W&B reprices DeepSeek V4-Flash 14x on Serverless Inference

Weights & Biases repriced the entry-level model in its Serverless Inference catalog. DeepSeek V4-Flash, an experimental 1M-context MoE model, moved from a flat /bin/bash.01 input / /bin/bash.01 output per 1M tokens to /bin/bash.14 input, /bin/bash.07 cached input, and /bin/bash.28 output — a 14x increase on input and 28x on output, plus a new discounted cached-input tier that the old two-part rate did not have.

The move lifts the floor of the whole catalog: the cheapest published input rate on W&B Inference is now /bin/bash.05 per 1M tokens (OpenAI GPT OSS 20B, IBM Granite 4.1 8B, JetBrains Mellum2 12B), where DeepSeek V4-Flash previously undercut everything at /bin/bash.01. The /bin/bash.01 rate read as a loss-leading introductory price on an “Experimental”-badged model; normalising it to /bin/bash.14//bin/bash.28 puts V4-Flash roughly in line with the rest of the cheap tier while still sitting well below DeepSeek V4-Pro at .74 input / .46 output.

Nothing else moved. The cloud pricing page is unchanged from the prior capture: Free at /bin/bash/mo, Pro from 0/month billed monthly, custom Enterprise, storage overage at /bin/bash.03/GB, Weave data ingestion overage at /bin/bash.10/MB, the /mo Pro Inference credit, CoreWeave Sandboxes credits (0/mo Free, 5/mo Pro), and ARIA still free “for a limited time” on Free and Pro.