Daily LLM news, Sunday, 1 November 2026

Good afternoon. Watermarking leads today: xAI will invisibly mark Grok API text by default from 7 November, and Anthropic is extending its own text watermark to four older Claude models, with a loophole on two of them. OpenAI has a biology version of its flagship model for vetted labs only, Sam Altman says the company has spent hundreds of millions of dollars on AI maths, SenseTime has opened up a small model that runs its middle layers twice, and DeepSeek has published a design for an agent that juggles many tasks at once.

LEAD: Grok and Claude text is getting invisible watermarks, ahead of EU AI Act rules

Background: A text watermark nudges which words a model picks so the text can later be recognised as machine-written, without adding any visible marks or characters. The EU AI Act requires providers of AI systems that generate content to mark their output.

What happened: xAI, whose developer docs now call it SpaceXSI, published an Output Watermarking page on 31 October. From 7 November, text from grok-4.7, grok-4.6, grok-4.5, grok-4.3 and the three grok-4.20 variants carries an invisible watermark by default on every API endpoint, covering ordinary text, structured outputs and tool-call arguments in both the Responses and Chat Completions APIs. A new x-watermark response header says whether a reply was marked. Team admins can opt out per model, but only after confirming a legal exception applies, and some models may be locked on. Detection is statistical and available only by application. TokenPost reports the consumer Grok apps aren't affected.

Separately, Anthropic emailed API customers that it will watermark Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, adding to newer models such as Opus 5.5 that already carry the mark, according to The Decoder. The Decoder adds that the rollout on Amazon Bedrock may take a few extra days. On Opus 4.6 and Sonnet 4.6, the mark only applies when temperature, top_p and top_k stay at their defaults, so changing any one of them produces unmarked text. Anthropic's own advice: "If you rely on the watermark for your own compliance, keep these settings at their defaults." Opus 4.7 already rejects non-default values for those settings, so the gap there is effectively closed. Anthropic's detection tool is in private preview for regulators, researchers, media and similar groups.

Why it matters: Machine-written text from most of the big model makers will now be statistically traceable through their APIs. Neither xAI nor Anthropic has published accuracy or false-positive rates, and statistical watermarks tend to weaken on short, paraphrased or translated text. Developers using these APIs for anything where marked text could be a problem should check their settings before the start dates.

Sources: xAI watermarking docs, The Decoder, TokenPost

OpenAI releases GPT-Rosalind-6, a biology model for vetted labs only

Background: GPT-Rosalind is OpenAI's line of models for life sciences, offered only to approved organisations because it is more willing to answer biology questions that could be misused.

What happened: On 30 October OpenAI published a system card addendum for GPT-Rosalind-6, built on its GPT-6 Astra flagship. It is a reasoning model for biology, drug discovery and translational medicine, with fewer refusals on dual-use biology questions, and it is offered only through trusted access. Access requirements are the same as for the earlier GPT-Rosalind-5.5, with a separate route for government biodefence work. Approved customers must use least-privilege access controls and have qualified humans review outputs. It is available in OpenAI's Work and Codex products, and through the API as gpt-6-rosalind only where separately approved; it isn't in the Chat tab. API use isn't allowed for customer-facing products, and OpenAI hasn't published a price.

Why it matters: It's the clearest example yet of a lab shipping its strongest model in a less-restricted form to a gated group rather than to everyone. How well the vetting holds up is the real test, since the safeguard here is who gets in, not what the model refuses.

Sources: OpenAI GPT-Rosalind-6 system card addendum, OpenAI Help Center

Altman: OpenAI has spent hundreds of millions on AI maths, and has more results unpublished

Background: OpenAI has released hundreds of AI-generated maths results this autumn, including a claimed Navier-Stokes proof, and mathematicians have complained about thin write-ups, weak attribution and too little verification.

What happened: On the Cheeky Pint podcast with Stripe's John Collison, released Thursday and widely shared over the weekend, Sam Altman said the Navier-Stokes work cost about $10 million, "maybe a little more", and agreed that OpenAI's maths effort overall runs to hundreds of millions of dollars. He said the company deliberately let the maths push "get out of control" for a week to probe the limits of its models, with staff throwing the hardest problems they could find at them in a Slack channel, and confirmed OpenAI holds more results than it has released. He said he has empathy for mathematicians but feels "unequivocal" that "humanity deserves fundamental knowledge". The Verge notes OpenAI is also being criticised for withholding details of a claimed Yang-Mills result on the advice of experts it picked.

Why it matters: These are rare hard numbers on what frontier labs spend to point models at research problems, and they raise fair worries that academic researchers could be priced out of their own fields. Expect more pressure for formal, checkable proofs and outside review before release.

Sources: The Verge, Cheeky Pint transcript

SenseTime opens a 9B "looped" model that runs its middle layers twice

Background: A normal transformer passes text through each layer once, so a deeper model needs more stored weights and more memory. Looped models reuse the same layers more than once to get extra depth without extra parameters.

What happened: SenseTime released SenseNova-Loop-9B-Preview on Hugging Face under the Apache 2.0 licence. It has 32 physical layers but runs a 24-layer middle block twice, giving 56 layer passes, and uses a two-stream design it calls mHC to keep the repeated passes stable. SenseTime's own model card reports scores such as 40.45 on Terminal-Bench 2.1 and 71.38 on BIRD-Dev, and says mHC adds 4.42 points on GPQA Diamond. No independent results have been published yet.

Why it matters: Memory, not raw compute, is often the limit when running models locally or cheaply, so trading a bit more compute for a smaller memory footprint is attractive. It is a preview, so treat the numbers as the lab's own until others test it.

Sources: Hugging Face model card, FrontWord

DeepSeek publishes a design for a "Super-agent" that runs many tasks at once

Background: Most coding and chat agents work in separate sessions, and DeepSeek's team says one person can only keep track of three to five at a time.

What happened: In an article on Saturday, DeepSeek's Harness team proposed a two-layer design: a Router that talks to the user, keeps shared context and hands out tasks, and Workers that run tasks in parallel and report back. Events such as chat messages, email, CI results and code-review comments act as triggers, and a separate Watchdog checks whether Workers are stuck or drifting off goal. DeepSeek says an internal version already reproduces and fixes bugs and responds to code reviews. It published a research prototype on GitHub but says it is not an official product, and the prototype leaves out the Watchdog and other parts of the article.

Why it matters: It shows one of the leading open-model labs putting serious effort into the system around the model, where task success, recovery from failures and context management increasingly decide which agents people actually use.

Sources: Global Times, DeepSeek prototype on GitHub, IT之家

Also worth a look

• Kimi K3.1 Pro weights still due by 4 November: As of 31 October there was still no K3.1 Pro repository on Moonshot's Hugging Face page, and no licence published. System Design Academy explainer
• GLM 5.5 rumour (unverified): The X account Lumina claims Zhipu's next model is well into post-training. It hasn't been confirmed by Zhipu. Advance Intelligence roundup
• A paid offer to train hacking models (unverified): Jun Song claims on X that a black-hat group approached him with an offer to post-train open-weight models on exploit data. The claim hasn't been independently confirmed. Advance Intelligence roundup