Together cuts Qwen3.8 inference pricing, adds GLM-5.3 to serverless lineup

WHY IT MATTERS

As open-weight releases accelerate, neutral inference platforms face margin pressure from both directions — model labs shipping cheaper, better weights, and hyperscalers undercutting on raw compute. Watch whether cached-input pricing keeps compressing faster than input/output pricing across serverless inference providers generally — that would confirm agentic and long-context workloads, not raw token volume, are becoming the real pricing battleground.

BY THE NUMBERS

• Together also added GLM-5.3 to its serverless catalog, priced at $1.40 input / $4.40 output / $0.26 cached input per 1M tokens
• The sharpest number here is the 50% cut to cached-input pricing, not the smaller trims to input/output

THEIR RATIONALE

The move reads as a straightforward pass-through of falling model-serving costs plus a same-day catalog expansion: cut the price on an existing model while adding a new frontier open-weight model at a competitive rate, keeping Together's serverless lineup priced ahead of substitutes.

THE SIGNAL

The sharpest number here is the 50% cut to cached-input pricing, not the smaller trims to input/output. Cache pricing is the lever inference platforms use to compete for repeat-context workloads — agents, long chat histories, RAG — without touching headline rates that get quoted in comparisons. Pairing a price cut on an existing model with a same-day new-model launch also shows Together competing on portfolio breadth: keep customers inside the serverless catalog by constantly refreshing which model offers the best price per token, rather than defending any single model's price point.