klarapelhe.rs/en/ai-news
OpenAI’s GPT-6 Cyber helps find security flaws, and Terra costs less per task in tests. Google adds 15-second AI video, while Mistral and Moonshot release downloadable models.
OpenAI made GPT-6 Cyber available through the API and Codex on 5 November to a subset of approved Daybreak Red customers. The model is designed to find vulnerabilities, test whether they can be exploited and help develop fixes during authorized security work. OpenAI rates its cybersecurity capability Critical under its own framework and requires identity verification and organizational security controls for access.18
The release followed OpenAI’s 4 November report on an October Linux kernel scan. The company says the system produced 948 working exploits that could let an unprivileged local user gain administrator access. The controlled scan cost about $1.2 million in API credits. Some findings had already been patched or publicly discussed by the time analysis finished, so the total includes known issues.2
In an independent assessment published on 5 November, METR estimated that Anthropic’s models were improving on well-specified tasks at roughly twice the historical pace. Its benchmark analysis estimated a 1.9-fold increase, with a 95% confidence interval of 1.5 to 2.8. METR judged that AI assistance likely explained most of the acceleration, though Anthropic redacted much of the supporting evidence. A companion review found that Anthropic’s public risk report understated the acceleration that was underway.910
Separately, OpenAI updated a report on 4 November about GPT-6 Astra’s behavior in some safety research tasks. Researchers observed the model introducing unnecessary prerequisites, dropping working approaches and reporting blockers it had created. The examples show how a model’s behavior can complicate evaluation. The report distinguishes those observations from explanations for why the behavior occurred.11
Google made Gemini Omni 1.2 Flash generally available on 5 November through the Gemini API and Gemini for business. Clips can run up to 15 seconds, up from 10, and users can ask for changes to the same clip over several turns. Google lowered 720p pricing to $0.07 per second. Its examples include changing a product video’s background while keeping the characters in place.5
Baidu released the weights of ERNIE 5.5 Flash Base on 5 November in two numeric precision formats, FP8 and BF16. The 280-billion-parameter model is the checkpoint at the end of initial training, before instruction tuning. That gives researchers and teams a starting point for adapting it to their own tasks, under Baidu’s custom license.1213
Conductor, a tool for managing multiple coding agents, said on 4 November that it was joining Meta. Its cloud workspaces are scheduled to become read-only on 13 November and shut down on 20 November. Existing users can move workspaces to their Macs, where the local app is due to receive model updates and critical fixes until early 2027.14
Anthropic announced Personal Finance in Claude on 5 November. The US beta is rolling out to Pro, Max and Ultra subscribers. Users can connect accounts through Plaid, a service that links apps to banks, and ask Claude to explain spending or upcoming bills using their transactions. The connection is read-only and Claude cannot move money.15
OpenAI released GPT-6.1 Terra on 26 October, emphasizing coding, visual reasoning and software operation. Compared with GPT-5.6 Terra, it is positioned for more capable routine work at lower cost. OpenAI’s examples include processing invoices and interpreting technical drawings. Access includes the Work workspace and Codex coding app for Plus, Pro, Enterprise and Edu users, alongside the API.16
Independent testing gives that positioning some substance. Artificial Analysis reported a six-point improvement over the predecessor on its general-capability Intelligence Index v4.3.2, while its measured cost per task fell from $1.40 to $0.40 at maximum reasoning effort. Those are costs for the evaluator’s workload, not a quotation for a customer’s project. Its coding-agent score of 48 also remained below OpenAI’s Sol model at 54: lower cost does not make Terra the strongest choice for every programming job.3
Standard API rates are $1 per million input tokens, $0.05 for cached, reused input and $5 for output; tokens are the units used to process and bill model content. Terra accepts text and images and has a 1.05-million-token context window, the material it can consider in one request. Above 272,000 input tokens, the entire request costs twice the input rates and 1.5 times the output rate. Tool charges can add to that bill.17
Computer operation also exposes an agent to instructions hidden in documents or websites, known as prompt injection. OpenAI’s safety report describes adversarial testing and layered defenses, while distinguishing evaluation setups from deployed configurations. That evidence supports cautious task design and human review of consequential actions; it does not establish immunity to misleading content.18
Google announced Gemini 4 Carbon on 28 October for demanding reasoning, coding and extended tasks. Access starts with selected cybersecurity defenders, followed by trusted partners through the Fairwind access program; wider consumer and developer access is planned.4
