Two language models launched within hours of each other — DeepSeek V4 Pro and Grok 4.6 — and together they have weakened the market rationale that justified slow response times or steep price premiums for top-tier models.
What happened
- DeepSeek V4 Pro scored 87.9 on the Terminal Bench tool benchmark, nearly matching Claude Fable 5’s 88.0. DeepSeek’s reported cost was about $0.87 per million tokens, compared with roughly $50 per million tokens for Claude Fable 5. DeepSeek’s team has warned that prices could rise later.
- Grok 4.6 struck at speed and cost-efficiency: according to the source, coding tasks that ran longer than 30 minutes on GPT-5.6 Sol finished in under 20 minutes on Grok, with some completing in three minutes. The per-task cost measurements cited were about $2 for input and $6 for output in those examples.
Why this matters
The AI model market has long been governed by an ‘‘impossible triangle’’ idea: you can have models that are smart, cheap, or fast — at most two at once. That dynamic allowed frontier models to command higher prices and tolerate slower responses because marginal quality advantages justified the premium.
DeepSeek broke open the price corner of that triangle, while Grok opened up the speed corner. Neither model claims to be the outright best in every metric, but together they eliminate the customary excuse that only superior raw quality warrants higher cost or slower service.
Concrete implications
- Organizations and developers who previously paid for the highest-performing models now have clear alternatives that can materially lower operating costs and shorten development cycles.
- Competitive pressure will shift not only to raw model quality but to how quickly and cheaply that capability is delivered in practice.
Summary
DeepSeek V4 Pro and Grok 4.6, released on the same night roughly two hours apart, attacked the frontier’s two traditional levers: price and speed. As capabilities that are ‘‘good enough’’ become both affordable and fast, remaining slow and expensive becomes a harder commercial position to defend.
(Names cited: DeepSeek V4 Pro, Grok 4.6, Claude Fable 5, GPT-5.6 Sol, Terminal Bench. Performance scores and price figures follow the numbers provided by the source; the launches were described as occurring "one night, two hours apart" in the source material.)
