The AI Wire ●

Tencent released the full version of Hy4 on November 4, adding vision and computer-use capabilities and making the model available through Tencent Cloud and OpenRouter and inside CodeBuddy, WorkBuddy, Yuanbao and ima, Unite.AI reported. The August preview was open-sourced at 770 billion total parameters, 49 billion active and a 1M-token context window; the full release keeps the listed API prices unchanged at $0.834 per million input tokens and $2.501 per million output, with OpenRouter listing a 1,048,576-token context and image input. The weights and a comprehensive system card are planned for a later release, and a cheaper, faster Hy4 Flash is planned after that.

The release's most notable choice is what Tencent left out. Reuters reported that Hy4 was trained with limited cybersecurity data — defensive content only — to reduce its ability to aid cyberattacks: the company filtered out exploit development, reverse engineering and penetration-testing material, keeping defensive proof-of-concept generation and vulnerability patching, and says red-team audits by its Xuanwu and Keen security labs found no elevated risk on tasks like defeating binary obfuscation. Moonshot AI took a similar step with its newest model last month, as Chinese labs follow the US labs in curbing offensive capability.

Tencent reports gains over the preview in instruction following, long-context understanding and long-horizon tasks, and says hallucinations in internal search evaluations fell from 15.76% to 5.45%. It also concedes a gap remains to top proprietary models on research-level mathematics, long-horizon coding and advanced visual question answering. The benchmark claims are company-run and internal; the hosted model, however, is live today, while open weights remain a promise.

OpenAI launched ChatGPT Images 2.6 on November 4, adding native animated GIF creation to ChatGPT and Codex. Users can turn a written prompt or an uploaded photo into a looping GIF, then change its speed, revise individual frames and share the result — a one-step route ChatGPT did not have before.

The release is available globally across all subscription tiers on web, desktop and mobile, and OpenAI says image volume now exceeds 4 billion images a week across ChatGPT, Codex and the API, up from more than 3 billion at the Images 2.5 launch on September 8. Images 2.6 also adds reusable reference images, aimed at keeping a subject consistent across successive edits.

For developers, OpenAI added two API models: GPT-Image-2.6 Flare, positioned for faster generation — RuntimeWire reports output in as little as six seconds and generation about 50% faster — and GPT-Image-2.6 Sunburst, positioned for higher-quality output and consistency across edits. The split mirrors the 2.5 lineup, separating fast draft work from premium production work.

Google launched direct checkout in Canada on November 4, letting shoppers buy products inside the Gemini app and AI Mode in Search without leaving for a separate store website. The first retailers on board are Lululemon, Shoppers Drug Mart, Simons and select Shopify merchants, following earlier rollouts in the United States and Australia.

Payments run through Google Pay using cards and shipping details already saved in Google Wallet, and retailers remain the seller of record. The feature is built on the Universal Commerce Protocol, the open standard Google launched with industry partners earlier this year to give shopping agents, retailers and payment providers a common language.

At launch the Canadian tool is limited to one item at a time, with multi-item carts and price alerts planned to match the US feature set. Google says 59 per cent of Canadian holiday shoppers plan to use an AI tool to help them shop this season, and that over the coming months it will work with retailers on extras such as applying loyalty rewards at checkout.

OpenAI formally showed ChatGPT Voice working inside Finances on November 3: after connecting accounts, users can talk through their money out loud — which subscriptions they are paying for, which to keep, and what cancelling the rest could save. OpenAI directs users to chatgpt.com/finances to try it, and its Help Center lists Finances for Free, Go, Plus and Pro users in the US on web, iOS and Android.

Finances itself is not new — it launched as a Pro preview in May, reached Plus and Pro through the autumn, and expanded to Free and Go users on October 2, which is also when voice support first became available to those tiers, according to CryptoBriefing's research. Accounts link through Plaid (12,000+ institutions) with credit data through Experian as a separate connection; the access is read-only, so ChatGPT can see balances and transactions but cannot move money, pay bills or trade.

The November 3 announcement is the showcase, not the first availability — but it moves voice from a quietly shipped capability to the way OpenAI is now pitching the feature. What OpenAI has not published is any date for Finances outside the US.

Google announced on November 3 that Dreambeans, its experimental Google Labs app that creates personalised daily story collections, is now available to everyone in India aged 18 and over. India is the app's first market outside the United States, and it is available on Android and iOS in 14 languages: English plus Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Gujarati, Kannada, Malayalam, Odia, Punjabi, Assamese and Nepali.

