Union Alpha not working?

The four failure modes

1. 1

   429, or replies that slow to a crawl

   A free stealth endpoint is one of the busiest endpoints on the internet on the day it launches. Rate limits and queueing are the endpoint protecting itself, not a problem with your account.

   • Retry with exponential backoff rather than a tight loop; a tight loop makes your own limits worse.
   • Move bulk work to an off-peak hour where you reasonably can.
   • Keep a second model wired up for the requests that cannot wait.
2. 2

   The model is missing from the picker, or requests 404

   Stealth previews get delisted, renamed or moved behind a real model name. When the identity is revealed, the codename often stops being the address.

   • Check our status endpoint first — it reads the catalogue live.
   • Confirm you are asking for the id, not the display name: exactly stealth/union-alpha.
   • If the listing is gone, the reveal has probably happened; check the tracker page for the new name.
3. 3

   Empty, truncated or self-contradicting answers

   Testers have reported inconsistent behaviour between sessions on this endpoint, which is part of why it is suspected to be several models behind one name. Some of it is the model; some of it is a saturated endpoint returning a partial generation.

   • Run the same prompt twice. If the two answers disagree in kind, you are seeing a routing or saturation effect, not a prompt problem.
   • Lower the output length and try again — long generations on a loaded endpoint truncate first.
   • Do not treat the model's statements about its own identity as evidence either way.
4. 4

   Tools, images or structured output do not work in your client

   The endpoint declares tools, tool_choice, response_format, temperature, top_p and max_tokens, and it accepts image input. A client that ignores those fields will silently behave as if the feature does not exist.

   • Check that your client passes the tool definitions through rather than stripping them.
   • Send a single small image as a test before blaming a long document.
   • Remember the output cap is 131,072 tokens: very large structured requests can hit it.

What nobody outside the maker can tell you

• Whether your specific account is limited, and to what — there is no account console to look at.
• Why a particular generation came back wrong. Without an operator, “the model was busy” stays a hypothesis.
• Whether the OpenCode route ever really served it. We could never check that from outside, and the window is closed now.
• How long the next codename will last. Three releases, three unannounced endings — that is the whole sample.

We keep those limits on the page instead of writing confident-sounding filler, because the point of a tracker is to be trustworthy on the day it matters.

FAQ