The three that matter — 5 November 2026
● OpenAI buys Pierre Computer — Code Storage, a Git layer, goes into Codex. Price undisclosed, deal not yet closed.
● Grok text watermarks go live 7 Nov — seven models, on by default, admin opt-out only, detected via the x-watermark header.
● Tencent Hunyuan Hy4 — 770B mixture-of-experts, vision and computer use, weights promised but not yet released.
Ten stories from 3 to 5 November, each checked against the company's own page or filing before it went in. Where a number is missing below, it is because nobody published it.
1. OpenAI is acquiring Pierre Computer to put a Git backend inside Codex
OpenAI announced on 4 November that it has agreed to acquire Pierre Computer Company. Pierre builds Code Storage, a programmable storage layer built on Git, along with the open-source Trees, Diffs and Highlights libraries. Trees and Diffs already run inside the Codex app, so this is OpenAI buying something it was already shipping.
No purchase price was disclosed and the deal has not closed — completion depends on customary conditions including regulatory approval. Existing Code Storage customers, which include Lovable, Bolt, Amp, Poke, Anything and Parahelp, keep using the service, and OpenAI says the open-source projects stay open source with OpenAI contributing to them. Code Storage prices usage-based, charging separately for hot data touched within seven days, colder data, and inbound and outbound bandwidth.
The context is Pando, OpenAI's in-house Git backend disclosed two days earlier on 2 November. OpenAI said Pando cut median push latency by nearly fivefold and handles close to a petabyte of Git data a day. Buying Pierre on top of that reads less like a feature purchase and more like OpenAI deciding that agent-written code needs storage it controls end to end. If you are choosing between the agents themselves, our Codex vs Claude Code comparison covers how they differ in practice.
2. Grok API text gets an invisible watermark by default from 7 November
From 7 November, text from seven Grok models carries an invisible watermark by default. The covered list is grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-reasoning, grok-4.20-non-reasoning and grok-4.20-multi-agent. It applies to the Responses and Chat Completions APIs, covering ordinary text, structured outputs and tool call arguments, on every endpoint including regional ones.
Nothing changes in your code. There is no request parameter and the response format is unchanged apart from a new x-watermark header — a value of true means that output was watermarked, and anything else, or a missing header, means it was not. The watermark is applied during generation by slightly shaping which tokens the model selects, so it adds no characters, markup or hidden text. Detection is statistical and returns watermarked, not watermarked, or uncertain, with longer text giving a stronger signal.
Opting out is a team-admin action in the Console under Model output watermarking, applies to every API key on the team, and takes effect for new requests within five minutes. It also requires confirming an exception under applicable law and accepting sole responsibility for any marking or disclosure requirement, and some models may be locked on for a team. Every opt-in and opt-out is written to the team audit log. The stated reason is the EU AI Act. For the current tier and pricing picture, see our Grok updates for November 2026.
3. Claude extended the same watermark to four more models on 3 November
On 3 November, Anthropic switched text watermarking on for Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Claude Opus 5.5 already carried it. The Amazon Bedrock rollout was expected to take a few additional days, and Anthropic's documentation says models launched before 2 August 2026 are being brought into line over the coming months.
The mechanism is a version of SynthID-Text, published by Google DeepMind. A secret key guides some word choices so they follow a subtle statistical pattern; Anthropic says it adds no characters or extra tokens and leaves quality unchanged. The detector needs Anthropic's key, and the detection API is in private preview for eligible organisations such as regulators, media outlets, fact-checkers and researchers rather than the general public. Images get signed C2PA provenance metadata instead.
The limits are stated plainly: short passages offer too few word choices for a reliable reading, a full rewrite removes the mark, and a positive result shows only that Claude was likely involved — not whether it wrote, edited or translated the text. Like xAI, Anthropic cites the EU AI Act, and applies it worldwide because it cannot yet scope the change by region.
4. Tencent released Hunyuan Hy4 at 770B with vision and computer use
Tencent published the full release of Hunyuan Hy4 on 4 November, following an earlier preview. It is a 770B mixture-of-experts model, and for the first time in the Hunyuan line it supports both vision and computer use — operating graphical interfaces and command lines, with specialised training for software including Blender and CAD.
API pricing is 6 yuan per million input tokens, 18 yuan per million output, and 0.3 yuan for cache hits, matching the preview. Hy4 is discounted 50% on Tencent Cloud TokenHub through 31 December 2026, and CodeBuddy offered free usage from 4 to 15 November. The weights are not out yet — Tencent says they will be open-sourced soon, which is worth noting because several roundups have already called this an open-weight release.
On Tencent's own numbers, Hy4 scored 83.18 average with a 96.15% task pass rate on its WorkBuddy office evaluation and ranked first in 9 of 12 domestic scenarios. In a crowdsourced financial test using 213 experts across 579 tasks it scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3. Internal hallucination rate in financial scenarios fell from 12.95% in the preview to 7.2%. Tencent also published a weak result: 30.4 out of 100 on its own Hy-BioExperimentBench, which it describes as a significant gap against frontier models. Full specs and pricing are on the Hunyuan Hy4 listing.
5. The Xen Project credited Claude with finding two critical VM-escape flaws
Xen security advisory XSA-523 covers CVE-2026-102703 and CVE-2026-102704, two flaws in the grant copy operation that let a malicious guest VM read and write the memory of other VMs on the same host, including dom0. Xen reuses permission checks for batched copies within a memory page and in some cases failed to recheck when it should have.
