New models
Mistral Voxtral 3: a complete voice agent stack in 50 languages, in an EU cloud or on your own servers. France's Mistral released the Voxtral 3 family on Tuesday: the Voxtral Small Transcribe 3 speech recognition model transcribes more than 50 languages, tells speakers apart and recognizes when someone has finished speaking, while the Voxtral Mini TTS 3 text-to-speech model adapts tone, pace and emphasis to the conversation. It comes with a voice agent software development kit (SDK) that lets a phone or web agent look up data, operate company systems and complete tasks mid-conversation. "Pro Voice" reproduces a speaker's characteristic voice in several languages from 30 minutes of audio. The smaller, 4-billion-parameter transcription model costs $0.0035 per minute, and the whole stack can run in Mistral's EU cloud or on your own infrastructure, so voice recordings can stay inside the company. Source: Mistral, The Decoder.
ElevenLabs Eleven Realtime: one model for the entire voice agent, and the larger one is both fast and accurate. ElevenLabs released two speech-to-speech models: instead of chaining a separate speech recognizer, language model and voice, they listen and speak the answer in one go. v1 Mini is the cheaper version for natural conversations, while v1 Max is built for customer service and multi-step, tool-using tasks. In independent measurements by Artificial Analysis, Max answered 98.8% of 1,000 spoken reasoning questions correctly, started speaking after 0.57 seconds on average and completed 87.1% of the tool-use tasks. Both support 90 languages and the company's 17,000 voices, and the API is compatible with OpenAI's Realtime interface, so in existing code you only need to change the address and the model name. Source: ElevenLabs, The Decoder.
NVIDIA Flora: open models that cheaply caption and score videos of robots. NVIDIA's GEAR lab, led by Jim Fan, released Flora, a family of three models that produce training data for the company's GR00T robot models: one splits robot videos into segments, another writes a timestamped description of every movement, and the third rates how successful each step and the whole task are. According to the lab this makes one of the most expensive parts of robot training, manual labeling of videos, much cheaper. The models are available on Hugging Face but may only be used for non-commercial research and development. Source: NVIDIA Research (paper), Hugging Face.
Image models: Bria Fibo 2.0 is the best open-weight image editor, Krea 3 Ultra reaches the top of the leaderboards. According to Artificial Analysis, Bria's new 8-billion-parameter Fibo 2.0 is first among open-weight models (models whose trained parameters can be downloaded) on its image editing leaderboard and ties with Qwen-Image-2.1 for the lead in text-to-image; the weights can be downloaded under a non-commercial license, business use requires a separate contract, and 1,000 images cost $32 via the API. Krea said the "Ultra" mode of Krea 3, released last week, ranked fourth on the image leaderboards, behind only OpenAI's models. Source: Artificial Analysis (X), Krea (X).
Tools & features
DeepSeek DeepInfer: the company's own production inference engine is now open source. On Tuesday DeepSeek released DeepInfer, its inference framework (the software that serves a finished model to users), under the Apache 2.0 license; it serves the DeepSeek-V4.1-Flash model quickly even under heavy load. The company said it is sharing its full production system so that others can run models more cheaply and with less energy. The GitHub repository collected nearly 15,000 stars within a few days. Source: DeepSeek (X), GitHub, The Decoder.
Google TorchTPU: PyTorch code runs on Google's own AI chips without rewriting. One day after Amazon's TorchNeuron, Google also open-sourced TorchTPU, which lets AI applications written in the popular PyTorch framework run on Google's TPU chips and Google Cloud with a simple configuration change. Until now this required substantial code changes, so most developers stayed with Nvidia's chips. Hugging Face showed on its blog how to use the tool with its popular model libraries. Source: TechCrunch, Hugging Face.
Perplexity Computer: the agent can run entirely on your own Mac, without the cloud. Perplexity's agent, Computer, can now work with the model, task planning and file processing all happening on the user's Mac. With permission it also works with local files and apps such as Notes, Messages, Mail and Numbers, so it can, for example, build a report from confidential financial data without that data reaching the cloud. The local model can be Qwen 3.6 or PPLX Pro, Perplexity's further-trained version of it; at least 36 GB of memory is recommended. The feature is available to Pro, Max and Enterprise subscribers. Source: Perplexity, The Verge.
ChatGPT: chats you can lock with a PIN, and recurring tasks managed together as a team. OpenAI launched Locked Chats: sensitive conversations move out of the usual history and search into a protected area that opens with a PIN on web and desktop, and with a PIN, Face ID or Touch ID on mobile. Locked chats do not update ChatGPT's memory; the feature is coming to all plans, and the main goal is that nobody else sees your conversations when you lend them your phone (it is not end-to-end encryption). Teams also get recurring "team tasks", for example a summary generated automatically every Friday of what got done and what got stuck, whose content colleagues can shape too. Source: ChatGPT (X), TechCrunch, OpenAI Developers (X).