Dreambeans brews its stories from the Google apps a user chooses to connect, including Calendar, Gemini, Gmail, Photos, Search and YouTube. The resulting daily collection can cover places to explore, arts and crafts, or reminders of upcoming local events, and users can dive deeper into a story, bookmark it for later, or share it with others.

The expansion is also a test of appetite. Dreambeans first appeared in the US in June, and Sensor Tower estimates cited by TechCrunch put US downloads at nearly 55,000 since then, with some stories now able to include short animations made with Google's Veo video model. Those same estimates suggest downloads fell sharply in October, so India — where Google says interest was strong — will show whether the daily-story format travels.

Since November 2, API requests naming xAI's grok-imagine-image-quality model no longer reach that model: they keep the same request and response shape but are served by grok-imagine-image-2.0 with quality set to low, and the older grok-imagine-image-pro name follows the same chain. Nothing returns an error, and the bill drops by a cent per image at every resolution — the replacement bills $0.04 against the old $0.05 at the standard size, AI Pricing Guru's daily-verified pricing tables show.

xAI did give notice: its migration guide went up with a 60-day window starting September 2 and recommends switching to the new model by name so customers choose the quality they pay for. Even so, this is the third xAI model retirement in five weeks handled by silent redirect rather than an error, one tracker noted, and xAI keeps no dedicated deprecations page and no written general notice policy.

The practical lesson for developers is to log the model that actually answered, not just the one requested — at xAI right now, the two can differ. The redirect lands in the same week xAI switches on a default text watermark for Grok API output, from November 7, which the Wire covered on October 31.

Anthropic released message threads in public beta on the Claude Platform on November 2. Instead of resending the full conversation history with every Messages API request, an agent sends the full conversation once, then only the new messages plus the previous response ID, while Anthropic holds the thread in memory between turns.

The gain is latency, not cost. With message threads on by default for most Claude Code users, Anthropic says requests carrying 500,000 to 1 million tokens see median time to first token fall by about 35%, and one reported long-conversation payload shrank from about 800 KB to 1.8 KB. Claude still reads the full conversation on every turn, so token usage, standard rates and rate limits are unchanged.

The limits are built into the design. Threads are held encrypted in memory, never written to disk, and expire five minutes after the last request, after which the client resends the full conversation to start a new one. The beta is available on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry and Claude Models on Google Cloud, aimed at long-running coding agents, computer-use agents and document- and image-heavy workflows.

Anthropic's Frontier Red Team, working with collaborators from EPFL, MIT and elsewhere, published FlashProve on November 2: an open-source research prototype for provable inference. Using a zkSNARK — a zero-knowledge proof system — a provider publishes a short commitment to one exact set of model weights without revealing them, then produces a proof for a completed prompt-and-response pair that anyone can check against that commitment.

The test case was large. Anthropic evaluated FlashProve on Meta's Llama 4 Maverick, a roughly 400-billion-parameter mixture-of-experts model in a 4-bit conversion, generating proofs at about 110 tokens per second on a single H100 GPU. A batch of 24 requests took about 223 seconds and produced an 18.78 MB proof per request, while verification took about 17 seconds on a 24-thread CPU with no GPU needed.

The cost is the catch. Proving took roughly 70 to 95 times the GPU time of simply running the model on vLLM, contexts are limited, and Anthropic frames the release — code, Lean proofs and documentation under Apache 2.0 — as a progress update, not a product. A proof also attests only that a specific model produced a response; it says nothing about whether that response was accurate, safe or unbiased.

OpenAI published a safety report on October 30 for GPT-Rosalind-6, a new version of its life-sciences model for biology, drug discovery and translational medicine, and added it to the API changelog the same day. Access is tightly gated: GPT-6 Rosalind Research is available to approved research organisations through ChatGPT Work and Codex, API access needs a separate approval, and buying a subscription alone does not provide it. Applicants face a trusted-access review of their proposed research, security controls and governance, and in Enterprise workspaces the model starts with a two-week administrator preview, off by default.

The model uses the same underlying weights as GPT-6 Astra, with no additional training for this release. OpenAI says it refuses fewer legitimate biological questions that could also have harmful uses, while still refusing biological weaponisation. It is classified High capability in the biological and chemical domain and, relying on GPT-6 Astra's assessments, Critical in cybersecurity; none of its four Critical biological capability evaluations reached the relevant threshold.

On OpenAI's own benchmarks, at the highest reasoning budget tested, GPT-Rosalind-6 scored 49.6% on a medicinal-chemistry evaluation against 28.9% for GPT-Rosalind-5.5, 37.7% against 16.7% on genomics and quantitative biology, and 61.3% against 44.0% on an applied life-sciences benchmark built from tasks contributed by working scientists. OpenAI notes one biological-data benchmark used a configuration different from the production system, and the scores do not show the model can diagnose patients or deliver a successful medicine.

