On July 30, 2026, OpenAI cut the input token price of GPT-5.6 Luna from $1 to $0.20 per million tokens, and the output token price from $6 to $1.20 per million tokens. A reduction of up to 80%. OpenAI itself explained this as being "made possible by improvements in model architecture efficiency," but most of the industry read it as a warning shot aimed at Chinese AI companies.
Just hours later, DeepSeek released V4 Flash 0731. Input priced at $0.14 per million tokens, output at $0.28. This undercuts Luna's just-reduced price by a further 30%, and for output pricing comes in at less than a quarter of Luna's rate. The model has 284 billion total parameters, but thanks to its Mixture of Experts architecture, only about 13 billion parameters are activated during inference. This efficiency underpinned the extremely low pricing.
| Model | Input Price (/1M tokens) | Output Price (/1M tokens) | Total |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | $0.14 | $0.28 | $0.42 |
| OpenAI GPT-5.6 Luna (post price cut) | $0.20 | $1.20 | $1.40 |
| Google Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.50 |
| DeepSeek V4 Pro | $0.435 | $0.87 | $1.305 |
| Anthropic Claude Fable 5 (converted per task) | approx. $3.15/task |
According to measurements by the benchmarking firm Artificial Analysis, V4 Flash's cost per task is approximately 3 cents. Compared to Anthropic's Claude Fable 5 at $3.15 per task, that's a price gap of more than 100x.
A flood of 7 trillion tokens a week swallowed 20,000 GPUs
The release of V4 Flash 0731 attracted demand that exceeded even DeepSeek's own expectations. In OpenRouter's weekly rankings, V4 Flash processed 7.22 trillion tokens between July 27 and August 2, taking the top spot. On August 1 alone, it processed 8 trillion tokens on the coding agent platform OpenCode. Of that, 5 trillion tokens came from free trial usage, and 3 trillion from paid usage.
Let's set this figure against DeepSeek's own computing resources. Founder Liang Wenfeng stated at an investor meeting that the company's available computing resources amount to roughly 20,000 units in H100-equivalent terms. Most of these, he said, "arrived in just the past month or two." Meanwhile, he noted that while the largest U.S. models activate roughly 800 billion parameters during inference, DeepSeek's experimental environment is limited to about a tenth of that—on the order of several billion parameters.
Handling over 7 trillion tokens' worth of inference requests per week with 20,000 GPUs. Looking at the numbers, it's clear this situation was unsustainable. On August 4, performance degradation occurred on the V4 Flash API, temporarily rendering the service nearly unusable. DeepSeek announced it had restored service that same day, but this outage appears to have been the direct trigger for the price increase notice.
In fact, DeepSeek had already introduced a mechanism back in July that doubled prices during peak hours (9am–12pm and 2pm–6pm Beijing time). Even so, it couldn't fully suppress demand, and in this latest notice indicated it would raise the baseline price itself—a price hike applied across all hours, separate from the peak-hour surcharge.
What it means for 284 billion parameters to "nearly match"
Here we need to be precise about the claim of being "on par with Opus 4.8," as suggested by the original wccftech article regarding V4 Flash 0731's performance. In the benchmark table DeepSeek published on Hugging Face, Opus 4.8 outperforms V4 Flash 0731 across all 9 metrics.
| Benchmark | V4 Flash 0731 | Opus 4.8 | Difference |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 85.0 | -2.3 |
| NL2Repo | 54.2 | 69.7 | -15.5 |
| Cybergym | 76.7 | 83.1 | -6.4 |
| DeepSWE | 54.4 | 58.0 | -3.6 |
| Toolathlon-Verified | 70.3 | 76.2 | -5.9 |
The gap ranges from 2.3 points to 15.5 points. Notably, on NL2Repo, which measures the generation of large-scale code repositories, the gap exceeds 15 points—hardly what one would call "on par." That said, reaching this level with only 13 billion active parameters is itself noteworthy in terms of architectural efficiency. DeepSeek itself describes the model in its model card as "broadly competitive," and the distance between "comparable" and "equivalent" is something worth measuring precisely.