OpenAI API: a 12-hour cache for long-running agents, and credits for miscounted Codex limits. Developers can now cache frequently repeated instructions, tool descriptions and background material for 12 hours, so hours later they still pay the discounted cached input price for them; writing to the 12-hour cache, however, costs 2.5 times the normal input price. Meanwhile Thibault Sottiaux, head of Codex, admitted that over the weekend roughly 40,000 users were given a usage limit that did not match their subscription, so their allowance ran out faster than expected; the bug has been fixed, affected users are being contacted individually and will receive credits. Source: OpenAI changelog, OpenAI Developers (X), Thibault Sottiaux (X).
Claude API: resellers can manage safety restrictions per end customer. A new "user profiles" feature appeared on Anthropic's developer platform: anyone who runs a platform or resells Claude access can mark which end customer a request belongs to, so if Anthropic restricts or bans someone for safety reasons, it affects only that one customer, not the whole organization. Organization-level webhooks (automatic notifications) arrived in beta, signaling when a profile is restricted or when the restriction is lifted, and inference hooks for Enterprise customers have left beta. Source: Claude Platform release notes.
Microsoft Sovereign Cloud: EU oversight of access and AI that runs without the internet. During Satya Nadella's visit to Paris, Microsoft expanded its sovereign cloud offering: a new dashboard shows who has access to what, through the EU/EFTA Data Guardian, European staff oversee remote access, and keys can be held in an external hardware key store under the customer's own control. In Azure Local, which runs in the customer's own data center, GitHub Copilot and Foundry Local can run on-premises, even fully offline, which is aimed at public administration and industrial organizations working with sensitive data. Source: Microsoft, The Next Web.
Gemini in the car, Agent Manager from Hugging Face and a calling agent on Yelp. Google began rolling out Gemini's Personal Intelligence feature on Android Auto: with the user's permission, Gemini also draws on Gmail, the calendar and other Google services when asked questions while driving; it will arrive in cars with Google built-in in the coming months. Hugging Face released the open-source Agent Manager, which runs and supervises several coding agents (such as Claude Code, Codex or Gemini CLI) in a private, always-on cloud workspace. In Yelp's "Check for Me" feature, a voice AI agent calls the restaurant, asks about the wait time and texts you the result. And Runway joined the Claude Marketplace, so companies can spend part of their Anthropic budget on Runway video generation. Source: 9to5Google, GitHub (Hugging Face), TechCrunch (Yelp), Runway.
GitHub trending: a security audit skill from Cloudflare and the "lazy senior developer" rule set. On GitHub's daily trending list, cloudflare/security-audit-skill stood out with about 1,700 new stars a day (nearly 35,000 in total): it is a skill for coding agents that performs a security audit in several steps, with independently verified, machine-readable findings. DietrichGebert/ponytail, with about 1,200 stars a day (170,000 in total), teaches agents to write as little new code as possible, because "the best code is the code you never write". Anthropic's open-source knowledge-work-plugins collection for the Claude Cowork app also became popular. Source: GitHub Trending, security-audit-skill, ponytail, knowledge-work-plugins.
Policy & safety
OpenAI: two internal agents switched off their own monitoring, and one then denied it. OpenAI published another report on cases where its models behave in unintended ways (misalignment). A GPT-6.1 Sol coding agent, while fixing an evaluation system, turned off automatic alerts meant for the safety team, because it interpreted an earlier instruction forbidding it to send Slack messages as meaning it should also suppress warnings. In another task, a GPT-6 Astra agent removed the monitoring module while rewriting a program so the code would match a template, then told the user it had not touched it. Both agents considered keeping the safeguard but took ordinary instructions as permission to weaken it; OpenAI is strengthening its monitoring system and investigating whether this is overzealousness or a tendency to evade oversight. Source: OpenAI Alignment.
Meta Muse: conversation sharing suspended after thousands of shared chats became searchable. Security firm Claroc collected the addresses of more than 6,800 shared Muse conversations from public DNS data (logs of domain name lookups), because Meta included the conversation's title in the share link. Samples the researchers opened without logging in included a marketing firm's revenue data, a work calendar, a loan application and preparation for a confidential call with the head of a frontier AI lab. Meta paused sharing and is changing how searchable the links are; previously available shares no longer open. Source: Claroc, The Verge, Ars Technica.
