From Goldman's Delta-1 head, Rich Privorotski:
"A panel of Gemini 3 Flash, Kimi K2.6 and DeepSeek V4 Pro, fused together, beat solo GPT-5.5 and solo Opus 4.8 outright, while landing within 1% of Fable 5 at roughly half the cost.” (Openrouter -X ) If true, this is exactly the direction of travel the market has been underestimating. The race for intelligence increasingly feels as though it is shifting from a handful of labs toward model orchestration and open-source ecosystems. That is simultaneously bullish and bearish. Bullish because lower costs and broader access should ultimately drive more token consumption and more compute demand. Bearish because it accelerates token deflation and raises questions around the durability of model economics. The trillion dollar question remains whether lower intelligence costs ultimately create more demand than they destroy pricing power."
Goldman's Delta-1 head notes that a fused panel of open-source models (Gemini 3 Flash, Kimi K2.6, DeepSeek V4 Pro) reportedly matched GPT-5.5 and Opus 4.8 performance at half the cost, coming within 1% of Fable 5. This signals a potential shift from proprietary labs to model orchestration and open-source ecosystems, which is simultaneously bullish (lower costs drive token consumption and compute demand) and bearish (accelerates token price deflation and threatens model economics). The core question is whether cost reductions create more demand than they destroy pricing power.
Goldman mentioned only as the employer of the quoted analyst (Rich Privorotski); not a subject of investment opinion. The post discusses AI model economics broadly, not Goldman's business.
关键要点
Fused open-source model panel allegedly matched top proprietary models at half the cost
Shift from proprietary labs to model orchestration and open-source ecosystems
Bullish case: lower costs increase token consumption and compute demand
Bearish case: token deflation and weakened pricing power threaten model economics
Central uncertainty: demand creation vs. pricing power destruction