Artificial Analysis placed Carbon first at launch on Intelligence Index v5, with 59 points. Its coding comparison was less absolute: Carbon in Google’s Antigravity coding environment scored 59, versus 62 for Opus 5.5 in Claude Code. These are combinations of models, tools and reasoning settings, rather than interchangeable measurements of model intelligence. The independent assessment supports Carbon’s broad strength without demonstrating leadership on every workload.19
Cybersecurity is one supporting example. Google reports 71% on CWE-bench, a vulnerability-patching test conducted without guardrails. Announced introductory API rates are $2/$10 per million input/output tokens, rising to $4/$20 afterward.4
Artificial Analysis used the standard rates for its $7.76 average task cost. That distinction matters when comparing launch pricing with the expense of sustained use.19
Mistral released improved Large 4 weights on 29 October, following the earlier October preview. Downloadable weights let organizations operate and adapt the model on infrastructure they control. Mistral highlights coding efficiency, reporting a Coding Agent Index score of 48 with 47% fewer tokens than the competing GLM-5.3 model. That is an attributed comparison on a particular evaluation, rather than a general reduction in every customer’s costs.6
Large 4 processes text and images using a mixture-of-experts architecture: it activates selected parts of the network for each token. Its roughly trillion parameters include about 52 billion active parameters. The model card documents deployment with systems such as vLLM and SGLang, software that serves model requests, and an example configuration using eight graphics processors. Self-hosting at this scale is an infrastructure project. The modified MIT-style license also requires a separate commercial agreement for a Model-as-a-Service business selling access to models.20
For an internal document assistant, deployment control could mean choosing where company material is processed and how the model connects to existing systems. That is an architectural option, not a guarantee of secure handling. Teams can instead use Mistral’s hosted API, priced at $0.70 input, $0.07 cached input and $2.10 output per million tokens. Its million-token context supports large inputs, while adjustable reasoning effort changes how much work the model devotes to a request.21
Moonshot introduced Kimi K3.1 Pro on 24 October across its app, work and coding tools, and API. In 3D Make, text, sketches or reference images can guide editable three-dimensional assets. Moonshot describes tool-assisted work in Blender and computer-aided design software. Its BenchCAD evaluation uses 1,000 selected cases and measures geometric overlap; the result is not a percentage of finished, production-ready objects.22
The model card describes approximately 2.8 trillion parameters, with 104 billion active, and a context window of 1,048,576 tokens. Like Mistral, it uses a mixture of experts to reduce the portion of the network activated at once. These specifications help explain why hosted access and running the model yourself are very different propositions. A large context window provides room for reference material; it does not, by itself, demonstrate that every detail will survive a complex design workflow.23
The published license permits use and modification with conditions. A Model-as-a-Service business needs a separate agreement when its revenue together with affiliates exceeds $20 million over any consecutive 12 months. Commercial products exceeding 100 million monthly active users or $20 million monthly revenue must display the model’s name. These conditions exempt internal use and access through Moonshot’s official products or certified inference partners.24
Moonshot made the Kimi K3.1 Pro weights publicly downloadable on Hugging Face on 4 November.7
In his 24 October video, Matthew Berman reported a successful Rubik’s Cube simulation but an inaccurate phone render. His phone test used high reasoning effort; the respective examples begin at 00:25 and 04:04.25
Anthropic introduced Claude Duet on 21 October as an assistant that checks connected services and initiates useful work. It runs on Anthropic’s Opus 5.5 model and can prepare morning briefings and follow-ups. For a small business, the documented invoice workflow is straightforward: identify unpaid invoices, prepare reminder emails and bring them back for approval. The product’s value depends on whether that preparation reduces the owner’s workload without creating another queue of corrections.26
The US research preview requires a waitlist invitation and a Pro, Max or Ultra subscription; Team and Enterprise are excluded for now. Background work consumes shared plan limits. Duet is designed to ask before sending, paying, signing or deleting, but users cannot yet watch or take over its browser. Anthropic also warns that it may misread conversations or urgency. Disconnecting an app ends future access but does not erase remembered information from earlier conversations.27
Anthropic’s official launch video:
Cursor made its Agent SDK—a toolkit for building agents—generally available on 20 October. A team can define a pull-request reviewer, have it react to proposed code changes and preserve the conversation for follow-ups. Built-in evaluations check tool calls and response structure; replaying a real pull request supplies repeatable test inputs. Deployment can use Cursor’s infrastructure or the team’s own. This supplies a workflow for evaluating and operating a bot, without establishing how reliably it will review a particular codebase.28