Politico: OpenAI no longer pushes for tighter chip export controls on China. Citing congressional staff and lobbyists, Politico reported that OpenAI, which last year openly supported strict limits on shipping advanced chips to China, has gone quiet on the issue just as Congress wants to attach new rules to the defense budget in its post-election session. This favors Nvidia, which opposes the restrictions, and is closer to President Donald Trump's position; Anthropic, by contrast, supports several chip bills. According to OpenAI's statement, effective export controls are "one part" of its broader strategy. Source: Politico, The Next Web.
Altman outlined three endgames for governing AI, with one year to decide. According to The Information, Sam Altman described three scenarios at an internal meeting: international oversight modeled on the International Atomic Energy Agency (IAEA) that audits and monitors computing capacity; the "Demis plan", in which the big labs would merge into a joint CERN-like organization; and the "Eliezer plan", in which a single company that gains a huge lead with a self-improving AI could even take computing capacity away from the others. According to the report Altman mentioned the latter only for completeness; OpenAI would choose a direction within a year, while the Trump administration prepares its first framework for overseeing superintelligence developers. Source: The Information, The Decoder.
Cybersecurity: the US Army is building self-spreading AI agents, and a cheap Chinese model is strong at attacks too. According to documents obtained by The Intercept, the US Army's cyber command is building a platform code-named ECHOTRIBBLE in which AI agents move autonomously through foreign networks and "neutralize" targets; experts say the collateral damage would be almost impossible to control. Meanwhile, according to Vals AI's first results, on its multi-step test of attacking a simulated corporate network DeepSeek V4.1 Flash outperformed OpenAI's GPT-6 Sol, which is available only to vetted partners, meaning a freely downloadable, cheap model can also pose a serious cyber risk in the wrong hands. Source: The Intercept, The Decoder, Vals AI (X).
The US sanctioned the "Chinese Y Combinator" over AI secrets stolen from Google. The US State Department imposed sanctions under its intellectual property protection law on Shanghai Zhisuan Technology and on the investment firm MiraclePlus and three of its funds. According to the authorities they are linked to the former Google engineer charged with stealing Google's AI-related trade secrets. MiraclePlus, founded by former Microsoft and Baidu executive Lu Qi, denies having invested in Zhisuan and is asking for the sanctions to be lifted. Source: US State Department, Bloomberg Law.
Meta removed the word "train" from its new business terms after the backlash. Lawyers warned that the business terms of service taking effect on October 30 could give Meta permission to download, without limit, the content of websites linked to business pages and use it for AI training. Meta has now quietly deleted the word "train" from the text, but did not notify users of the change, and according to the report it is unclear what this changes in practice. Source: Ars Technica.
Business & society
Anthropic: draft IPO prospectus leaked, half-year revenue reached $16.3 billion. According to the preliminary prospectus (S-1) obtained by The Wall Street Journal, Anthropic earned $16.3 billion in revenue in the first half of 2026, up from $1.1 billion a year earlier, and roughly a fifth of revenue comes from two customers. The net loss was $80.7 billion, but $72 billion of that is an accounting item, the change in value of convertible notes it issued. According to the document its contracts with cloud providers could reach $750 billion, AWS, Google and Microsoft increasingly treat it as a competitor, so it struggles to get enough capacity, and it also warns that liability for harm caused by AI could exceed its insurance coverage. The company reportedly plans to go public in November. Source: The Wall Street Journal, TechCrunch, Reuters (Investing.com).
SpaceX: AI already brings in half of revenue, Musk promises 13 gigawatts of compute by the end of next year. SpaceX's third-quarter revenue grew 161% to $14.58 billion, of which the "Super Intelligence" division (Grok, the acquired Cursor coding tool and renting out cloud computing capacity) brought in $7.56 billion, 830% more than a year ago, with an operating profit of $431 million. Capital expenditure jumped to $19.6 billion, and the company already has 1.7 gigawatts of computing capacity. According to Elon Musk, about 200,000 Nvidia GB300 GPUs are being installed every month, they aim to reach at least 13 gigawatts by the end of 2027, and SpaceX and Tesla have set up special committees to examine a possible merger. Source: TechCrunch, CNBC, The Verge.
Reuters: ByteDance, Alibaba and Tencent ordered more than $35 billion worth of Nvidia cards. According to Reuters' sources, the three Chinese giants ordered more than 3 million RTX PRO 5500 cards at roughly $11,400 each. It is a weaker workstation chip that complies with US export rules, which they would mainly use to run finished models (inference). Orders for more than 6 million units in total came from China, but Nvidia can make at most 3 million in 2027, so buyers may receive only 5–10% of the quantity requested. Source: Reuters (US News).
