OpenAI ships text watermarking: opt-in for the API, auto in the EU
Off by default in the API, automatic for EU ChatGPT and Codex, and a rewrite erodes it. Plus Nova 2.5 Sonic, Mistral Large 4, and an open image model.
Copy markdown
Turn it on at the org level, not per request
OpenAI's textGrain nudges the model's word choices into a statistical pattern a detector can read. It's off by default in the API — flip it on per project or organization and choose which models carry it, with no change to individual requests. OpenAI says a watermarked Astra showed no measurable hit on coding and agent benchmarks (DeepSWE, Terminal-Bench), and it plans to open-source the scheme.
The EU gets it automatically — a rewrite strips it
Eligible ChatGPT and Codex text for EU users will be watermarked automatically 'in the coming weeks' under the EU AI Act, with no opt-out — mirroring Anthropic's global Claude watermark from August. But the signal is fragile: swapping 10% of words in a 400-token passage dropped detection from 92% to 66%, and 25% took it to 17%. Code is harder to mark, and detector access is gated to approved researchers, so you can't verify your own output.
Amazon Nova 2.5 Sonic goes GA for voice agents
Amazon's speech-to-speech model is generally available on Bedrock across four regions, with AWS citing better reasoning, instruction-following, and tool-calling accuracy plus lower latency than Nova 2 Sonic. If you're building a real-time voice agent, it's a single model from audio in to audio out — no separate STT and TTS stack to stitch together.
Mistral Large 4 preview — open weights land late October
Mistral opened a public preview of Large 4, a ~1T-parameter multimodal model with roughly 49B active per token, positioned as an open-weight frontier model. Weights are slated to drop later this month, so you can prototype against the preview now and plan to self-host once they land. Behavior may still shift before the release.
Google rolls out Nano Banana 2.1 image model
Google's latest image generation and editing model in the Gemini 3 family is live via Gemini, AI Studio, and the Gemini API, taking text and image inputs. It's the fastest path to swap image gen or in-place edits into an app you already wire to Gemini — no new SDK, just a model string.