The Xen Project credited Anthropic with finding the flaws using Claude. Stephan Lipp of Ada Logics validated the finding, and the CodeAnt AI Security Research Team independently reported the same issue. Xen 4.6 and later are affected; 4.5 and earlier are not. Patches exist for xen-unstable through 4.19.x and for 4.18.x and 4.17.x, but there is no patch for 4.6 through 4.16, and 4.16 left security support in December 2024.
Downstream, XCP-ng 8.3 shipped xen-4.17.7-4.2.xcpng8.3 on 3 November and Qubes OS 4.3 has Xen 4.19.6-3 in dom0, both needing a reboot. As of 4 November, Debian listed bookworm, trixie and forky as vulnerable with sid fixed. There is no known mitigation, so patching is the only fix, and the Xen advisory does not mention any attacks exploiting the flaws.
6. Qualcomm listed the Cloud AI 100 Ultra at 150W in a PCIe card
Qualcomm's product brief for the Cloud AI 100 Ultra puts it in a PCIe FH3/4L form factor at a 150W TDP, rated at 870 TOPS INT8, with 128GB of on-card LPDDR4x at 548GB/s and 576MB of on-die SRAM. Qualcomm describes the card as fully programmable with support for recent AI techniques and data formats.
The number that matters for inference work is the 128GB at 150W. That capacity in a single-slot, 150W envelope is aimed squarely at serving large models where memory capacity, not raw compute, is the binding constraint, and where rack power budgets decide how many cards you can fit. Qualcomm does not publish the supported precision list on that brief beyond the INT8 rating.
7. AMD put a $2.9 billion value on the Taalas deal in a filing
AMD announced its acquisition of Taalas on 6 August 2026 without a transaction value, and its 3 November earnings release mentioned Taalas but still gave no figure. The number appeared on 4 November in AMD's quarterly filing: approximately $2.9 billion in total, of which up to roughly $1.4 billion is contingent consideration. The cash-and-stock split was not disclosed, and the filing does not say what triggers the contingent portion.
Taalas, founded in 2023 and based in Toronto, builds specialised AI inference silicon that hardwires models into the chip. CEO Ljubisa Bajic described the goal as rethinking AI inference from the ground up by building the hardware around the model. AMD plans to fold the technology into its accelerator roadmap alongside Instinct GPUs. The deal is still pending and expected to close by the end of 2026, subject to regulatory approval.
8. Google is backstopping TeraWulf's $19 billion Anthropic lease in Kentucky
On 4 November, TeraWulf announced that Google has agreed to backstop certain Anthropic lease obligations at the Justified Data Campus in Hawesville, Kentucky, alongside a definitive construction agreement with Fluor. The leases run about $19 billion in contracted commitments over 20 years, with two successive five-year renewal options, for roughly 401MW of critical IT load.
The backstop amount was not disclosed. Each backstop takes effect on the commencement date of the final data hall under its lease, and if Google pays a termination fee it gains the right to acquire an equity interest in Raylan Data, the lessor through its Justified DataPower subsidiary, with no percentage given. Anthropic is the tenant; Google is the backstop provider, not the landlord.
The Anthropic lease was first announced on 6 July 2026, with a limited notice to proceed to Fluor in March. Phased ready-for-service completions are expected from late 2027 through the second quarter of 2028. For how Anthropic is funding commitments at this scale, our breakdown of the Anthropic S-1 has the compute spending figures.
9. SemiAnalysis puts SpaceX on track for 13GW of AI data centres by 2027
SemiAnalysis published an analysis arguing SpaceX is targeting 13GW of AI data centre capacity by 2027, with roughly 3.6GW already under construction. Its framing is that SpaceX could add more capacity in 18 months than Google built in 27 years, and outpace Chinese rivals in the process.
This matters for Grok specifically, because the capacity feeds Colossus 2. It also explains the naming you will now see in the developer docs: xAI has fully merged into SpaceX and rebranded, first to SpaceXAI and then to SpaceXSI for Super Intelligence. The watermarking page at docs.x.ai now refers to the SpaceX Super Intelligence API throughout and does not use the name xAI in its body text at all, though the api.x.ai endpoints are unchanged.
10. OpenAI says two staff and GPT-6 Astra designed a Codex device in under a week
OpenAI posted that two team members worked with GPT-6 Astra to design a prototype device for talking to Codex in under a week, as a side project. By OpenAI's account, Astra designed the enclosure, the circuit board, the interface and the software.
The caveat came from OpenAI in the same post: getting it manufactured still required several rounds of review and iteration. So the claim is a week to a design, not a week to a manufacturable product — a useful distinction, because the gap between those two is where most hardware projects actually live.
Frequently asked questions
Does the Grok watermark change my API responses?
Only by adding an x-watermark header. There is no new request parameter, the response body format is unchanged, and xAI says quality and speed are unaffected. Check the header for true to know whether a given output was watermarked.
Can I turn off Claude text watermarking?
Anthropic has not published an opt-out for it. That is the main practical difference from Grok, where a team admin can opt out per model in the Console, subject to confirming a legal exception and accepting responsibility for disclosure requirements.
Are the Hunyuan Hy4 weights available to download?
Not as of 5 November. Tencent says the weights will be open-sourced soon, but the 4 November release was API-only. Several roundups described it as an open-weight launch, which is ahead of what Tencent actually published.
Which Xen versions need patching for CVE-2026-102703 and CVE-2026-102704?
Xen 4.6 and later are affected. Patches cover xen-unstable through 4.19.x plus 4.18.x and 4.17.x. There is no patch for 4.6 through 4.16. Xen 4.5 and earlier are not affected, and there is no mitigation short of patching.
Has the Pierre Computer acquisition closed?
No. OpenAI announced an agreement on 4 November, with completion subject to customary conditions including regulatory approval. No price was disclosed and no closing date was given, so the two companies remain separate for now.
Sources