DeepSeek is pushing its cheaper Flash model to compete directly with its own Pro version through aggressive price cuts and engineering improvements, while U.S. federal agencies simultaneously accuse the company of running industrial-scale campaigns to extract proprietary features from leading American AI models. The company launched a closed beta of Flash V4.1 with a new architecture and native multimodal capabilities, then asked users directly whether the model could fully replace Pro. Meanwhile, the Cybersecurity and Infrastructure Security Agency (CISA), National Security Agency (NSA), and Federal Bureau of Investigation (FBI) released a joint advisory on September 9 accusing DeepSeek and five other Chinese AI firms of extracting billions of tokens from U.S. systems since late 2024.

Why Is DeepSeek Trying to Make Flash Compete With Pro?

The economics are compelling. During peak hours, Flash costs 9 yuan per million output tokens, compared with 27 yuan for Pro, roughly one-third the price. On September 10, DeepSeek announced another round of price cuts for Flash on its open platform. Cached input tokens will drop 60 percent to 0.02 yuan per million, uncached input will fall about 33 percent to 1 yuan, and output will decrease roughly 11 percent to 4 yuan. For developers running high-volume applications, the savings could be substantial if Flash can handle the work.

The challenge is that cheaper models often require more computational steps to reach the same result as larger ones. A small model might need to make multiple API calls, each incurring input and output charges, while a larger model solves the problem in one pass. DeepSeek is trying to solve this by improving what it calls "intelligence density," the efficiency of its reasoning chains. The goal is to let Flash accomplish what Pro does, but faster and with fewer wasted steps.

How Is DeepSeek Improving Flash's Efficiency?

DeepSeek has been working on multiple technical improvements to make smaller models more capable. In a research paper published in July called DSpark, the team reported that under live V4-Flash user traffic, per-user generation speed improved by 60 to 85 percent compared with a baseline approach. The improvements focused on eliminating wasted verification steps in a technique called speculative decoding, allowing the system to deliver answers to users faster without reducing overall service capacity.

The company is also addressing a fundamental shift in how AI models are evaluated. Researchers have begun incorporating user wait times into optimization targets, recognizing that the most compute-efficient approach does not necessarily deliver the fastest response. DeepSeek must balance two competing demands: making Flash more capable, and making it respond quickly enough that users accept it as a Pro replacement.

What Are the Real-World Costs of Using AI Agents?

The price cuts matter most for developers building AI agents, systems that can break down complex tasks into multiple steps and call external tools. A user on a Chinese tech forum named "Jueqian" recently used the new Flash model to generate a maintenance page for scheduled downtime. The model performed impressively, finding deprecated requirements in project guidelines, planning documentation updates, and modifying code to match existing styles. But the total cost was 15.5 yuan.

This reveals why per-token pricing alone does not tell the full story. An agent might make dozens of API calls to complete a single user request: reading project files, checking existing code, modifying templates, and reviewing results. Each call adds input and output charges. A low per-token price does not guarantee that the overall job will be cheap.

How to Reduce Agent Costs and Improve Efficiency

• Orchestration Tools: DeepSeek's open-source Harness framework helps developers structure agent workflows by separating model decisions from tool execution. The model decides what to do, while Harness manages the tools, sessions, and runtime environment, reducing unnecessary API calls.
• Batch Processing: Harness includes a "PTC mode" that allows models to write programs chaining multiple tool operations in a single pass. For example, instead of asking the model to decide what to do after each file read, the model can write a program to read and filter files in batches, returning only relevant results.
• Context Efficiency: Developers should avoid sending intermediate results back to the model step by step. Instead, tools should handle routine tasks like reading, filtering, and organizing material according to established rules, keeping context usage low.

What Are U.S. Agencies Accusing DeepSeek Of?

On September 9, CISA, the NSA, and the FBI released a joint advisory accusing DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI of running organized campaigns to extract proprietary features from American AI models.

According to the agencies, DeepSeek has been conducting an organized campaign since at least late 2024 to feed synthetic training data derived from other companies' models into its R1 and V3 systems. The targets included multiple versions of Claude, Gemini, GPT-4, GPT-4o, GPT-5, and Grok 4. Officials called DeepSeek's widely cited $5.6 million training cost misleading, arguing it excludes the cost of data obtained through what the agencies termed "malicious" knowledge distillation.

"We strongly urge AI companies to take immediate steps to safeguard their platforms against knowledge distillation campaigns that threaten to close the gap in advancements made by American companies," said Nick Andersen.

Nick Andersen, Acting Director, Cybersecurity and Infrastructure Security Agency

Knowledge distillation is a standard research technique where a smaller model learns from the outputs of a larger one. The three agencies differentiated between legitimate research and what they described as "aggressive, malicious, and targeted distillation activities at an industrial scale." The advisory stated that the campaigns "allow those actors to deliberately strip away security protocols from the resulting models and undo mechanisms that ensure those AI models are ideologically neutral and truth-seeking".

What Specific Models Were Targeted?

The agencies detailed distillation campaigns targeting a range of American AI systems. Moonshot AI, also known as Beijing Moonshot Technology Co. Ltd., was accused of pulling substantial Claude Fable 5 data for its Kimi-K3 model and GPT-4o data for Kimi-K2 since at least mid-2025. Alibaba distilled Claude-4, Claude Opus, Claude Sonnet, and GPT-5 in late 2025 to improve software engineering, customer service, and image creation for its Qwen family of models. MiniMax used Claude Code, Claude Sonnet 4, Claude Opus, and several Gemini versions to improve its M2 model, and even attempted prompt injections to convince Claude Code it was actually a MiniMax product.

StepFun distilled multiple Claude and GPT-5 variants between late 2025 and early 2026 for its Step 4 model. By mid-2026, Z.AI had taken billions of tokens of GPT-5.5 and Claude Opus 4.8 data for chain-of-thought reasoning capabilities.

"Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable," stated Kratsios.

Kratsios

What Recommendations Did U.S. Agencies Make?

The advisory from CISA, NSA, and FBI recommended three defensive steps for U.S. model providers. First, companies should hunt for anomalous prompts, accounts, and usage spikes that might indicate distillation campaigns. Second, they should quietly degrade answers when a distillation campaign is suspected, making the stolen outputs less useful. Third, they should share intelligence across companies, cloud providers, and API aggregators to identify patterns and coordinate defenses.

The timing of these accusations coincides with DeepSeek's push to make Flash a viable alternative to Pro. If the allegations are accurate, DeepSeek's cost advantage may partly reflect not just engineering efficiency, but also the use of training data obtained from competitors without permission. The company has not publicly responded to the agencies' claims as of this writing.