DeepSeek低价风暴打服硅谷！海外平台争相倒贴V4 Flash
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - 量子位 reports that DeepSeek's V4-Flash-0731 release delivers near-frontier coding/agent performance at roughly 1% of competitors' API prices, triggering a pricing shock where overseas platforms are stacking additional subsidies on top of an already ultra-cheap model. It matters because frontier-adjacent capability at commodity pricing shifts the economics of building AI applications.
- V4-Flash-0731 reportedly keeps the identical architecture and model size as the April V4 Preview — the only change was a redone post-training run, which lifted DeepSWE from 7.3 to 54.4, plus Cybergym 76.7 and Terminal Bench 2.1 82.7.
- Artificial Analysis intelligence index of 50 (+10 over the prior generation), said to surpass DeepSeek's own V4 Pro and approach Opus 4.8.
- Pricing cited at $0.14 per million input tokens (cache miss) and $0.28 per million output tokens; OpenCode reportedly served 8T tokens of DeepSeek Flash on Aug 1 alone, while Nous Portal ran a 90%-off promo and Cline tripled its free tier.
- Article also flags an upcoming "DeepSeek Code" agent harness, with lead 崔添翼 recruiting closed-beta testers from open-source agent-harness projects.
Note: benchmark and pricing figures are as reported by the article; treat vendor/media-sourced numbers as unverified.
Sources (1)
DeepSeek低价风暴打服硅谷！海外平台争相倒贴V4 Flash
TL;DR - 量子位 reports that DeepSeek's V4-Flash-0731 release delivers near-frontier coding/agent performance at roughly 1% of competitors' API prices, triggering a pricing shock where overseas platforms are stacking additional subsidies on top of an already ultra-cheap model. It matters because frontier-adjacent capability at commodity pricing shifts the economics of building AI applications.
- V4-Flash-0731 reportedly keeps the identical architecture and model size as the April V4 Preview — the only change was a redone post-training run, which lifted DeepSWE from 7.3 to 54.4, plus Cybergym 76.7 and Terminal Bench 2.1 82.7.
- Artificial Analysis intelligence index of 50 (+10 over the prior generation), said to surpass DeepSeek's own V4 Pro and approach Opus 4.8.
- Pricing cited at $0.14 per million input tokens (cache miss) and $0.28 per million output tokens; OpenCode reportedly served 8T tokens of DeepSeek Flash on Aug 1 alone, while Nous Portal ran a 90%-off promo and Cline tripled its free tier.
- Article also flags an upcoming "DeepSeek Code" agent harness, with lead 崔添翼 recruiting closed-beta testers from open-source agent-harness projects.
Note: benchmark and pricing figures are as reported by the article; treat vendor/media-sourced numbers as unverified.