Checking IBM's claim: open weights are cheaper, but not uniquely so

A cost check of IBM CEO Arvind Krishna's claim that open-weight models deliver ~90% quality at 10–25% of frontier token costs. On an 8,000-input, 500-output job, DeepSeek V4.1-Flash costs 2.9% of Claude Opus 5 and 7.1% of GPT-6 Sol, but closed small tiers GPT-6 Luna and Claude Haiku 5.5 undercut Flash by 30%.

• DeepSeek V4.1-Flash off-peak costs 2.9% of Claude Opus 5 and 7.1% of GPT-6 Sol
• GPT-6 Luna and Claude Haiku 5.5 cost $1.05 per 1,000 requests, 30% below Flash
• Own hardware beats GPT-6 Sol above 7% utilization but never beats Flash
• The 90% quality figure is unverifiable: IBM named no benchmark for 70, 50 and 90