DEV Community

Davis George
Davis George

Posted on

Your Model Disappeared Overnight. Here's How to Vet a Relay Before the Next One Does

Your Model Disappeared Overnight. Here's How to Vet a Relay Before the Next One Does

On November 19, a developer opened Codex to keep working on a project and found GPT-6.2 Sol gone. Not deprecated, not renamed — just gone, mid-project. Within a minute the post was on the front page of linux.do: "My 6.2 Sol stopped working". The same day, a free relay announced its GPT-6.2 Astra was "temporarily delisted."

This was not a one-off. Sol has vanished or been rate-limited in every generation — 6.0, 6.1, 6.2 — and the related threads go back months. The lesson developers are drawing is not "pick a better model." It is: the model you build on can disappear overnight, and the question to ask any relay is not only what it costs, but whether it will still be there tomorrow.

The demand is already public

The same day, in a thread titled "Google has stood up", users were openly asking for a cheap Gemini relay. Their reasons were specific, not price-only: opening an account themselves is a hassle, Google's IP and payment checks are strict, and Codex Fast burns through quota too fast. Two more users immediately replied "+1, I need this too."

That is the shape of real demand: not "cheapest model wins," but "I want this model, I don't want to maintain the account, and I want the cost to be predictable." Any relay that answers only the price part is answering the wrong question.

Three checks before you depend on a relay

1. Is the price public, dated, and checkable?

A price you cannot verify is a rumour. Look for a pricing page that shows when it was last synced and lists each model with input and output rates against the vendor's official price.

As a worked example, using prices listed on zhuan.xyz/pricing today (page sync 2026-11-18 21:21 UTC, 23 models): GPT-6.2 Sol is listed at $0.24 / $1.19 per million tokens against the official $2 / $10 — 88.1% below official. GPT-6.2 Astra is listed at $0.95 / $4.75 against $8 / $40, also 88.1% below. Claude Fable 5.2 is listed at $2.48 / $12.42 against $8 / $40, 69% below.

Note what this does and does not prove. It proves those models are listed at those prices today. It does not prove they will be listed tomorrow — availability changes, and any honest relay should say so. Before you depend on a model, check the listing the same day you need it.

One more honesty note: that same pricing page lists no Gemini models at all as of today, so no Gemini price is cited here. If Gemini 4 is the model you need, check availability first and do not assume a relay carries it because it carries Claude and GPT.

2. How does it handle failure in public?

Every relay hits 429s, overloads, and outages. The difference is what happens next. In one linux.do thread this week, a user reported a relay returning 429s and the only answer was "join the group and DM us" — no cause, no fix, no follow-up in the thread.

Ask for the opposite: a public record of what broke, why, and how long recovery took. A provider that publishes its incidents is giving you the one thing a price list cannot — evidence about what happens on a bad day. A provider that only handles failures in private DMs is asking you to trust it with no evidence at all.

3. Can you verify the model is really that model?

"Downgraded model" is now the default suspicion in relay discussions, and users have started verifying it themselves with observable tests. In one thread about Gemini 4 Carbon, a user suspected their fast replies meant they had been routed to an old, small model — and the community's test was to time a specific image-generation task. If it finished in under ten seconds, the verdict was: you are not talking to the model you paid for.

That particular test is a community anecdote, not a benchmark, and it should not be treated as proof either way. But the instinct behind it is right: prefer relays that give you something checkable — model fingerprints, response headers, cache-hit rates, uptime logs — over relays that ask you to take "full-blooded, not downgraded" on faith. Speed alone proves nothing; a reply can be fast because the route is good, or fast because the model is small.

The real cost is switching, not the token

When a model disappears mid-project, the cost is not the tokens you already burned. It is the afternoon you lose re-pointing your code, re-testing your prompts, and finding out whether the replacement model behaves the same way. That is why "stable and available" has become a selling point in its own right, alongside price.

So the practical checklist is short: a dated public price list you can re-check, a public incident record, and a verifiable way to confirm the model behind the endpoint. If a relay cannot show you all three, its price is not the number that matters.

zhuan.xyz

Top comments (0)