NVIDIA: Nemotron 3 Ultra (batch)

The nvidia/nemotron-3-ultra-550b-a55b:batch row left the OpenRouter catalog in the 2026-09-03 snapshot, a day after it was last listed, active and without an expiry date. A live fetch of the catalog the same day (https://openrouter.ai/api/v1/models) confirms the delisting: no batch variant for this model, while the parent nvidia/nemotron-3-ultra-550b-a55b and :free rows remain. The bound facts above no longer render a dated last-known value: the row has left both the current and previous snapshots, so the data layer records no as-of date for it.

The model is not retired and was not renamed. NVIDIA still serves Nemotron 3 Ultra on its own NIM API (build.nvidia.com, live on 2026-09-03), and OpenRouter still lists the standard and free rows. No nvidia/ row was added and none other was removed in the 2026-09-03 snapshot, so no successor slug has appeared — but the snapshot was not otherwise still in that namespace: three of the nine nvidia/ rows changed, as observed on 3 September 2026, including the standard row's price, whose prompt rate fell from $0.625 to $0.60 and whose completion rate fell from $3.125 to $2.40 per million tokens — a move the 2026-09-04 snapshot reversed, listing $0.625 / $3.125 again.

OpenRouter publishes no reason for withdrawing the batch variant specifically. OpenRouter's batch offering is alive as a whole: 66 :batch rows remained in the same catalog, as observed on 3 September 2026, and the Batch API quickstart documents an async batch endpoint that takes a standard model slug. But this model's batch page (nvidia/nemotron-3-ultra-550b-a55b:batch) still identifies itself as the batch variant — the slug and the page's own variant state both name it — but its variant list now holds only the free and standard groups, with no batch group and so no providers or pricing at all; nothing anywhere on the page announces a deprecation, and the standard model's page carries no mention of a batch option or its removal. NVIDIA's own pages say nothing about batch availability through OpenRouter.

The provider behind the row did publish its side. The row's last recorded prompt, cached-input and completion rates are exactly the ones Together AI's serverless model list gave for nvidia/nemotron-3-ultra-550b-a55b (archived 18 August 2026), and none of the three providers on the standard row's endpoint listing (DeepInfra, BaseTen, Venice; fetched 2026-09-25) quotes that pair of prompt and completion rates. Together's changelog entry of August 27, 2026 lists nvidia/Nemotron-3-ultra-550b-a55b among models that "have been deprecated and are no longer available on serverless", naming no replacement, and its Deprecations page tables the 2026-08-27 removal with on-demand dedicated endpoints marked No (both fetched 2026-09-25). By Together's own batch overview, its batch API runs on serverless models or on dedicated model inference, and this model now has neither there. So the withdrawal was not a fold into another tier: the provider that served the batch row ended all of its inference for the model, and no other provider took the batch entry over — the row's endpoint listing answers "endpoints":[] (fetched 2026-09-25). What the sources still leave open is the six-day gap: the row stayed listed until 2026-09-02, after Together's removal date, and neither publisher says why it outlived it or which party dropped it. Together's Deprecations page notes only that during a transition a model "can still appear in catalog listings while its route-level availability changes".

The standard row is no substitute for batch work. The router's async batch endpoint takes the base model slug in its model field, but the same quickstart states that "Batch runs on :batch endpoint variants" and that "A submit for a model with no :batch endpoint returns 400" (fetched 2026-09-25). With the batch row's endpoint list empty, a batch submission for this model has no route to run on, and the standard row's providers cannot stand in for it. The binding is kept, so a re-listing of this row would re-mint against this entry.