Did DeepSeek V4 Pro Secretly Use Claude? Here's What We Know

A fact-check explainer on the latest AI industry controversy — and why it's more complicated than it looks.

The Tweet That Started It All

In July 2026, a tech blogger known as Leo on X (formerly Twitter) posted a thread that quickly went viral in AI circles. The claim was provocative: DeepSeek's latest model, V4 Pro, might be secretly routing certain user queries to Anthropic's Claude Fable 5 behind the scenes.

Leo came with receipts. Using OpenCode, a tool for direct API interaction, he ran tests against DeepSeek V4 Pro with a simple task: generate a 3D game. What came back was anything but ordinary.

Portions of the generated code bore an uncanny resemblance to Claude Fable 5's output patterns — not just in syntactic style, but in reasoning structure. The problem breakdown, formatting, even comment styles all deviated from what developers had come to expect from DeepSeek models.

Then came the twist. When Leo mixed in prompts on cybersecurity and biohacking — topics that trigger safety filters — the game output quality dropped noticeably. The knowledge range contracted, and the reasoning reverted to standard DeepSeek behavior.

The implication: DeepSeek might be running a routing layer that sends "high-value" queries to Claude, while keeping sensitive queries on its own infrastructure. Users get Claude-level performance at DeepSeek prices — and DeepSeek potentially gets high-quality training data for model distillation.

The Price Gap That Makes You Think

To understand why this theory caught fire, look at the numbers. As of mid-2026, the pricing differential is staggering:

Model	Input Price	Output Price
DeepSeek V4 Pro	¥6/M tokens (off-peak)	¥12/M tokens (peak)
Claude Fable 5	$10/M tokens	$50/M tokens

At current exchange rates, Claude Fable 5 is roughly 50 to 60 times more expensive than DeepSeek V4 Pro. For a developer building a product on top of these APIs, the math is unambiguous. If DeepSeek V4 Pro genuinely delivers Claude-level performance even part of the time, the value proposition is overwhelming.

This price gap is the economic engine behind the entire controversy. If someone were going to arbitrage AI models, this is exactly the spread they'd exploit.

How the Routing Theory Works

Let's break down the technical mechanism that Leo and others have proposed. It's not as simple as "DeepSeek forwards everything to Claude." The theory is more nuanced, involving a multi-step routing architecture:

It's an elegant theory. But is it true?

The Evidence: What We Actually Know

Let's separate what has been demonstrated from what is being inferred.

What has been shown:

• Output from DeepSeek V4 Pro's API sometimes exhibits stylistic patterns that closely resemble Claude Fable 5's code generation style.
• When safety-sensitive content is mixed into prompts, the output quality and knowledge scope degrade.
• The degradation pattern is consistent with the theory that a different model — or a different configuration — handles these queries.

What has not been shown:

• No direct evidence that DeepSeek's servers are actually calling Anthropic's API endpoints.
• No packet tracing or network-level evidence of cross-company API calls.
• DeepSeek has not acknowledged any such arrangement, and Anthropic has not commented on this specific incident.

The observed behavior could have alternative explanations. DeepSeek may have trained V4 Pro on a dataset that includes Claude-generated outputs (a form of offline distillation, which is different from the live routing being alleged). The model could be using an internal mixture-of-experts architecture that routes different types of queries to different sub-models — all within DeepSeek's own infrastructure. Or the prompt sensitivity could simply reflect genuine safety training that DeepSeek applied to its own model.

None of these alternatives require DeepSeek to be secretly paying Anthropic's API bills.

The Elephant in the Room: Anthropic's February 2026 Accusation

This is not the first time DeepSeek has been accused of tapping into Claude's capabilities. In February 2026 — just five months before Leo's tweet — Anthropic made a dramatic public accusation.

Anthropic alleged that three Chinese AI companies — DeepSeek, Moonshot AI (月之暗面), and MiniMax — had orchestrated what they called an "industrial-scale distillation attack." The numbers were staggering: approximately 24,000 fake accounts had been used to conduct over 16 million interactions with Claude's models.

"They train on everyone's public data, then object when someone tries to learn from their outputs." — Satya Nadella, Microsoft CEO, on the double standard in AI training data debates

The accusation touched a nerve in the AI industry. But it didn't go unanswered.

Elon Musk fired back, accusing Anthropic of hypocrisy — essentially "a thief crying thief." The charge stung particularly because Anthropic, like most AI labs, has built its models on vast quantities of publicly available internet data, much of it created by humans who never consented to having their work used for AI training.

