# DeepSeek API tests show Claude-like behavior under selected prompts > Source: > Published: 2026-07-18 17:34:38+00:00 An independent test of DeepSeek-V4-Pro found sharply different behavior depending on whether prompts contained cyber or biology content, raising new questions about whether some requests were being answered with outputs from Anthropic's Claude Fable 5. In [a nine-post thread on X](https://x.com/synthwavedd/status/2078514339552628880?s=46) on July 18th, [leo (@synthwavedd)](https://x.com/synthwavedd) said a DeepSeek API session accessed through OpenCode initially answered a knowledge test as if its training data ended around May 2025. After the tester combined the knowledge questions with a request to build a 3D game, the response cited products and events from later in 2025 and closely matched an answer produced separately by Fable 5. The behavior changed again when the prompt included cyber or biology questions that the tester said activated Fable 5's safety classifiers. Under those conditions, the API returned to the older knowledge profile attributed to DeepSeek V4 and produced a substantially weaker game. Removing the sensitive material restored the stronger game output. The screenshots support the narrower finding that the tested endpoint behaved inconsistently across related prompts. They do not establish that DeepSeek collected Fable 5 outputs, identify the system that generated any individual response, or show that those outputs entered a training dataset. Prompt sensitivity, an undisclosed model update, server-side routing, or differences introduced by combining several tasks could also produce divergent results. Similar code and visual design from a common game prompt are suggestive evidence, rather than a model fingerprint. A stronger test would require repeated trials, identical prompts with one variable changed at a time, API response metadata, network records and comparisons against locally hosted DeepSeek V4 weights. The thread supplies screenshots from a small number of runs and does not provide request logs or a reproducible test suite. ### Anthropic had already accused DeepSeek The allegation lands against a documented dispute between the two AI developers. On February 23rd, [Anthropic said it had identified an industrial-scale campaign](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks?via=free) by DeepSeek that generated more than 150,000 exchanges with Claude. Anthropic said the traffic targeted reasoning, grading and censorship-related tasks. Anthropic attributed the activity to DeepSeek through request metadata, IP correlations, infrastructure indicators, synchronized traffic and shared payment methods. Anthropic also said prompts sought step-by-step reconstructions of internal reasoning that could serve as chain-of-thought training data. Those findings came from Anthropic's own systems and concerned an earlier campaign. They provide credible precedent for the July 18th claim, while leaving the new Fable 5 allegation unverified. Distillation is a standard model-development technique in which a smaller or cheaper model learns from responses generated by a stronger model. The dispute concerns access and authorization. Anthropic says competitors have used fraudulent accounts and proxy networks to extract Claude's capabilities at scale, violating its terms and regional restrictions. Anthropic said in February that it was building classifiers, behavioral fingerprinting, stronger account verification and model-level countermeasures. The company described the traffic patterns, volume and coordinated infrastructure as the evidence separating an extraction campaign from ordinary API use. ### The timeline narrows what the test could mean DeepSeek [released V4-Pro and V4-Flash on April 24th](https://api-docs.deepseek.com/updates/), according to its API change log. Anthropic introduced Fable 5 on June 9th, more than six weeks later. Anthropic describes Fable 5 as the same underlying model as Mythos 5, with additional safeguards that route many cyber and biology requests to Opus 4.8. That sequence means Fable 5 outputs could not have been part of DeepSeek V4's original pre-release training under the public timeline. If the observed behavior came from Fable 5, the plausible mechanisms would involve activity after V4's April release: continued optimization, an updated API model, live request routing, or output collection for a future system. DeepSeek's public materials do not resolve that issue. Its [V4 model card](https://fe-static.deepseek.com/chat/transparency/deepseek-V4-model-card-EN.pdf) says optimization data is produced manually or automatically by DeepSeek researchers, with a small portion potentially based on user inputs. DeepSeek says its data acquisition complies with intellectual-property and privacy laws, but the document does not identify outside models used to generate question-answer pairs. DeepSeek also distributes V4 as open weights under the MIT license and through its hosted API. That distinction matters because API behavior can diverge from the downloadable weights through routing, system prompts, safety layers or unannounced serving changes. DeepSeek's current documentation lists V4-Pro as a 1.6-trillion-parameter mixture-of-experts model with 49 billion parameters active per token and a one-million-token context window. The July 18th test has therefore surfaced a testable anomaly rather than a completed attribution. Anthropic's earlier evidence established that DeepSeek had sought Claude outputs at scale. Determining whether Fable 5 is involved in the current DeepSeek API requires telemetry that screenshots alone cannot supply.