What's this about?
DeepSeek's V4 Flash (0731) became the Chinese AI community's obsession this week. A viral thread from a heavy user claims it did in ~1.5B tokens what previously took "many tens of billions of GPT-5.6 Sol tokens" — praising its efficiency while calling Sol "a high-IQ obsessive who's slow and overthinks." OrangeAI documented resellers (Command Code, OpenCode, Cola) selling V4 Flash *6× below* DeepSeek's own official price, with the punchline "you can't use it up." Then the twist: DeepSeek announced it's about to raise prices sharply — "good times never last."
What exactly is a "Flash" model?
A smaller, faster, cheaper variant in a model family, tuned for high throughput and low cost rather than maximum capability. The claim driving the hype is that for real coding work, V4 Flash's speed and low cost make it *more* productive per dollar than a slower, more expensive frontier model — efficiency beating raw intelligence.
What's the community saying?
- "The world underestimates V4 Flash" — the core claim is that its impact only shows in direct comparison: same work, a fraction of the tokens and time.
- Efficiency-over-IQ: repeated complaints that slow "overthinking" frontier models are less productive than a fast good-enough one.
- Pricing whiplash: resellers dumping it below cost, then DeepSeek hiking — a live demonstration that "the price of a model" is unstable and mostly about who's subsidizing whom.
- OrangeAI's "industry inflection" claim: more people will realize cheap-fast models change the economics of building.