Together drops four serverless inference models six days after adding NVIDIA B200

WHY IT MATTERS

For buyers routing inference workloads, Together's growing churn means pinned model availability can't be assumed — endpoints disappear roughly as often as plans change, so contracts or pipelines referencing a model by name carry real deprecation risk. Watch whether the dropped capacity (Gemma 4 31B, DeepSeek V4 Flash 0731, Llama 3.3 70B, MiniMax M2.7) resurfaces as newer variants, and whether other inference providers start pruning older checkpoints at a similar clip rather than just adding to their catalogs.

BY THE NUMBERS

• 19 pricing/catalog changes in the last 90 days, about one every 8 days
• Sixth model-catalog removal since tracking began; previous was Sep 11, 2026 (LFM2.5-8B-A1B)
• Comes 6 days after Together added NVIDIA B200 post-training at $11/GPU/hr on Oct 20, 2026
• Feature removals are rare: just 3.7% of 2,028 tracked pricing changes in the last 180 days

THEIR RATIONALE

Together hasn't stated a rationale for dropping these four models, but the timing points to catalog churn rather than retrenchment: this sits inside a run of 19 pricing and catalog changes in the last 90 days — about one every 8 days — and follows closely on new serverless additions like the NVIDIA B200 post-training option, Reflection Beam and the Tev1 4B Experimental model. The likely read is routine pruning of older or lower-demand models to make room on the catalog, consistent with Together's recurring pattern of model removals, most recently on Sep 11, 2026 (LFM2.5-8B-A1B) and Jun 29, 2026 (GLM-5).

THE SIGNAL

The sharper read: feature removals like this are uncommon across the index — only 76 of the 2,028 pricing changes tracked in the last 180 days (3.7%) involved dropping a feature or model, spread across 74 companies, with Beehiiv and Supermetrics the only ones to do it twice. For Together, this is already its sixth such removal, more than almost anyone else in the pattern, which fits a company iterating on its model catalog faster than the market resets prices — a sign that in the inference-pricing race, speed of catalog turnover is becoming as much a competitive lever as the per-token rate itself.

IN CONTEXT

Together has pruned its serverless model catalog six times since tracking began, most recently removing LFM2.5-8B-A1B on Sep 11, 2026 and, before that, GLM-5 on Jun 29, 2026. This cut lands just six days after Together added NVIDIA B200 post-training at $11/GPU/hr, part of a cadence of roughly one pricing change every 8 days over the past year. Across the index, that kind of removal stays rare — Flatlogic and Krea made similar cuts in the past week, among just 76 removal-type changes out of 2,028 tracked in the last six months. For Together, dropping older models looks less like retrenchment and more like routine upkeep to keep the catalog current with its steady stream of new releases.