Pricing update, November 2: OpenAI's API pricing page, checked on November 1, lists gpt-6-rosalind at $15 per million input tokens, $1.50 cached and $75 output for prompts up to 272K input tokens — rising to $30, $3 and $112.50 above that — against GPT-6 Astra's $10 and $50. That is a 1.5x premium for, on OpenAI's own October 30 system-card addendum, the same Astra weights with no added training data and no additional post-training: what trusted-access customers are paying for is a looser refusal boundary on dual-use biology, paired with stricter cyber refusals. The Astra for Law configuration, by contrast, still carries no premium, at the standard $10 and $50.

Google has moved its Gemini Computer Use tool into general availability, giving developers a production release of the software that lets AI agents interact with browser, mobile and desktop screens. The change appears in the October 30 Gemini API changelog and applies to both the Interactions and GenerateContent APIs. It is a change in release status, not the technology's first introduction — the changelog records earlier preview versions.

Computer Use lets a model analyze screenshots and propose actions such as clicking, scrolling and typing; developers must supply the software that carries out those actions and sends updated screenshots back. Google lists repetitive data entry, website form filling, web application testing and research across websites among the uses. Its documentation also describes configurable safety policies and optional screenshot scanning to detect hidden adversarial instructions, with an action system that can require user confirmation or block an action outright.

General availability is not a clean bill of health: Google warns Computer Use may contain errors and security vulnerabilities, recommends close supervision for important tasks, and advises against using it for critical decisions, sensitive data or actions where serious errors cannot be corrected. It is also a developer tool — not an announcement that Gemini chat users now have an autonomous computer operator.

Anthropic has quietly added a server-side rule to Claude that blocks users from recreating a specific existing game — even if it is renamed or called original. The rule, delivered under a flag called 'tengu_heron_brook,' first appeared in Claude Opus 5.5's changelog on October 27 and was confirmed in reporting on October 30; there is no setting to turn it off.

Original work is still allowed. The rule expressly permits building original games, games in a genre, and games from the user's own designs and assets — the line is drawn at recreating a particular existing title.

The change lands after weeks of users pairing Claude with reverse-engineering tools to decompile and rebuild existing games, and users are now reporting that related new safety classifiers seem to be limiting other work too, from security research to published coding books. Anthropic has not published a detailed explanation of the rule's scope.

Mistral released its Large 4 model to the public on Thursday, October 29, Reuters reported. Unlike leading AI systems such as Google's Gemini, which are mainly accessed through online services, Large 4 can be downloaded and run independently by anyone, on their own systems.

The released version is an upgrade from the one Mistral unveiled three weeks ago, following testing with governments and corporate users that helped improve its performance. Mistral described it as its most capable openly available model for cybersecurity use — designed to help organisations identify and fix vulnerabilities — competing in a field dominated by Chinese AI models.

"I'm not comfortable living in a world in which open weights do not exist," Mistral Vice President of Science Pierre Stock said. Once the model is out, he added, "it's impossible to contain access" — that protects users from being cut off, but it also means anyone can use it.

Meta has launched Waddle World, an AI virtual-pet app now available on iOS and Android in the Philippines, with wider availability unconfirmed. Users hatch and care for a pixelated companion in a 16-bit style, feeding it, playing with it and watching its mood change.

The twist is conversation: users talk or type about their day, and the pet — which remembers what they share — turns those conversations into illustrated journal entries. The store listing describes companions with moods such as happiness, hunger, sleepiness and boredom, learning more about the user's routines over time, and specifies users must be at least 18.

Waddle World is one of several experimental consumer AI apps Meta has shipped in recent months, alongside Style Sketch, Pocket and Vibes. Meta has not confirmed any wider release or said whether the app connects to its broader Muse AI companion work.

Meta has quietly released Style Sketch, a free AI fashion app now available on iOS and Android in Indonesia, Singapore and Thailand. The app was first spotted by social media consultant Matt Navarra on October 28, and Meta did not respond to TechCrunch's request for comment or say whether a wider launch is planned.

Style Sketch builds a virtual closet from photos a user has already posted to Instagram, or from items added manually by photo, video or camera-roll upload. It then suggests outfit combinations matched to the user's style, with virtual try-on suggestions.

The launch is part of a run of small standalone AI apps from Meta in recent months, alongside Vibes for AI video, Pocket for making games, and the AI pet app Waddle World. TechCrunch notes that digital-closet rivals such as Alta and tools in Google Photos already exist, so Meta's edge is the Instagram connection rather than the idea itself.