DeepSeek Accused Of Routing API Prompts To Claude For Distillation
An X investigation says some DeepSeek V4 API prompts produced Claude-like outputs, then fell back when safety filters appeared to trip.
Entities: DeepSeek, Anthropic, synthwavedd
A long X thread from @synthwavedd argues that DeepSeek's API may have routed some harder prompts to Anthropic's Claude Fable 5, then returned outputs close enough to suggest distillation rather than normal model behavior. The claim spread through quote-posts and replies on X, with posters pointing to coding outputs, chain-of-thought style and an apparent snap back to more typical DeepSeek behavior when prompts touched safety-sensitive topics.
Combined views
55.1K
12 posts, first seen 5h ago
Reactions from ranked influencers
12 posts
This sounds pretty silly and desperate. Using your first party API as a wrapper for a transfer station to Claude? I've never got Fable-shaped outputs from DS API, in style or in quality Devastating if gets confirmed
🧵 DeepSeek appear to have engaged, or be engaging in, a large-scale operation to collect outputs from proprietary models (including Claude Fable 5) for certain requests via their API as part of a distillation effort. After seeing such claims circulating earlier today, we conducted an investigation into them on our Discord. We found that, when "Deepseek V4" was used within OpenCode - via their official API - for complex prompts (i.e. 3D games) and combined with a knowledge-related query, the model provides virtually identical outputs to Fable 5. CoT structure is also very different from what is typically expected from Deepseek models. Both of these behaviours revert to what is expected for V4 when simpler prompts were used. When "Deepseek V4" was asked to incorporate answers to questions related to cyber or bio tasks that we verified hit Fable's classifiers into its 3D games, outputs tanked in quality. This is very difficult to explain unless the request was routed to Fable and fell back after hitting a classifier. For complex code prompts without anything else mixed in, like 3D games, outputs were remarkably similar to those produced by Fable 5. We were able to produce these results most consistently via OpenCode and the official Deepseek API combined with a prompt that specifies a complex code task. Deepseek have continued to modify their routing system since, as we have observed changes in behaviour and CoT style compared to those seen previously. Our investigation was conducted from 7AM-8AM PT.
OK I've seen enough Everyone using OpenCode is sus but I'll believe them The charitable read is that DeepSeek is indeed running the transfer station on their official first party API to collect trajectories, under the guise of "grayscale testing". It's an insane move. Tbh I don't mind much, that's smarter than A/B on your own checkpoints, though they likely still won't get CoTs. The worst possible version is there's no V4 GA in a meaningful sense, they've hit some roadblock that doesn't let them even reach ≈GLM 5.2 level (which would suffice to make V4 competitive), are playing for time and coping pathetically, burning money on proxies. In the end, this isn't Llama Reflection, this is a company with the best reputation in Chinese AI, worth > $50 billion. There must be weights, and these weights must have certain performance in independent deployments. This will undermine trust no matter how it works out, but if the model lands on huggingface and is decent, the damage will be mitigated. This is bad beyond DeepSeek, because Kimi's weights are also unavailable, and similar doubts are now possible. But I am pretty sure K3 is real and hope people won't generalize.
@teortaxesTex It would be pretty easy to test if you want to, there are spicy words you can add to any prompt that will set off the classifier. "alpha cobratoxin" is an antivenom that anthropic hates in any context. https://xlr8harder.github.io/archive/fnord/
Yikes
🧵 DeepSeek appear to have engaged, or be engaging in, a large-scale operation to collect outputs from proprietary models (including Claude Fable 5) for certain requests via their API as part of a distillation effort. After seeing such claims circulating earlier today, we conducted an investigation into them on our Discord. We found that, when "Deepseek V4" was used within OpenCode - via their official API - for complex prompts (i.e. 3D games) and combined with a knowledge-related query, the model provides virtually identical outputs to Fable 5. CoT structure is also very different from what is typically expected from Deepseek models. Both of these behaviours revert to what is expected for V4 when simpler prompts were used. When "Deepseek V4" was asked to incorporate answers to questions related to cyber or bio tasks that we verified hit Fable's classifiers into its 3D games, outputs tanked in quality. This is very difficult to explain unless the request was routed to Fable and fell back after hitting a classifier. For complex code prompts without anything else mixed in, like 3D games, outputs were remarkably similar to those produced by Fable 5. We were able to produce these results most consistently via OpenCode and the official Deepseek API combined with a prompt that specifies a complex code task. Deepseek have continued to modify their routing system since, as we have observed changes in behaviour and CoT style compared to those seen previously. Our investigation was conducted from 7AM-8AM PT.
At least one big Chinese lab seems to be routing hard requests through Claude to obtain distillation data I suppose 🙃
🧵 DeepSeek appear to have engaged, or be engaging in, a large-scale operation to collect outputs from proprietary models (including Claude Fable 5) for certain requests via their API as part of a distillation effort. After seeing such claims circulating earlier today, we conducted an investigation into them on our Discord. We found that, when "Deepseek V4" was used within OpenCode - via their official API - for complex prompts (i.e. 3D games) and combined with a knowledge-related query, the model provides virtually identical outputs to Fable 5. CoT structure is also very different from what is typically expected from Deepseek models. Both of these behaviours revert to what is expected for V4 when simpler prompts were used. When "Deepseek V4" was asked to incorporate answers to questions related to cyber or bio tasks that we verified hit Fable's classifiers into its 3D games, outputs tanked in quality. This is very difficult to explain unless the request was routed to Fable and fell back after hitting a classifier. For complex code prompts without anything else mixed in, like 3D games, outputs were remarkably similar to those produced by Fable 5. We were able to produce these results most consistently via OpenCode and the official Deepseek API combined with a prompt that specifies a complex code task. Deepseek have continued to modify their routing system since, as we have observed changes in behaviour and CoT style compared to those seen previously. Our investigation was conducted from 7AM-8AM PT.
🧵 DeepSeek appear to have engaged, or be engaging in, a large-scale operation to collect outputs from proprietary models (including Claude Fable 5) for certain requests via their API as part of a distillation effort. After seeing such claims circulating earlier today, we conducted an investigation into them on our Discord. We found that, when "Deepseek V4" was used within OpenCode - via their official API - for complex prompts (i.e. 3D games) and combined with a knowledge-related query, the model provides virtually identical outputs to Fable 5. CoT structure is also very different from what is typically expected from Deepseek models. Both of these behaviours revert to what is expected for V4 when simpler prompts were used. When "Deepseek V4" was asked to incorporate answers to questions related to cyber or bio tasks that we verified hit Fable's classifiers into its 3D games, outputs tanked in quality. This is very difficult to explain unless the request was routed to Fable and fell back after hitting a classifier. For complex code prompts without anything else mixed in, like 3D games, outputs were remarkably similar to those produced by Fable 5. We were able to produce these results most consistently via OpenCode and the official Deepseek API combined with a prompt that specifies a complex code task. Deepseek have continued to modify their routing system since, as we have observed changes in behaviour and CoT style compared to those seen previously. Our investigation was conducted from 7AM-8AM PT.