On July 31, 2026, DeepSeek released DeepSeek-V4-Flash-0731 and began a public beta of its API. Pricing per million tokens is $0.14 for cache-miss input and $0.28 for output. Since OpenAI had cut GPT-5.6 Luna's price to $0.20 for input and $1.20 for output just the day before, DeepSeek's move might easily be read as an immediate counter-cut.
But DeepSeek hasn't changed its pricing. V4 Flash's $0.14 and $0.28 rates have been in place since before, and what was updated in 0731 is the model's capabilities and API features. The architecture also remains unchanged—280 billion total parameters, 13 billion active—with only re-post-training applied. With this update, the axis of comparison has expanded from unit price alone to how much work can be completed at the same price, i.e., task performance.
Furthermore, DeepSeek plans to introduce peak pricing for a total of 7 hours per day, doubling all billing items during that window. The effective date has not yet been set. Even though the low-price banner remains unchanged, the ranking against OpenAI could reverse depending on the time of use and cache-hit rate.
Price Held, but Post-Training Repeated
V4 Flash 0731 inherits the architecture and scale of the Preview released on April 24. Total parameters stand at 280 billion, but under the MoE (Mixture of Experts) approach, only 13 billion are active during inference for each token. The context window reaches 1 million tokens, with a maximum output of 384,000 tokens.
DeepSeek's changelog describes 0731 as having been "re-post-trained." It does not claim to have redone large-scale pretraining or to have made the model larger. The update also applies only to the V4 Flash API—the V4 Pro API, the app, and the web version remain as before. Simply setting the model name to deepseek-v4-flash in an existing API call switches users to the latest version without any change to how they invoke it.
On the feature side, native support for the Responses API has been added, along with configuration for Codex. Reasoning intensity can be selected from three levels: low, high, and max. The model weights are also released under the MIT License, allowing self-hosting via vLLM or SGLang. However, the vLLM configuration example DeepSeek provides is a single node loaded with four GB300 units. A low API price and light self-hosting infrastructure are not the same thing.
"Opus-Class" Claims Vary by Benchmark
According to DeepSeek's official table, 0731's Terminal Bench 2.1 score rose to 82.7, up 20.9 points from Preview's 61.8. DeepSWE also jumped 47.1 points, from 7.3 to 54.4, surpassing V4 Pro Preview's 12.8. Even with the same architecture, tightening agent-focused post-training and the execution environment can substantially move performance on tool-using tasks.
That said, it can't be lumped together as equivalent to Claude Opus 4.8. Across the 9 items DeepSeek listed, 0731 scored below Opus 4.8 in every case. The gap is small for some—Agents' Last Exam is 25.2 vs. 25.7, and Terminal Bench 2.1 is 82.7 vs. 85.0—but widens elsewhere, with NL2Repo at 54.2 vs. 69.7 and the internal DSBench-Hard evaluation at 59.6 vs. 71.7.
Evaluation conditions also warrant caution. For the public code-agent benchmarks, DeepSeek used the not-yet-released minimal mode of DeepSeek Harness with reasoning intensity set to max. The two DSBench items are internal tests. The capability of the model alone and the capability of the system including the harness need to be read separately.
In Artificial Analysis's independent evaluation as well, 0731's Intelligence Index rose from the old Flash's 40 to 50. However, this still falls 1 point short of GPT-5.6 Luna's 51 and 6 points short of Opus 4.8's 56. On the other hand, Artificial Analysis calculated—based on actual input and cache usage, and tokens consumed for reasoning and answers—that task cost was approximately 60% lower than Luna's. What 0731 has closed in on is not Opus overall, but first, Luna-level composite scores and the cost of reaching them.
The $0.14 Rate Becomes $0.28 After Peak Pricing Takes Effect
Currently, DeepSeek V4 Flash charges $0.14 per million tokens for cache-miss input, $0.0028 for cached input, and $0.28 for output. Luna's corresponding rates are $0.20, $0.02, and $1.20. Comparing standard pricing alone, DeepSeek is cheaper across all three categories.
The gap becomes visible when comparing equivalent workloads. Suppose you process 100,000 input tokens and 10,000 output tokens ten times, with zero cache hits. The total comes to 1 million input tokens and 100,000 output tokens. DeepSeek costs $0.168, while Luna costs $0.32—DeepSeek is 47.5% cheaper. Since each input is 100,000 tokens, this also avoids the long-context surcharge that kicks in for Luna above roughly 272,000 tokens.
However, DeepSeek has announced plans to double all billing items during Beijing time 9:00–12:00 and 14:00–18:00—corresponding to 10:00–13:00 and 15:00–19:00 Japan time. Running the same workload during peak hours after this takes effect would cost DeepSeek $0.336, exceeding Luna's $0.32 (assuming Luna's price stays unchanged) by 5%.
This is a projection based on planned pricing, not a current billing figure—and the effective date has not been announced. Still, it's clear that which time slots batch processing can be shifted to will become a factor in model selection. While DeepSeek's standard $0.0028 cached-input rate has a strong effect, sending a long prompt in a single call would incur Luna's doubled input rate and 1.5x output rate. Monthly billing cannot be predicted from the minimum values on a price sheet alone.
What a ~20,000-Unit Contract Reveals About the Upstream of Low-Cost APIs
What sustains the competition among low-cost models is not just sparse architectures and post-training, but also access to large-scale computing infrastructure. On July 31, Bloomberg reported that Moonshot AI has a compute contract with Alibaba covering roughly 20,000 Nvidia chips. According to anonymous sources, this infrastructure accounts for a significant portion of the total compute power used for the Kimi model family.
What Bloomberg reported is a usage contract—not a claim that Moonshot purchased and owns the chips. The published summary doesn't specify the chip model, nor does it state that the 20,000 units are dedicated exclusively to Kimi K3. Even so, a contract connecting to Nvidia compute resources via Alibaba's cloud offers concrete evidence of the supply chain underpinning China's open-weight development.
Kimi K3 has 2.8 trillion total parameters and 104 billion active parameters, selecting 16 out of 896 experts to run. The sparse architecture avoids computing all parameters every time, but Moonshot itself recommends supernodes equipped with 64 or more accelerators even for inference. The company explains that at this scale of sparsity, routing and optimization become the main challenges, with uneven load distribution across experts slowing down processing.
Both DeepSeek's $0.14 pricing and Moonshot's open weights rest on large-scale compute infrastructure. What has been confirmed this time is that Moonshot uses roughly 20,000 units under contract, and DeepSeek plans to double pricing during congested hours. When comparing model performance against unit price, such supply-side conditions also change the actual billing amount and available capacity. When will DeepSeek's peak pricing begin, and how will task costs measured under the same harness and quality standards change? These two questions—more than the next price cut—will make the real shape of this price competition clear.