GLM-5.3-Flash's 50%-off launch pricing ends Sept 9; cost doubles
Open-weight MoE that nears Opus 4.8 on coding: input goes $0.075→$0.15 per 1M at 16:00 UTC Sept 9. Upstage's Solar Pro 4 jumps 10x a day later.
Copy markdown
The deadline: half-price promo ends 16:00 UTC Sept 9
GLM-5.3-Flash's 50%-off launch promo ends at 16:00 UTC on Sept 9 (midnight Sept 10 in Singapore, 8 PM in Dubai). Input rises from $0.075 to $0.15 and output from $0.25 to $0.50 per 1M tokens; cached input doubles from $0.015 to $0.03. If you batch against it, lock in the promo rate before then or plan for 2x.
What doubles: a 320B open-weight multimodal MoE
The model losing its discount is MIT-licensed open weights: 320B total / 18B active, a ~1M-token context, and the first natively multimodal model in the GLM-5 line (text, image, video). Even at the post-promo $0.15/$0.50 it undercuts nearly every hosted frontier API, and because the weights are open you can self-host and skip the per-token bill entirely.
The benchmarks: agentic scores approaching Opus 4.8
GLM-5.3-Flash reports 63.4 on DeepSWE (up from GLM-5.2's 46.2), 84.3 on Terminal-Bench 2.1, and 78.4 on Toolathlon — close to Claude Opus 4.8 on coding and agentic tasks. On vision it tops DeepSeek's Vision-Exp on Chartography (78.0 vs 64.3). For agent builders, that's frontier-adjacent tool use at flash pricing.
Elsewhere: Upstage's Solar Pro 4 jumps ~10x on Sept 10
One day after GLM's promo closes, Upstage's agent-focused Solar Pro 4 (Intelligence Index 42) reportedly leaves launch pricing, rising from $0.03/$0.12 to $0.30/$1.20 per 1M tokens at 23:59 UTC Sept 10. Two low-cost API windows shut 24 hours apart — if either model is in your stack, re-price this week.