DeepSeek v4 Flash shows how cheap models create capacity cliffs
The post is a concrete operator-side view of model demand elasticity: when a coding model gets cheap enough, usage can jump faster than serving capacity and pricing can become the control knob.
Links: Original source
Logged at IST: 2026-08-21 10:12 IST
What it is: Jay from OpenCode explaining what happened after DeepSeek v4 Flash became the cheap default for a lot of coding-agent traffic.
Gist: The useful part is the operator view of a model getting cheap enough to change user behavior. Jay says DeepSeek v4 Flash launched on August 1 and grew on OpenCode from roughly 3T tokens/day to 18T tokens/day in two weeks, close to doubling OpenRouter's daily volume and possibly 30 to 50% of DeepSeek's own volume.
The cause, in his telling, was not just model quality. It was price. Users got a first taste of AI that felt "too cheap to meter": he describes DeepSeek Flash as far cheaper than several alternatives, while also being a marked improvement over the prior Flash model.
Then the economics snapped back. On August 16, DeepSeek raised prices by 5x during peak hours and 2.5x off-peak. Jay says growth stopped and daily tokens fell to less than half the peak. His explanation is capacity: DeepSeek may not have been able to absorb the 6x demand jump, and higher GPU prices would make new capacity hard to serve at the old price.
The infrastructure detail is the most interesting bit. Jay suspects DeepSeek is doing unusually aggressive long-lived token caching, which is why only a few providers can match the original price. OpenCode has been trying to fill a near 10T token/day gap, but even a provider that can match the price would need enough capacity. He estimates roughly 1000 B300s for their throughput.
The takeaway: model adoption is being decided by a three-way constraint between capability, token price, and available inference capacity. If a model is competent and cheap enough, demand can appear almost instantly. If serving cannot keep up, pricing becomes the throttle. That gap is now an opening for every lab that can make coding-agent inference feel abundant rather than metered.
Newsletter angle: Strong follow-up to the DeepSeek value-per-intelligence thread. This is less about another benchmark and more about the market signal: cheap competent coding models create real, measurable demand cliffs.