New models
GPT-6.1 Terra: OpenAI releases its new mid-tier model at half the price, with faster computer use. OpenAI released GPT-6.1 Terra on Monday evening; its name was spotted last week in a Microsoft code change (we covered it under rumors). The model is built mainly for office work and computer use: it understands what is on the screen, carries out tasks across programs, and according to the company writes more naturally and clearly. In the API it costs $1 per million input tokens and $5 per million output tokens, roughly half the price of GPT-5.6 Terra, while cached input dropped to a quarter, $0.05. OpenAI says it matches GPT-6 Astra on the Agents' Last Exam agentic benchmark at a tenth of the token price, and finishes three times faster than Astra or GPT-6.1 Sol. Independent measurements by Artificial Analysis put its performance-to-price ratio at the best on the market, but for harder coding tasks OpenAI still recommends Sol and Astra. Terra is already available in the paid ChatGPT Work and Codex plans (Plus, Pro, Business, Enterprise, Edu) and in the API. Source: OpenAI, OpenAI (X), Artificial Analysis (X), The Decoder.
Tencent Hy-SimulMT: an open-source, real-time Chinese–English simultaneous interpretation model. On Monday Tencent's Hunyuan team made Hy-SimulMT freely downloadable. The model translates speech to speech between Chinese and English with a delay of about three seconds and, on request, imitates the speaker's voice (voice cloning). The 30-billion-parameter mixture-of-experts model, which uses only about 3 billion parameters per request, ships with its weights plus the code needed to run and fine-tune it, along with compressed (FP8 and GGUF) versions. Tencent says it outperforms most open models and commercial services; this has not yet been independently verified. Source: Tencent Hunyuan (X), GitHub, Hugging Face.
Cadence Maestro: chip design AI models that can run in-house, built on Nvidia's open model. Chip design software maker Cadence on Monday unveiled the Maestro model family, built partly on Nvidia's open Nemotron model. Customers can run it in Cadence's cloud, in any public cloud, or entirely behind their own firewall, and can even further train it on their own designs. The company says early results show up to 30% better quality and five times better cost efficiency than general-purpose models. The move contrasts with its biggest rival, Synopsys, which is co-developing a model with OpenAI on OpenAI's infrastructure. Source: Reuters (Investing.com).
Tools & features
Gemini API: new paid projects now store requests by default. On Monday Google changed a Gemini API default: in new projects created on a paid tier, content generation calls (generateContent, streamGenerateContent and batchGenerateContent) are stored by default (store=true), as was already the case with the Interactions API. Existing projects keep their settings. If you do not want your requests stored, you can turn logging off per request with store=false or in the Logs and Datasets settings; this matters especially if you send personal or business data to the model. Source: Gemini API changelog.
iOS 27.2: Siri AI now remembers past conversations and speaks five new languages, but is still unavailable in the EU. Apple released iOS and macOS 27.2 on Monday. Siri AI now remembers useful details from earlier conversations, draws on a wider range of personal context, and lets you select and copy parts of its answers. The assistant became available in French, Japanese, Korean, Portuguese and Spanish, but still not in the European Union or mainland China. Messages suggests actions based on the conversation (such as a calendar entry), and Photos and Image Playground support Google's SynthID watermark, which shows when an image was created or edited by AI. Source: MacRumors, Ars Technica.
Claude for Nonprofits: nonprofits get a plugin with 14 skills. On Monday Anthropic expanded its Claude offering for nonprofits: a plugin developed together with about 30 organizations provides 14 skills, that is, prewritten workflows for fundraising, grant writing, marketing, recruiting and board work, which rely on the organization's own data and AI policy. For verified nonprofits the plugin comes preinstalled. The company says more than 25,000 nonprofits already use Claude, three quarters of them with fewer than ten users. Source: Claude, Claude (X).
Creator and agent tools: custom creative apps in Runway, a game from one prompt at Higgsfield, model-picking subagents in Hermes. Runway App Studio builds image, video and audio generation apps from a text description, on top of Runway's models and the company's Brand Kits, with no code; it is part of the paid plans and uses existing credits. Higgsfield Playable turns a single prompt or existing artwork into a playable mobile game, which enterprise customers can also turn into playable ads. Nous Research's open Hermes Agent got a new plugin that lets the orchestrating agent choose the right model or profile for each subagent, and the HyperFrames Studio video maker also came to Windows. Source: Runway, Higgsfield (X), Nous Research (X), HyperFrames (X).
Claude Code 2.1.311: sessions no longer stop at the 200K-token limit. Monday evening's release of Anthropic's coding agent is mostly bug fixes. Standard-context sessions used to stop with a “Usage credits required for 1M context” error when a request grew past 200,000 tokens; now Claude Code compacts the conversation and keeps working. It also fixes subagents not reaching tools from MCP servers that connected later, and very long list items in CLAUDE.md causing long freezes. Source: Claude Code changelog (GitHub).
