Updated 2026-09-20

Openference: ten minutes of thinking, then an error. I am calling that a quasi-scam

Thinking ran for 10m 39s and the request ended with "The model provider's stream was interrupted." I paid and got nothing back. The 85-second delay I measured earlier was not slowness — it was a preview.

Ten minutes thirty-nine seconds

One request against GLM 5.3 Flash. This is what came back.

Thought: 10m 39s

Error: The model provider's stream was interrupted.
Please retry.

Ten minutes and thirty-nine seconds of thinking, then the stream died. No result. The money went out and nothing came back. If "please retry" is the answer, what happened to the ten minutes?

Once could be bad luck. But everything measured below points the same way. This is a pattern, not an accident.

What I measured before that

Streamed with reasoningeffort=low and maxtokens=400, asking for a 40-line answer.

GLM-5.3-Flash            85.3s   4294 tok   ~50 tps
  thinking 14,616 chars -> body 3,108 chars

GLM-5.3                  57.3s   5021 tok   ~88 tps
  same pattern

DeepSeek-V4-Flash-0731   19.7s   1468 tok   ~75 tps
  body starts far sooner (same key)

I asked Flash for forty lines and it poured out fourteen thousand characters of thinking before any of it appeared. Decode itself runs around fifty tokens a second, so the speed is not dead. But the screen shows you none of it — and this time the body never arrived at all.

A real session is worse

The default thinking is max. The config reads openference/glm-5.3-flash:max.
Thinking blocks are hidden.
A one-line question: 18,285 input tokens, 9.6s to first token, 67 output tokens of which 58 were reasoning.
A Mario 1-1 request ran 97 seconds with zero visible body text. Thinking churned, then I aborted.
And then, after 10m 39s, the stream was interrupted.

Is it a scam

It is not a wallet-drainer. The models are real and decode speed is normal. But it is a quasi-scam: it takes the money, eats the time, and delivers no result. What those three things mean to someone who paid does not need spelling out.

Calling it "just slow" does not survive a request that returns nothing. Slow you can wait out. Zero is zero however long you wait.

So how do you use it

That is why this is in the scam section rather than the reviews. It earned a warning.
If you insist on using it, drop thinking to low. Leave it at max and you get the above.
Otherwise use DeepSeek-V4-Flash-0731 on the same key. 19.7 seconds against 85.3, and at least it does not die.
Worth checking whether a request that returned nothing is refundable. Ten minutes thirty-nine seconds is not free.

The models are not bad. The operation is. And what you pay for is the operation, not the model.