Google: Gemma 4 31B (batch)

The google/gemma-4-31b-it:batch row left OpenRouter's catalog between the 2026-09-14 and 2026-09-21 snapshots, and a live fetch of the catalog on 2026-09-21 confirms it is still absent: the standard and free rows are listed, the batch row is not. The Pulse recorded the withdrawal on 2026-09-21, and the row was last seen on 2026-09-14.

That date is not incidental. Together AI had published the removal in advance. Its changelog entry of September 1, 2026 lists google/gemma-4-31B-it among the models "deprecated and will be removed from serverless on September 14, 2026", recommends zai-org/GLM-5.3-Flash as the replacement, and says the model remains available through on-demand dedicated endpoints; the September 15 entry records it "no longer available on serverless". The same publisher's Deprecations page tables the 2026-09-14 removal and a 2026-09-15 history row.

The match is not only the date. As observed on 21 September 2026, three :batch rows left this catalog in the same transition — this one, openai/gpt-oss-20b:batch and thinkingmachines/inkling-small:batch, all last seen 2026-09-14 — and they were exactly the three chat models in that one Together notice. Each one's last recorded prompt and completion rate, context length and completion limit equal Together's endpoint for the same model in OpenRouter's endpoint listing (this model, gpt-oss-20b, Inkling Small, all fetched 2026-09-21), and as observed on 21 September 2026, in each of the three listings Together was the only endpoint that matched on the full tuple. OpenRouter publishes no per-row reason, so which party dropped the entry — the router, or the entry losing the endpoint it described — is not stated; the dated removal of the serving model is.

OpenRouter's own records agree the entry is gone. …/models/google/gemma-4-31b-it:batch/endpoints answers HTTP 200 with "endpoints":[], and the variant guide defines that as the resolved catalog entry not existing — the lookup does not fall back to the base model, and the batch quickstart says a submission for a model with no :batch endpoint returns 400. The single-model route answers HTTP 404, "Model not found: google/gemma-4-31b-it:batch" (both fetched 2026-09-21). The human page still answers 200 and still prints the batch slug, but lists no provider and no price; in the standard row's endpoint listing linked above, Together's endpoint is the only one reported without uptime, latency or throughput, and Together's Deprecations page notes that during a transition a model "can still appear in catalog listings while its route-level availability changes".

The row did not move to another tier, and no Gemma row has a batch entry now. No arrival in the transition shares its google/gemma-4-31b-it-20260402 canonical stem, so nothing here was renamed or superseded.

The model is not retired either. Gemma 4 31B is in Google's current Gemma 4 family — the Gemma 4 model card, last updated 2026-07-30, lists the 31B Dense variant, and the weights remain published — and OpenRouter still lists the standard and free rows that serve it. The batch offering as a whole was alive: as observed on 21 September 2026, the Pulse's snapshot carried 74 :batch rows, and Together's recommended replacement, z-ai/glm-5.3-flash, was listed with a batch row of its own. The timeline's 2026-09-21 retired line records the catalog row leaving; the model was not retired.

The feeds: binding is kept. If google/gemma-4-31b-it:batch is ever a catalog row again it binds straight back to this entry.