There’s nothing more exhilarating, exhausting and tumultuous to be covering in the world right now than AI. There’s a major development, planned or unplanned virtually every day. You know you’re obsessed when you’re spending the evening of your birthday obsessing over Chinese model “mini” releases. But what DeepSeek released today, essentially a *post training* corporate focused upgrade of its v4 release with weights, via API, and frontier level capabilities, feels pretty revolutionary and kind of obsession worthy.

The architecture did not change. The agent benchmark results did.

DeepSeek released the official version of DeepSeek-V4-Flash on July 31, turning the smaller member of its V4 family into one of the strongest and least expensive agentic models currently available.

The release, formally named DeepSeek-V4-Flash-0731, is unusual for what DeepSeek did not change. It has the same architecture and model size as the V4-Flash preview released in April. DeepSeek says it only repeated the post-training stage. This is a bit like taking a suit in to be tailored and getting back a Versace.

The resulting model substantially outperforms the preview and beats the much larger V4-Pro preview across DeepSeek’s selected agent benchmarks. It is a striking demonstration of how much operational capability can now be added after pretraining, without increasing the underlying model’s scale.

That changes the economics of model development. Pretraining V4-Flash required more than 32 trillion tokens; DeepSeek did not repeat that run. It reused the pretrained base and manufactured a major operational capability gain during post-training. If the result generalizes, the marginal contest shifts from financing ever-larger base models alone toward building better training trajectories, reward systems, specialist models and agent environments afterward. The base model becomes reusable infrastructure. Post-training now becomes the product cycle.

And the early cracks in American AI are beginning to show. Only weeks ago, one of the purest investment theses in the space, a company lauded breathlessly for turning a $225 million bet on AI infrastructure into $5 billion, then $20 billion, Situational Awareness seemed visionary and nearly invincible. The hedge fund founded by 25-year-old former OpenAI researcher Leopold Aschenbrenner had grown from a few hundred million dollars to approximately $20 billion by converting his AI-infrastructure thesis into heavily leveraged market positions. Built around the thesis in his influential 2024 essay that an enormous AI infrastructure buildout was both imminent and inevitable, the fund concentrated heavily in companies such as Broadcom, Intel and CoreWeave and amplified those positions with borrowed money. The strategy produced a reported 439 per cent return during the first six months of 2026 and helped the fund grow to roughly $20 billion.

But then the Chinese models hit, and the trade reversed. AI-related stocks fell sharply, Situational lost about 67 per cent in July, and margin pressure turned a market correction into a liquidity event. After seeking additional capital, the fund sold the bulk of its public-equity portfolio to Ken Griffin’s Citadel, which acquired the broker-financed portion of the book. Citadel did not take over Situational itself: the fund retained roughly $10 billion in other assets, including its private stake in Anthropic (which it reportedly also at one point tried to sell at least part of). Reuters reported the transaction and fund figures.

The sale shows that the first structural cracks are becoming visible. Situational Awareness was the financial expression of the strongest version of the American AI thesis: rising capability would require extraordinary quantities of chips, energy and data-centre capacity, making the companies supplying that buildout an almost one-way investment. One violent month was enough to expose how much of that confidence had been reinforced by leverage and rising asset prices. DeepSeek’s V4-Flash release makes the episode more consequential. An open-weight Chinese model is now delivering near-frontier agent performance at a fraction of American API prices just as public markets are beginning to question whether every layer of the US infrastructure buildout can earn an adequate return. The pressure on American AI is therefore arriving from both directions: falling costs and stronger competition at the model layer, alongside weakening tolerance for unlimited spending at the capital layer.

What DeepSeek released

V4-Flash remains a mixture-of-experts model with 284 billion total parameters and approximately 13 billion activated for each token. It retains a one-million-token context window and now offers low, high and max reasoning-effort settings. A speculative decoding module called DSpark is attached to improve generation speed.

The official hosted API is available in public beta under the existing deepseek-v4-flash model name. It supports OpenAI-compatible and Anthropic-compatible interfaces, along with the Responses API used by coding agents such as Codex. The DeepSeek consumer app and website have not been upgraded to the new model.

DeepSeek has also published the model weights under an MIT licence. The Hugging Face repository is approximately 167 GB, making the release openly deployable, although practical local use still requires substantial hardware or quantization.

The official version of V4-Pro has yet to be released.

The agent results

DeepSeek reports gains across terminal use, software engineering, tool use, automation and cybersecurity environments:

The DeepSWE result is particularly dramatic: the smaller model rises from 7.3 to 54.4 *after post-training*. Its CyberGym score nearly doubles.

Without sounding breathless, this is a revolution in what is possible. Artificial Analysis gives V4-Flash-0731 an Intelligence Index score of 50, ten points above the April preview and six points above V4-Pro. That puts it one point behind GPT-5.6 Luna at max reasoning effort on the same index. Its score on Artificial Analysis’s agentic real-world work evaluation rose from an Elo rating of 1,189 to 1,559. While results don’t quite establish that V4-Flash is universally equal to the leading closed models, they do show that the upgrade extends beyond DeepSeek’s internal tests. Did I mention this is post training.

The benchmark is a system result

DeepSeek’s evaluation conditions are relevant. For its public coding-agent benchmarks, the company used the minimal mode of an unreleased DeepSeek Harness, selected max reasoning effort and set temperature to 1.0 and top-p to 0.95.

This makes comparisons between vendor reasoning labels increasingly less meaningful. DeepSeek now exposes low, high and max settings, while other companies use their own scales and allocate different amounts of test-time computation. DeepSeek’s strongest published agent results were produced at max. A shared label does not represent a shared reasoning budget, cost profile or probability of success.

The price is part of the release

DeepSeek is charging $0.14 per million uncached input tokens and $0.28 per million output tokens for V4-Flash. Cached input costs $0.0028 per million tokens. The company says it plans to introduce peak pricing at twice the standard rate during specified daytime periods in Beijing.

At either rate, the model applies significant price pressure to proprietary agent APIs. The combination of a one-million-token context window, open weights, strong coding-agent performance and extremely low hosted pricing makes it particularly relevant for high-volume development work and private or sovereign deployments.

More capability also means more operational reach

The release has a second implication. The same model architecture can now perform far more effectively inside terminal, automation and cybersecurity environments. If its permissions remain constant, its effective operational reach has increased.

None of the published benchmarks establishes a corresponding improvement in specification fidelity, boundary recognition or safe recovery from an ambiguous instruction. They measure successful task completion. An agent that completes more tasks can also travel farther after selecting an unintended path.

DeepSeek-V4-Flash-0731 is much more consequential than an ordinary point release. It shows that architecture and parameter count no longer describe the full capability transition. Post-training can substantially change what a model can accomplish, while reasoning allocation and agent scaffolding determine how much of that capability appears in practice.

The official V4-Pro release will show whether DeepSeek can reproduce the same gains at its larger scale. For now, V4-Flash has already made the central point: the base model is only the beginning of the deployed system.

Sources

Artificial Analysis evaluation

DeepSeek API change log

DeepSeek-V4-Flash-0731 model card and weights

DeepSeek API pricing