Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

OpenAI extends text watermarking to API, off by default — diverging from Anthropic

OpenAI began phased deployment of textGrain text watermarking on October 5, automatic in the EU for ChatGPT and Codex, opt-in for API customers globally.

OpenAI began its phased deployment of textGrain on October 5, 2026, opening opt-in text watermarking to API customers globally and announcing that automatic watermarking for ChatGPT and Codex users in the European Union will follow in the coming weeks6,14. The company also opened applications for access to its text watermark detector on the same date.

The rollout responds to the EU AI Act's transparency rules, which took effect on August 2 and require AI companies to mark AI-generated content in a way other systems can identify2. Specifically, Article 50 of the EU AI Act obliges generative AI providers to ensure that synthetic audio, image, video, and text outputs are marked in a machine-readable format. OpenAI said it is working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.

Astra benchmark scores

OpenAI published benchmark scores showing near-identical performance from its Astra model with and without textGrain watermarking9. OpenAI reported Astra scored 93.94% on GPQA Diamond with watermarking versus 94.44% without, 71.68% on DeepSWE v1.1 versus 72.80% without, 56.06% on Terminal-Bench 4.0 versus 53.90%, 60.00% on Terminal-Bench Science 0.1 versus 56.90%, 34.86% on AutomationBench versus 34.09%, 87.35% on BrowseComp versus 87.92%, 64.60% on HealthBench Professional versus 64.27%, and 49.76 on the Artificial Analysis Intelligence Index versus 49.573. The company said it saw no meaningful change in its models' performance with watermarking switched on.

Opt-in versus on-by-default

The design choice sets OpenAI apart from Anthropic. Text watermarking will remain off by default in OpenAI's API, letting customers decide how watermarking fits their transparency obligations, the company said1. Anthropic, by contrast, said it would apply watermarking globally to supported Claude models at the model level, meaning it is present regardless of which Claude product or surface the text comes from. Anthropic does not describe an equivalent opt-out for API developers, according to The New Stack. Anthropic also said it did not yet have a reliable way to limit watermarking by region.

OpenAI said watermarking can be enabled at either the project or organization level, and no changes to individual API requests are required once it is turned on. Customers who opt into textGrain do not automatically gain the ability to detect its watermark, the company said. Detector access is initially limited to approved research and academic organizations studying text provenance and detection reliability. OpenAI also said the detector will report whether it detects an OpenAI watermark without identifying the user or revealing prompts or conversations.

Detection rates and fragility

OpenAI said its detector catches 95% of watermarked 400-token passages at a target false-positive rate of 1%. For shorter 200-token passages, detection drops to around 80% at the same false-positive rate. The company cautioned that shorter passages may contain too little material for reliable identification, and that detection is lower for more constrained material such as mathematics.

The watermark is also fragile under editing. In OpenAI's tests, replacing 10% of the words in a 400-token passage with synonyms reduced detection from around 92% to 66%, and replacing 25% brought it down to 17%. Source code is particularly difficult to watermark because there are fewer plausible word choices than in ordinary prose, the company said. OpenAI nonetheless plans to automatically watermark eligible Codex text output in the EU. OpenAI warned that high performance in ideal conditions does not guarantee reliable detection in everyday use4.

OpenAI said textGrain matched or exceeded SynthID's detection performance in its testing. The company published a technical report co-written with researchers from the University of Pennsylvania and Yale. OpenAI said textGrain gives it more control over the balance between watermark detectability and the variety of responses generated from the same prompt.

Broader provenance stack

The text watermark joins an existing provenance toolkit. OpenAI began adding Content Credentials, based on an open standard developed by the Coalition for Content Provenance and Authenticity, to generated images in 2024. It later added Google's SynthID watermarks to supported images in May 2026 and audio in July. OpenAI now offers a Content Provenance API for checking supported images and audio for those signals. OpenAI is a signatory to the EU's AI-content code alongside Google, Anthropic, Meta, Microsoft, and Mistral5.

OpenAI said it plans to open-source textGrain.

ANALYSIS The opt-in default for API customers means the vast majority of API-generated text will ship without a watermark unless individual organizations choose otherwise. The EU-specific automatic rollout for ChatGPT and Codex ties the feature's mandatory deployment directly to the EU AI Act's transparency obligations, creating a two-tier regime: watermarked consumer output in Europe, unwatermarked API output everywhere else unless developers act.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.