Together Rebuilds Fine-Tuning Pricing From 3 Size Tiers to 29 Per-Model Rows

WHY IT MATTERS

For buyers, fine-tuning costs on Together now depend on which specific model is chosen rather than which size bucket it falls into, so cost comparisons that used to work off three tiers need to be redone per model, and job minimums as high as $60 change the math for small fine-tuning runs. For the category, it signals that as model catalogs balloon past two dozen options, size-tier pricing stops scaling and providers default to per-model tables instead — watch whether Together's ~8-day change cadence continues into Q4, and whether rival model marketplaces follow with similarly granular fine-tuning pricing.

BY THE NUMBERS

• 3 size tiers replaced with 29 per-model price rows
• Job minimum shifts from a flat $4 to $4-$60 per model
• 4th pricing restructure by Together since March 2026
• 18 pricing changes in the last 90 days, ~1 every 8 days

THEIR RATIONALE

Together changes its pricing pages roughly every 8 days on average and has made 18 changes in the past 90 days alone, so this restructure lands squarely inside its normal rapid-iteration cadence rather than as a one-off event — in fact it arrives just one day after a separate Qwen3.7-Max price hike. Collapsing three broad size buckets into 29 named-model rows lets Together price each model against its own compute cost instead of lumping it with others in the same parameter range, and the shift from a flat $4 minimum to per-model minimums up to $60 lets it capture more from larger or more expensive fine-tuning jobs that the old flat floor was undercharging.

THE SIGNAL

Pricing restructures are common but not universal across the index: 749 of the last 5,456 tracked changes (13.7%) over 180 days involved one, led by companies like Datadog and Twilio doing it repeatedly. This is Together's fourth restructure in that stretch — after reshuffling its serverless model lineup and renaming Dedicated Endpoints/GPU Cloud in March, and overhauling Provisioned Throughput pricing in July — so the pattern here is a company continuously re-cutting its pricing surface as its model catalog grows, not a single correction. The move to granular per-model rows also tracks a broader shift among AI infra vendors away from coarse tiers toward pricing that follows actual per-model economics.

IN CONTEXT

This is Together's fourth pricing restructure since reshuffling its Serverless Inference lineup in March, following July's Provisioned Throughput pricing overhaul, and it lands just one day after a 60% Qwen3.7-Max price hike — part of a run of 18 changes in 90 days. Restructures aren't rare across the index: 749 of the last 5,456 tracked changes (13.7%) involved one, with Datadog and Twilio the most frequent restructurers. The push toward per-model granularity echoes this week's repricing at DeepSeek, which just reversed a model retirement to keep billing steady past its own deadline. For Together, replacing broad size buckets with 29 individually priced models suggests its fine-tuning catalog has outgrown tier-based pricing.