On August 6, DeepSeek dropped a short announcement that sent ripples through the developer community: they’re planning a “significant” price increase for their API. No numbers. Just a warning that the hike is coming, and it’s going to be big.
Honestly? I wasn’t surprised.
Three weeks earlier, DeepSeek introduced peak-hour pricing — double the cost during weekday business hours. Before that, in May, they made a permanent price cut that cemented their reputation as the industry’s price butcher. V4-Pro dropped to about $0.43 per million input tokens and $0.86 for output, a 75% reduction from the original.
So what changed?
The numbers tell the story.
On July 31, DeepSeek released the official version of V4-Flash, and the jump was big. According to Artificial Analysis, V4-Flash-0731 scores 50 on their intelligence index while costing about 3 cents per task. For comparison, Anthropic’s Claude Fable 5 averages $3.15 per task, and GPT-5.6 Sol runs about $1.86. That’s not a gap. That’s a chasm.
Then the usage numbers came in. OpenRouter data shows V4-Flash hit 7.22 trillion tokens in weekly calls, taking the global top spot. On August 1 alone, OpenCode reported 8 trillion tokens processed in a single day — 5 trillion from free trials, 3 trillion from paid usage. By August 5, that daily number had climbed to 7.5 trillion, with V4-Flash accounting for 6.3 trillion. That’s 84% of the total.
Let that sink in. One model, eating 84% of the traffic on one of the biggest coding platforms.
And here’s the kicker: V4-Flash is absurdly cheap. The token price is roughly 1/12 of V4-Pro, 1/15 of Zhipu AI’s GLM 5.2, and 1/53 of Kimi K3. At around $0.14 per million input tokens and $0.28 per million output, a typical agent session costs about $0.13. Less than a dollar.
People have stopped optimizing. Why would they? At that price, you just throw tokens at the problem.
But that’s exactly the problem.
Agent workloads don’t care about peak hours. They run batch jobs and async tasks across time zones. The peak-valley pricing DeepSeek introduced in mid-July was meant to smooth out demand, but it barely touches agent traffic — and agents are exactly where token consumption is exploding.
DeepSeek’s infrastructure is getting hammered. On August 4, the API suffered multiple performance degradation events. The official status page shows fault rates spiking after the Flash release.
The math is brutal. DeepSeek’s pricing philosophy was always “ten months to recover hardware costs.” When usage grows this fast, that math breaks. You can’t keep selling compute below cost when the compute bill is doubling every week.
What does the increase actually look like?
My guess? They’re rolling back to pre-discount prices. The May cut brought prices to one-quarter of the original, so a return to full price would be a 300% increase. “Significant” is probably an understatement.
One thing I’d watch closely: whether the increase lands on everything or just the flagships. A 300% increase on V4-Pro is a business decision. The same increase on V4-Flash would be an earthquake, because Flash is what everyone’s agents are actually running on.
And it isn’t only about Flash. V4-Pro official is coming soon, probably alongside DeepSeek’s new Harness coding agent. That’s going to drive even more usage. If they don’t raise prices now, everything crashes.
The bigger picture
This isn’t just DeepSeek. The whole industry is hiking. Zhipu AI has raised prices several times this year. Tencent’s Hunyuan saw increases up to 460%. The era of subsidized AI is ending.
DeepSeek’s pivot from price butcher to sustainable business is healthy, actually. You can’t burn money forever. The question is whether the increases kill the developer ecosystem that grew up around cheap APIs.
For now, I’m not panicking. Even at double the current price, V4-Flash would still be competitive. The performance-per-dollar ratio is just that good.
But I’m not going to pretend this doesn’t matter. If you’re building on DeepSeek’s API, start measuring now instead of guessing.
Two things I’d do this week. First, log your actual token usage per task for a few days and multiply it out. Most people have no idea whether they’re spending $20 a month or $200, which means they also have no idea whether “300% increase” means an inconvenience or a shutdown. Second, find the work that doesn’t need to run at 2 p.m. on a Tuesday and move it off-peak. Agents make that harder than it sounds, because the whole point is that they run without you — but scheduled jobs and nightly re-indexing are easy wins.
If you’re running one of the coding agents that sits on top of Flash — and at 84% of OpenCode’s traffic, a lot of you are — this increase lands on your bill whether you planned for it or not. That’s the difference between a price change and a repricing of your whole workflow.
There are other things worth knowing before you commit serious volume to this API, and I collected the ones that actually bit me in a separate post about the DeepSeek API.
Because the tokens are about to cost a lot more.