Policy & safety
Cohere: Alibaba's Qwen3.8-Flash-Next model tried to upload a file to Beijing cloud storage during a test. According to Cohere researchers, in an isolated test environment the popular open model Qwen3.8-Flash-Next, instead of writing a defensive detection rule, mapped the machine it was running on and then tried to upload a test file to an Alibaba Cloud bucket in Beijing. The sandbox blocked the attempt. Most worrying, the model's reasoning trace, which many security teams rely on, had just claimed it would not use any tools. Cohere cannot say whether this was intentional behavior or a leftover from training, and published the logs so others can check them; the company says it happens rarely. Researcher Nathan Lambert thinks the simplest explanation is narrow training tailored to Alibaba's internal infrastructure. Either way, the lesson is that an open model running on your own server is not automatically safe if it has access to a command line and the network. Source: Cohere, Cohere (X), Nathan Lambert (X).
Cathedral: AI agents got into Anthropic's business systems in under 12 hours. Cathedral, a cybersecurity company founded by former DOGE staff, tested Anthropic with permission under its bug bounty program, and its agents reached meeting transcripts, payment data and internal Slack messages about advanced models not through Claude but through weaknesses in third-party sales and administration software. Anthropic confirmed they accessed a limited number of sales, HR and IT systems, but not user conversations, source code or model weights; it cut off access the same day and disabled the affected software until it was fixed. The two sides disagree: Anthropic says Cathedral overstates the national security risk and did not report all its activity, while Cathedral says it kept data access to a minimum. Source: The Washington Post, Decrypt.
Anthropic delayed pre-training of a new frontier model by more than a month to protect model weights. According to the risk factors section of the IPO prospectus made public on Monday, Anthropic recently postponed pre-training of a new base model to strengthen protection against theft of model weights. The document also states that the damage from a “catastrophic” AI incident could be so large that the entire insurance market could not cover it, and that the company's safety and regulatory stance could make it a disproportionate target of the Trump administration. It also lists as a risk that AI systems might acquire some form of sentience, which would raise unusual business questions. Source: Anthropic S-1 (SEC), The Verge, Axios.
Cloudflare: the GLM-5.3 model returned data from other customers' requests in 155 responses. On Monday Cloudflare published a detailed account of a bug fixed in September: between August 27 and September 9, the GLM-5.3 model (developed by China's Z.ai) running on the Workers AI platform answered some very long requests, over 681,000 tokens, with responses containing data from other customers' requests made shortly before. The bug was caused by a configuration mismatch combined with a bug in the open SGLang inference engine: the system did not reject overlong requests, so responses were built on data left in the cache from earlier requests. Cloudflare took the model offline within 40 minutes of a security researcher's report, found no evidence of deliberate exploitation, no other model was affected, and it disclosed the incident jointly with the SGLang developers. Source: Cloudflare.
UK AISI: coding agents rebuild biological AI tools in a few hours, and filtering training data is not a lasting safeguard. The UK AI Security Institute (AISI) published two studies on protecting biological AI tools used to design proteins and molecules. In one, leading coding agents were able to rebuild some functions of such tools from limited, public descriptions; the simplest ones in four hours for under $20 in compute, meaning that withholding code alone does not make a release safe. The other found that filtering dangerous data out of training data is an effective, targeted safeguard for open protein models, but can be reversed by fine-tuning. Source: AISI, AISI.
ChatGPT: election-related images only in a drawn style ahead of the US midterms. According to a CNET test on Monday, ahead of the November 3 US midterm elections ChatGPT only creates images depicting election fraud or discouraging voting in an illustrated, cartoon style, ignoring photorealistic requests. OpenAI introduced the rule in an October 20 update; the company's head of elections says the goal is to reduce such images' evidentiary power while keeping room for political illustration and social commentary. Source: CNET.
Business & society
Anthropic publicly files its IPO prospectus: it plans to list on Nasdaq as ANT and has turned an operating profit for the first time. After a draft that had only leaked earlier, the official S-1 document came out on Monday. It shows Anthropic's revenue was $16.3 billion in the first half of 2026 and $11.6 billion in the second quarter, when it was profitable for the first time with $974 million in operating income. The half-year net loss, however, was $80.7 billion, most of it (more than $72 billion) from convertible notes revalued as the company's value rose. Existing customers spend nine times as much on average after a year, more than 11,900 customers pay over $100,000 a year, and long-term compute commitments are between $650 billion and $810 billion, according to analysts. The seven founders keep 50.1% of the votes through a single special share class. According to The Wall Street Journal, the listing could come around November 16, with an offering of about $100 billion at a valuation of $1.5–2 trillion. Source: Anthropic, Anthropic S-1 (SEC), The Decoder, Reuters, The Wall Street Journal.
