CROSS-CATEGORY: Inference Pricing Bifurcated in 72 Hours — DeepSeek Went Metered-and-Up While Z.ai Went Flat-Rate Subscription at a Tenth of US Rates

Decrypt / Z.ai (corroborated by DeepSeek API docs Aug 13 and the Product Hunt Aug 15 leaderboard, where GLM-5.3 placed #2 with 217 upvotes)·high signalXBlueskyLinkedInCopy link

Z.ai shipped GLM-5.3 on Aug 14 — same 743B base as GLM-5.2, moved by post-training alone (Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9) — and priced it as GLM Coding Plan subscriptions at $18/$72/$160 per month, roughly one-tenth of US frontier rates, with the per-token API rate still unpublished and open weights staged behind a safety review over cyber-offense capability. Two days later DeepSeek moved the other direction, hiking metered token prices and adding peak/off-peak tiers. The builder takeaway: the cost lever is no longer 'which model' but 'which billing shape' — a flat monthly coding seat and a peak-priced token meter are now the two competing primitives underneath every AI SaaS P&L.