OpenAI: gpt-oss-20b (batch)

The openai/gpt-oss-20b:batch entry left OpenRouter's catalog between the 2026-09-14 and 2026-09-21 snapshots, and a live fetch on 2026-09-22 finds it still gone — the standard openai/gpt-oss-20b entry is listed, the batch variant is not. OpenRouter's own routes say the variant has ceased to exist rather than merely stopped serving: …/models/openai/gpt-oss-20b:batch/endpoints answers 200 with "endpoints":[], and …/model/openai/gpt-oss-20b:batch answers 404, Model not found: openai/gpt-oss-20b:batch (both fetched 2026-09-22). The variant guide says exactly what that pair means: "When the resolved catalog entry does not exist, for example :free on a model with no free entry, the single-model route returns 404 and the endpoints route returns 200 with "endpoints": []. Treat either as an unavailable model, not as a signal to retry with the base slug."

The provider published the withdrawal, dated, a fortnight before the entry went. Together AI's changelog entry of September 1, 2026, under "Model deprecations": "The following models are deprecated and will be removed from serverless on September 14, 2026" — first on that list, openai/gpt-oss-20b, "Recommended replacement: Qwen/Qwen3.5-9B", with the note that all but one model named "remain available through on-demand dedicated endpoints". The September 15, 2026 entry records it done, under "Serverless model deprecations": "The following models are no longer available on serverless", openai/gpt-oss-20b first again. Together's Deprecations page tables both halves — a scheduled row reading 2026-09-14, openai/gpt-oss-20b, Qwen/Qwen3.5-9B, Yes for on-demand dedicated endpoints, and, in the history of "all models removed from serverless inference, most recent first", 2026-09-15, openai/gpt-oss-20b, Yes (all fetched 2026-09-22). September 14 is the day this entry was last seen.

That the withdrawn entry was Together's is measured, not assumed. OpenRouter's batch quickstart states that "A batch runs on one provider. By default OpenRouter picks the cheapest eligible batch endpoint for the model" — so a :batch entry stands or falls with the single provider behind it, and each of three surviving batch variants probed on 2026-09-22 resolves to exactly one endpoint. Of the thirteen endpoints, across twelve providers, that the standard entry lists when fetched 2026-09-22, exactly one carries the prompt rate, completion rate, context length and completion cap that were pinned for this entry's feed row when it went missing: Together's. Exactly one carries the same fifteen supported parameters, as a set: Together's again. Nothing else on that listing matches on either, and the pinned values are kept in data/vanished/answered/openrouter-models--openai-gpt-oss-20b-batch.md.

What did not happen is worth saying plainly, because a vanished feed row reads like a death and this is not one. The model is not retired: its weights are still published under Apache 2.0 at openai/gpt-oss-20b, whose API record reported the repository live, ungated and last modified 2025-08-26 when fetched on 2026-09-22, and OpenRouter still routes the standard entry across twelve providers. Nothing was renamed: the standard entry keeps both its id and its openai/gpt-oss-20b canonical slug, and none of the arrivals in the same transition shares that stem. The batch tier was not folded elsewhere for this model either — no gpt-oss-20b batch variant exists under any other slug.

What is still unpublished is the router's own reasoning. OpenRouter gives no per-entry reason for a withdrawal, so whether it dropped this variant because the endpoint behind it went away, or on a schedule of its own, is stated nowhere; the provider's account is the dated one, and the dates are one day apart. Two observations keep that gap honest. OpenRouter still listed a Together endpoint on the standard entry a week after Together says the model left serverless, and that endpoint reported no uptime, latency or throughput at all — a state Together's Deprecations page anticipates in a note of its own: "During a transition, a model can still appear in catalog listings while its route-level availability changes. If a listed model returns unavailable or disabled errors, treat the runtime API behavior and this page as the source of truth." And the batch tier of this catalog is moving on its own account: openai/gpt-oss-120b:batch, and the batch variant of Together's own recommended replacement, qwen/qwen3.5-9b:batch, were both in the 2026-09-21 snapshot and both absent from the live catalog on 2026-09-22, while neither model appears in any Together deprecation notice. One provider's dated removal explains this entry's timing. It does not explain everything happening to batch variants around it.

The feeds: binding stays. If openai/gpt-oss-20b:batch is ever a catalog entry again it binds straight back to this page. The timeline's 2026-09-21 retired line, written by the Pulse when it recorded the withdrawal, marks the catalog entry leaving — not the model, which is alive, open-weight and widely served.