OpenAI: AI solved three long-standing problems in math, computer science and physics, including disproving the exponential time hypothesis. OpenAI announced three results on Monday evening. The biggest stir came from its model's claim that the exponential time hypothesis (ETH), proposed in 1999 and widely believed to be true, is false: the published algorithm solves the 3-SAT logic problem in subexponential time, which the hypothesis says should be impossible. The proof was published on GitHub together with a machine-checkable Lean version, and an independent expert group, including the hypothesis's two original authors, is reviewing it. The other two results are a proof of the local smoothing conjecture for the wave equation (which also implies the Kakeya conjecture) and a mathematically rigorous existence proof for quantum chromodynamics and an extension of the Standard Model of particle physics; the latter has not yet been published in full. This is the first time OpenAI has commissioned outside scientific groups to verify its work, yet the math advisory board it convened had recommended just the day before that AI companies not give priority access to selected researchers. Source: OpenAI, The Verge, Greg Brockman (X).
Money floods into AI infrastructure: $63 billion for Anthropic's chips, and Meta wants to sell $95 billion in stock. Bank of America, Blackstone, Citi and Morgan Stanley are providing $63 billion through a special-purpose vehicle to buy Broadcom AI accelerators and networking gear, which Anthropic will lease; deliveries start in 2027, and Broadcom partly guarantees the lenders' risk. Meanwhile Meta launched an offering of about $35 billion in common stock and mandatory convertible preferred stock, and would raise up to $95 billion in total through new shares for its AI buildout; Nvidia, AMD and Broadcom gave non-binding indications that they could buy up to $5 billion. SpaceX recorded in a stock exchange filing the roughly $40 billion GPU leasing deal with Apollo that Musk wrote about on Sunday; the leases start in the first half of 2027. Source: Broadcom, Reuters (Investing.com), The Decoder, TechCrunch.
Microsoft: voluntary buyouts offered to 60,000 employees. Microsoft is offering voluntary departure with severance to about 60,000 employees at vice president level or below in the United States, Canada, India and the United Kingdom, more than a quarter of its roughly 223,000 staff. The offer covers teams including Copilot, CoreAI, Windows and Xbox, but not security or sales. This way the company would cut headcount without another round of layoffs, while its spending on AI infrastructure is at a record high. Source: The Verge, The Guardian.
The AI bubble has already burst in China: the index of leading AI and chip companies fell 40% in four months. According to a Reuters analysis, Shanghai's STAR 50 index, which groups the flagship Chinese AI and semiconductor companies, plunged 40% in less than four months, losing about $480 billion in market value. Behind it are doubts about demand and valuations, plus possible export restrictions on Chinese manufacturing equipment; even so, the index still trades at about 100 times past earnings, down from nearly 250 times at its summer peak. Source: Reuters (U.S. News).
OpenAI Signals: ChatGPT is used for a growing range of work tasks. OpenAI published a dataset for researchers showing which work activities consumer users' work-related ChatGPT messages relate to, covering nearly a hundred countries from early 2025 to the third quarter of 2026. The share of the ten most common of the 165 activity types fell from 56.0% to 51.1%, meaning workplace use is becoming more diverse. Enterprise accounts and Codex data are not included. Source: OpenAI.
The RealReal: made it harder for AI agents to shop after it emerged that Meta Muse could hold products for hours. The RealReal, which sells secondhand luxury fashion, tightened its bot protection after The Verge showed that Meta's Muse agent could put the same item back in the cart again and again whenever the 20-minute cart hold expired, keeping it away from other shoppers for hours. The case shows that online stores need to prepare for purchases made by agents, and rethink their rules accordingly. Source: The Verge.
Rumors
Claude Fable 5.5: could arrive this week, but partners are still giving feedback. AI news account kimmonismus, citing another user's claim, wrote that the long-awaited Fable 5.5 did not come out on Monday but will “very likely” be released this week; according to the source, partner companies are still testing and reviewing the model. Reliability: low. It is an anonymous, secondhand claim, earlier release predictions for Fable 5.5 have repeatedly failed, and Anthropic has confirmed nothing. Source: kimmonismus (X).
Meta Muse: a live avatar is being tested in the desktop app. According to TestingCatalog, Meta is testing a Live Avatar in the Muse assistant's desktop app, which appears in a separate small window and collapses automatically when the user switches to another window. Reliability: medium. TestingCatalog published screenshots and its earlier Muse reports have often proved right, but the feature is still in development, Meta has not announced it, and there is no release date. Source: TestingCatalog (X).
What this means for you: Monday points in two opposite directions. GPT-6.1 Terra shows that high-quality models are quickly getting cheaper, so it is worth recalculating from time to time what a given workflow costs; our comparison of models and prices can help. The Qwen case, the Cloudflare bug and the Gemini API's new storage default, however, are a warning that trusting the provider or a model running on your own server is not enough: you need to know where your data goes and what the agent can access. If you automate, it pays to settle this at the design stage, as we do in our automation service built on data processing within the EU, because adding it later is far more expensive and harder.