www.latent.space/p/ainews-sci-fi-with-a-touch-of-madness

This sentence mixes up the model maker with the evaluation creator.

• Claude Opus 4.6 is an Anthropic model. Anthropic’s official model page is titled “Claude Opus 4.6” and says “Claude Opus 4.6 is our most capable model to date.”
• Andon Labs is the company behind Vending-Bench Arena, the benchmark/eval environment. On Andon’s own page, Round #4 lists the participants as “Claude Opus 4.6, Gemini 3 Pro, Claude Opus 4.5, GPT-5.2”—showing Andon is evaluating external models, not claiming Opus 4.6 as its own model.

So the article’s wording is incorrect: Andon Labs created the benchmark, but Opus 4.6 itself is made by Anthropic.

This attribution is incorrect.

• Neuronpedia’s own site says it is an open source interpretability platform and states: “Neuronpedia was created by Johnny Lin” and is “supported by Decode Research … Anthropic, and others.” Support is not the same thing as ownership.
• Neuronpedia’s privacy page is even more explicit: “Neuronpedia is built and operated by Decode Research.”
• Anthropic’s circuit-tracing announcement also distinguishes the two organizations. Anthropic says its open-source circuit-tracing release includes “a frontend hosted by Neuronpedia” and that “The Neuronpedia integration was implemented by Decode Research.”

So Neuronpedia may collaborate with Anthropic, but it is not “by Anthropic.”

This repeats the same attribution error elsewhere in the post.

According to Neuronpedia’s own site, “Neuronpedia was created by Johnny Lin” and is supported by Decode Research … Anthropic, and others. Its privacy page states even more directly that “Neuronpedia is built and operated by Decode Research.” Anthropic’s own research post on circuit tracing says the Neuronpedia integration was implemented by Decode Research and describes Neuronpedia as the frontend hosting the graphs.

So Anthropic has collaborated with Neuronpedia, but Neuronpedia is not 'from Anthropic.'