.
- Going over does not cut you off — usage continues from credits at catalog pricing with your plan discount.
So why the warning
None of the above is the problem. The pricing is not the scam. Taking the money and not returning a result is the scam. Default thinking is set to max and it is hidden, so a request that looks dead is actually reasoning where you cannot see it. One Mario 1-1 prompt ran 97 seconds with zero visible body text before I aborted it.
Measured on the same key: an 18,285-token input with a 9.6-second TTFT returned 67 output tokens, of which 58 were reasoning. DeepSeek-V4-Flash-0731 answered the same kind of request in 19.7 seconds for 1,468 tokens.
The models are fine. The operation is not. And what you pay for is the operation, not the models.
The full measurement set is in the scam section, under Openference.
Questions
- Is the pricing itself the problem?
- No. Billing per request is a normal model and the caps are stated plainly. The problem is that a request you paid for does not return a result.
- Can I get a refund?
- Subscriptions are non-refundable under their own terms. Annual billing is about 17% cheaper on the same non-refundable basis.
- Are the models bad?
- The models are real and the decode speed is normal. On the same key, DeepSeek-V4-Flash-0731 answered the same kind of request in 19.7 seconds for 1,468 tokens.
The terms, as verified
Priced per requestRead from the vendor's own pricing page. The review above is a judgement about these numbers; it does not restate them, so the two cannot drift apart.
- $15.00
$15/month · Lite
400 requests per 5 hours, 7,500 a week
What actually limits you: The weekly cap, not the 5-hour one
Excludes Kimi models
- $30.00
$30/month · Pro
800 requests per 5 hours, 13,000 a week
What actually limits you: The weekly cap, not the 5-hour one
Includes Kimi models
- $60.00
$60/month · Max
1,600 requests per 5 hours, 26,000 a week
What actually limits you: The weekly cap, not the 5-hour one
Sold out when checked
Models GLM, DeepSeek, Qwen, Llama — Lite excludes Kimi
Worth knowing
- The weekly cap binds before the 5-hour one at every tier. 1,600 requests per 5 hours is 7,680 a day if you run continuously, but 26,000 a week is 3,714 a day, so the week is what you hit. Lite is 1,920 a day against 1,071. Pro is 3,840 against 1,857.
- One request is one API call regardless of size. Send two-hundred-thousand-token contexts and this is the best value on the page; send thousands of small calls and it is the worst.
- The $15 Lite tier excludes the Kimi models. If Kimi is why you are here, you need the $30 tier.
- Max was sold out at check time — every tier is capacity-limited.
- Hitting the quota does not cut you off. Usage continues from credits at catalog pricing with your plan discount, so a busy week costs money rather than downtime.
- Provider errors and capacity failures do not count against your allowance.
- Subscriptions are non-refundable. Annual billing is about 17% off.
Source: openference.comChecked 2026-09-20