What stands out is that V4 Flash 0731 outperforms the company's own higher-tier model, V4 Pro (Preview), on all 9 metrics. On Terminal Bench 2.1, Flash scores 82.7 against Pro's 72.1; on DeepSWE, Flash's 54.4 against Pro's 12.8. Through post-training improvements alone, a smaller model surpassed a larger one. If this is reproduced by third parties, it would force a revision of conventional assumptions about the relationship between model size and performance.
What the price hike notice didn't reveal
The notice DeepSeek posted on its developer platform is remarkably sparse on details. It states only that a "significant" price increase will be implemented "in the near future," without disclosing the new pricing figures or an effective date. It urges developers to adjust their usage plans ahead of the formal announcement.
It's also worth noting that this price increase affects only API users, not those who run the open-weight models on their own infrastructure. DeepSeek's models are released under the MIT license, and downloading the weights to run them on one's own GPUs remains free. In other words, this price hike is a repricing of the hosting service DeepSeek itself operates, not a shift in its open-source strategy per se.
So why hasn't DeepSeek disclosed the reasoning? Infrastructure strain from surging demand is the most natural explanation, but according to Bloomberg reporting, plans are also underway to build a 1-gigawatt-scale data center in Ulanqab, Inner Mongolia. The company reportedly hopes to bring at least part of this facility online between late 2027 and early 2028. Is the price hike meant to curb demand, to secure funding for infrastructure investment, or both? DeepSeek remains silent.
The shape of the price war is shifting
Laying out this sequence of events chronologically reveals how the dynamics of the AI price war reversed within a single week.
| Date | Event |
|---|---|
| July 30 | OpenAI cuts GPT-5.6 Luna prices by up to 80% |
| July 31 | DeepSeek releases V4 Flash 0731, pricing it even lower than Luna |
| August 1 | Processes 8 trillion tokens in a single day on OpenCode |
| August 4 | Performance degradation occurs on V4 Flash API due to capacity shortage |
| August 5–6 | DeepSeek notifies developers of a "significant price increase" |
OpenAI's price cut was widely interpreted as intended to drag DeepSeek into a price war. DeepSeek responded by offering an even lower price. But winning a price war and having the computing resources to sustain that price turned out to be two different things. Liang Wenfeng, DeepSeek's founder, is reported to have said at an investor meeting that "the biggest difference from the U.S. lies in resources. Differences in talent, model performance, and applications all trace back to differences in computing resources." These words carry heavy weight behind the scenes of the price hike notice.
In China's AI market, Beijing has been warning tech companies against "involution"—that is, attritional price-cutting competition. There's an intent to avoid a repeat of the price collapses seen in solar panels and electric vehicles. DeepSeek's price increase aligns with this policy direction as well.
Questions that remain
Once the specific figures for the price increase become known, we'll finally get a sense of where DeepSeek is trying to find its equilibrium point. How many times will the current $0.14/$0.28 multiply? Combined with the peak-hour doubling mechanism, where will the effective price ultimately land? Depending on the answer, the usage rankings on OpenRouter could shift significantly. It remains unclear how the current situation—where Chinese models occupy nearly all of the top spots in weekly token usage—will change as a result of the price increase.
Another question concerns the feasibility of the 1-gigawatt plan. Ulanqab already has a concentration of data centers from companies like Alibaba and ByteDance, and across Inner Mongolia as a whole, 28 large-scale computing facilities are reportedly under construction or renovation in the first half of 2026, with total investment of roughly 117.6 billion yuan. How much priority DeepSeek's plan holds within this landscape, and which chips it will use (NVIDIA or Huawei), remain undetermined.
DeepSeek made its name as "the company that made AI cheap." That same company has now entered a phase where it must pay the price for that cheapness itself. As long as there's a physical ceiling on computing resources, a strategy of competing on price alone has its limits. Can the gap between the token volume that 20,000 GPUs can process and the token volume the market demands be closed by a price increase alone? Or will developers wait out the time it takes for the 1-gigawatt data center to be completed? The answer has yet to be determined.