Then came a more technical rebuttal. Salvatore Sanfilippo, the creator of Redis (a widely-used open-source database), argued that distillation through public APIs is "impossible" in any meaningful sense. The reason is technical: true model distillation requires access to the full logits — the complete probability distribution over all possible output tokens — not just the final text response. What a public API returns is just the surface-level output, stripped of the rich internal signal that makes distillation effective.

The Chinese Internet Reacts: "Reverse Subsidy"

While the Western tech press and AI community debated the ethics and implications, Chinese social media took a distinctly different tone.

The prevailing sentiment on platforms like Weibo and Zhihu could be summarized as amused pragmatism. Users joked about "反向百亿补贴" — a term that roughly translates to "reverse billion-dollar subsidy," playing on the Chinese e-commerce practice of platforms subsidizing purchases to attract users. The joke: getting Claude-quality output at DeepSeek prices was essentially Anthropic unwittingly subsidizing Chinese developers.

"Wait, there's actually this kind of deal? Better use it more, then." — Popular comment on Chinese social media

"Spending DeepSeek money to get Claude's computing power," another user quipped. "This is the best deal of 2026."

The contrast reveals something deeper about how AI ethics debates play out across cultures. In the West, the focus was on the alleged violation — that if the accusation were true, it would represent unfair exploitation. In China, the focus was on the practical outcome: if the service is better and cheaper, who exactly is losing?

Neither reaction is entirely right or wrong. They reflect different baseline assumptions about competition, intellectual property, and the relationship between AI companies and their users.

DeepSeek's Silence — and What It Might Mean

As of writing, DeepSeek has not responded to the API routing allegations. In the AI industry, silence is often the most strategically ambiguous option — it avoids lending credibility to the accusation while also avoiding statements that could later be proven false.

DeepSeek has historically been tight-lipped about its architecture. Its models are open-weight but not fully open-source: weights are available, but detailed training methodology, data sources, and infrastructure decisions are not publicly documented. This is not unusual — OpenAI, Anthropic, and Google are similarly opaque — but it means external observers must infer behavior from output rather than disclosed architecture.

The Bigger Picture: Distillation, Open Source, and the Rules That Don't Exist

This controversy doesn't exist in a vacuum. It's the latest flashpoint in a broader, unresolved debate about the rules of AI development.

Model distillation has no clear legal status. The technique — training a smaller model to mimic a larger one — is well-established in the research literature. But when the "teacher" model is a commercial product accessed through a paid API, the legal and ethical boundaries become murky. Is it fair use? A terms-of-service violation? Industrial espionage? Different jurisdictions would likely give different answers.

The regulatory landscape is evolving. In May 2025, the White House issued an "anti-distillation" policy memorandum, signaling that the U.S. government views unauthorized model distillation as a national security concern. But the memo stopped short of creating enforceable law, and no equivalent framework exists internationally. The global AI industry is operating in a regulatory vacuum.

The open-source vs. closed-source battle is intensifying. DeepSeek has positioned itself as a champion of open-weight models, releasing its models publicly and enabling widespread access. Anthropic, by contrast, has maintained a more controlled approach. The distillation debate is, in part, a proxy for this larger philosophical divide. If you believe model weights should be freely available, you're likely to see distillation as a natural part of the ecosystem. If you believe models are proprietary products that require controlled access, distillation looks more like theft.

The uncomfortable reality: Distillation — in some form — is almost certainly standard practice. Every major AI lab trains on data that includes outputs from other models. The question is not whether distillation happens, but at what scale, through what methods, and with what degree of transparency.

So, Did DeepSeek Secretly Use Claude?

Based on the currently available evidence, the honest answer is: we don't know.

What we have is a set of behavioral observations that are consistent with the routing theory, but not exclusive to it. The output patterns could be explained by several alternative mechanisms — training data contamination, internal routing within DeepSeek's own model family, or safety training effects — none of which require live API forwarding to Claude.

What we can say is that the AI industry is entering a phase where the boundaries between models, companies, and capabilities are increasingly blurred. The tools that made the internet what it is — scraping, remixing, repurposing — are being applied to AI models themselves, and the legal and ethical frameworks haven't caught up.

Until DeepSeek speaks, or until someone produces network-level evidence of cross-company API calls, the controversy remains in the realm of informed speculation. But informed speculation, in this industry, often turns out to be a preview of the next big story.