Same family, different implementations
Both embed a keyed statistical signal in word choice during generation. Neither adds visible marks or metadata. Both are therefore carried by copy and paste, both weaken with heavy editing, translation and short length, and both need a key to detect.
What OpenAI claims, and what it has not shown
OpenAI says textGrain matched or exceeded the approaches it tested, including SynthID for text, and reports no meaningful benchmark loss from watermarking. These are vendor figures from OpenAI’s own evaluation, and the full methodology is in a technical report that OpenAI says it will extend. No independent head-to-head is available yet.
Detection access
OpenAI’s detector is limited to approved researchers and expert organisations. OpenAI also marks images and audio with SynthID and Content Credentials, and those verification tools are public, which is a separate matter from text. See /guide/chatgpt-vs-claude-watermark for the two chat assistants side by side.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.