OpenAI text-moderation

text-moderation was the model OpenAI gave away. Introduced with the moderation endpoint on 10 August 2022, it read a piece of text and returned a score for each harm category rather than a sentence of prose — free of charge, on the condition that it was used to screen traffic going through OpenAI's own API. Developers wired it in front of and behind generative models, which is how a classifier nobody paid for ended up on the critical path of a large part of the industry. The final snapshot, text-moderation-007, arrived on 25 January 2024 alongside the new embedding models, and the text-moderation-latest and text-moderation-stable aliases were pointed at it. By then the category list had grown to eleven: hate, hate/threatening, harassment, harassment/threatening, self-harm, self-harm/intent, self-harm/instructions, sexual, sexual/minors, violence and violence/graphic. Input was text only, up to 32,768 tokens of output, with a knowledge cutoff of September 2021. Its successor, omni-moderation, shipped on 26 September 2024 and added image input plus two categories the older model never had — illicit and illicit/violent. OpenAI notified developers of the deprecation on 28 April 2025 and switched off all three identifiers, -007, -latest and -stable, on 27 October 2025. Ten months after that shutdown, the model still occupies two of the ninety-six entries in OpenAI's published model catalogue, with no note that it has stopped answering. This catalogue keeps the profile as a historical entry; the working model is omni-moderation.

▸Videos

No videos yet.

▸Reviews