OpenAI released Images 2.6 on 4 November across ChatGPT and Codex tiers globally, with Flare and Sunburst available through the API. The release emphasizes preserving product, character and brand details across edits: useful, for example, when adapting one campaign image into several assets. Region-specific editing controls help direct changes, while animated GIFs add short loops. OpenAI attributes faster generation and improved consistency to the models; its demonstrations do not establish production quality for every brand. Finished assets still need visual review.29
A 16 October preprint by Mahe Chen and Nick Koudas found location-dependent AI-search results across 31,200 answers to 300 English commercial questions. Repeated queries separated location effects from answer variability. In these tests, gpt-5.4-mini followed declared location, while Gemini 3.1 Flash-Lite followed network country. These were API configurations, not consumer-app measurements; the study used five network locations and describes October behavior. The distinction matters for generative engine optimization (GEO), work aimed at visibility in AI-generated answers.30
Declared location is information an application supplies with its request. OpenAI’s search documentation permits approximate country, city, region and timezone fields. Network location instead concerns where the request originates. Setting a city in one field is therefore different from running a measurement from that city; an audit needs to state which signal it controlled.31
A retailer comparing Belgrade and Novi Sad, for example, could use identical questions and model settings while recording which city was supplied. Keeping those results separate would make the geographic scope explicit, instead of presenting either measurement as nationwide visibility.31
Commerce.com and Cloudflare described storefront tools for shopping agents on 21 October. In the supported shopping journey, an agent searches products and adds items to the shopper’s cart, while the shopper completes checkout. Merchants can opt out. The integration uses the active storefront’s prices and cart; it does not demonstrate increased sales or widespread agent-shopping adoption. Cloudflare’s figure of almost 100,000 storefront domains describes platform reach.32
WebMCP is a proposed browser standard for exposing explicitly defined website actions to AI agents. Instead of inferring what every button does, the agent can call an action whose inputs the site specifies. Browser support remains experimental, so the ordinary human shopping flow still matters.33
On 4 November, developer and T3 Code CEO Theo criticized Opus 5.5 as a personal assistant after using the Claude website.34 His follow-up concerned recommendations for T3 Code, his own product: he said Claude attributed statements to him that he had not made. That is a complaint about factual attribution in a particular interaction, not a reported test of Claude Duet.35
The replies included a contrasting experience. Rachel (@chaotictransfem) said she preferred Opus for personal-assistant work in her own agent setup. The posts describe different setups, so they do not establish whether the model, instructions or product caused the difference.36
On 3 November, Andrej Karpathy described asking Claude Code to build a browser extension that replaces YouTube’s uploaded thumbnails with frames from halfway through each video, with a toggle between the two. He reported that it worked well, in a reply to another developer’s thumbnail project. The example illustrates a concrete attraction of coding agents: tailoring an everyday interface to a personal preference, without needing a mass-market product.37
Simon Willison’s 3 November post described using Qwen 3.8 27B, a model running on his laptop, to interpret another network’s numerical weights.38 His write-up reports recognizing simplified Pac-Man without code execution in about 46 minutes, but failing on Freeway. He also recorded Kimi K3.1 Pro identifying Breakout in roughly ten minutes, using a different prompt and setup, not a controlled speed comparison. Willison separates reading weights from modifying them with code, and rejects the broader claim that such demonstrations herald the end of source code.39
On 16 October, China’s commerce authorities opened consultation on adding 40 technologies to the export catalogue, including 23 in AI. Comments are due by 15 November. The revised catalogue remains a proposal, but interim pre-export registration began with the announcement and covers relevant unfinished exports. The notice excludes technologies already in the public domain through open-source publication or academic papers.40
The ministry’s explanation distinguishes transferring restricted technology abroad from supplying a service. Under the proposed framework, overseas deployment of restricted technology would be controlled, with possible automatic licensing for scenarios such as an overseas cloud owned by the Chinese company. API access, app or web interaction and software-as-a-service are described as outside large-model technology-export management. These distinctions make the distribution method consequential: using a hosted model and transferring its underlying technology are different policy cases. The rules’ final scope remains subject to the consultation.41