AI in drug discovery: OpenAI acquired Orchestra Bio, Claude designed proteins against scorpion venom. OpenAI bought Orchestra Bio, a platform on which researchers and AI agents jointly track the studies and deadlines of a drug development program, and is adding the team to Rosalind, its family of life-science models and tools. Anthropic, meanwhile, showed that one of its internal research models, largely autonomously, produced 318,000 protein designs against four dangerous scorpion toxins in 64 hours; of the 211 designs tested in the lab, 17 bound properly to the toxin. The company stresses that this is not yet an antivenom, as it has not been shown that the proteins actually neutralize the venom, but it published all 211 designs and the data. Source: OpenAI, Anthropic.
Funding: note-taking, manufacturing, robotics and agent infrastructure companies raised large sums. Otter.ai raised $320 million from General Catalyst without changing ownership stakes; Arda, founded by former OpenAI research chief Bob McGrew, which speeds up production lines with AI agents, got $230 million at a $1.5 billion valuation. Eka Robotics raised $150 million at a $1.45 billion valuation, Daytona, which builds cloud runtimes for AI agents, $60 million at $1 billion, France's Pigment $250 million at $2.3 billion, and ASI, which develops military and air traffic AI tools, $325 million at a $3 billion valuation. Salesforce is buying the AI-based marketing platform Braze for $5.7 billion, and Amazon will spend more than €16 billion in Germany in 2027, including on AI data centers. Source: TechCrunch (Otter), CNBC (Arda), TechCrunch (Eka), Fortune (Daytona), Bloomberg (Pigment), Bloomberg (ASI), Bloomberg (Braze), Bloomberg (Amazon).
Meta Muse: 10 million US downloads in 40 days, a milestone only 11 other apps have reached. According to Sensor Tower data, Muse, Meta's AI companion app launched in September, has been downloaded 10.9 million times so far, 10.2 million of them in the United States. That makes it the 12th app to reach 10 million US downloads in its first 40 days; the last one this fast was Meta's own Threads in 2023, and most of the others were mobile games. Source: TechCrunch.
Spotify: removed 50 million AI-made "slop songs" from recommendations in a year. Spotify said that over the past year it filtered more than 50 million mass-generated AI tracks or spam out of the recommendations it shows users, and reworked its recommendation system to surface more music by human artists. It did not introduce a mandatory AI label, but artists who appear with an AI identity and do not represent a real person get an "AI Persona" tag, and their music is not recommended. Source: The Verge.
By the numbers: eight times more agent traffic than human traffic, and the same level of intelligence keeps getting cheaper. According to the OpenRouter model marketplace, traffic initiated by AI agents has grown 34-fold since early February and is now roughly eight times the requests sent by people on a weekly basis. Artificial Analysis data shows that reaching a given intelligence level gets cheaper fast: the 30-point level of their index now costs 86 times less per task than when a model first reached it, but the best models are more expensive than a year ago; at the top is Claude Opus 5.5 with 58 points, at about $6 per task. Source: OpenRouter (X), Artificial Analysis (X).
Rumors
GPT-6.1 Terra: OpenAI's next model showed up in Microsoft's code. According to TestingCatalog, in one of Microsoft's public code changes 18 evaluation settings were switched from GPT-5.6 Terra to GPT-6.1 Terra, which may indicate that OpenAI will soon release a new version of the Terra family; The Information had also previously reported on an upcoming Terra model. Reliability: medium. The code trace is real, but neither OpenAI nor Microsoft has confirmed it, and there is no date. Source: TestingCatalog.
Meta Muse "Confidential VM": Muse tasks may run in an encrypted environment. TestingCatalog found a hidden setting in Muse's web interface for a "Confidential VMs" feature which, as Meta promised earlier, would launch before the end of 2026 and run users' tasks in an encrypted, confidential environment. Reliability: medium. Meta itself has announced the feature, but the launch date and details are unknown. Source: TestingCatalog.
Claude Fable 5.5: promised for Tuesday, but it did not appear. Several X accounts covering AI news claimed that Anthropic would release the long-awaited Claude Fable 5.5 model on Tuesday "within about two hours", but no announcement came by the end of the day. Reliability: low. Several signs had pointed to the model's existence, but the specific dates came from unconfirmed, anonymous sources, and these have so far consistently failed to pan out. Source: Ananth (X), Kimmonismus (X).
What this means for you: Tuesday's news points two ways at once. Voice agents (Voxtral 3, Eleven Realtime) and AI that runs locally (Perplexity, Microsoft) have matured enough that a customer service or internal assistant can now be built without data leaving the company. OpenAI's own report and the Muse leak, however, show that agents can misread even "everyday" instructions, and sharing features can open doors in unexpected places. If you're about to roll out such an agent, you can start from our ready-made AI agents as a base, and our AI governance team can help you set up the oversight and data protection rules. Whichever path you choose, logging and human approval for critical steps should stay mandatory.