Saturday, 31 October 2026

Anthropic has renamed its two top individual Claude plans. Max 5x is now called Max and Max 20x is now called Ultra, with prices unchanged at $100 and $200 a month, according to [1]. Existing Max 20x subscribers move to Ultra automatically, with no price change [2].

The change that matters is the limit. Claude plans have run on a rolling five-hour window plus a weekly cap [1]. Verified Ultra accounts no longer have the five-hour limit; they keep a weekly limit of 10 times the Pro plan's usage, and a separate weekly limit for Fable [2]. Anthropic verifies Ultra accounts automatically, using factors such as account and payment history, and until that happens a temporary limit applies [2]. Verifying a phone number at checkout does not remove it [2]. To check, open Settings > Usage: if there is no Current session bar, the account is verified [2]. Anthropic says the change applies to most existing subscribers starting now, according to [1].

Max keeps the old shape: five times the Pro plan's usage every five hours and every week, at $100 a month [3]. The pricing page lists $200 a month in API credits on Ultra and $100 on Max, and Fable included up to half the weekly limit [4]. The credits work on the Claude API, Claude Managed Agents and the Claude Agent SDK, not on interactive Claude Code, and unused credits expire at the end of each billing cycle [5]. Both plans are sold monthly only, and the listed prices are for web subscriptions [2][3].

OpenAI, Google, Microsoft and Anthropic say they will honor a shared "policy token" in robots.txt that tells their crawlers a site does not want its content used for AI training, according to The Verge [1]. Until now a site has had to name each company's crawler separately [1][2].

The proposal is an Internet-Draft by Kevin Kelley of Anthropic, dated 30 October 2026, written for the IETF's AI Preferences working group [3]. It adds no new directive. A site writes an ordinary robots.txt group headed `User-Agent: __usage_ai-training` and uses Disallow and Allow as usual, so it can, for example, refuse training on everything except a press-kit folder [3]. A wildcard `User-Agent: *` group must not be read as matching the token, so an existing blanket rule does not count as a training preference [3]. The draft is an individual submission and has no formal standing in the IETF standards process yet [3].

Timing differs by company. Anthropic has already adopted the signal, Google and Microsoft plan to support it next year, and OpenAI says it will implement it in the coming months [1][2]. OpenAI says the signal will not replace its existing crawler controls, which stay in force [2]. Microsoft says sites will still appear in Bing results whatever they choose, according to [1]. Meta's communications head Andy Stone welcomed the idea but did not say whether Meta will honor it, according to [1].

The draft describes the token as a way to express a preference for or against AI training on given paths, using the robots.txt rules crawlers already read [3].

Cisco Talos has documented a Windows infostealer that calls itself "VOLTA-AUTODOXER" and hands one job most stealers leave to a person to a language model: deciding which victims are worth the most [1]. After taking browser passwords, cookies, payment cards, crypto wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud-provider keys and Wi-Fi profiles, it uses the stolen sessions to dig further. It searches the victim's Gmail for receipts to confirm a home address and pulls Uber Eats and DoorDash order history [1].

The bundle goes to OpenRouter with a system prompt that casts the model as "a forensic intelligence analyst drafting a target assessment" [1]. The default model is Claude Opus 4.6. If a reply contains refusal phrases such as "I can't", Volta falls back to Kimi K2 and then Gemini 2.5 Flash [1]. The output ranks what the victim has by monetization value and assesses their password habits [1]. The malware still works without the AI step, according to [2].

Talos says Volta looks like a builder product sold to other criminals, with six swappable slots filled in at build time [1]. Wired reports that Talos did not track whether Volta has been used in the wild [2]. The sample Talos analysed was first seen on 2026-10-11 [1].

Talos's advice: block, or at least alert on, Chrome launched with --headless by anything other than known developer tools, since that is how Volta gets past Chrome's App-Bound Encryption; revoke tokens and change passwords after an infection; and treat any wallet whose seed phrase sat in plaintext on disk as fully compromised [1].

Researchers at Parameter showed that a public ChatGPT share link could run attacker JavaScript inside the signed-in session of anyone who opened it [1]. In their test the code listed the victim's recent conversations, read a private chat that had never been shared with the attacker and sent part of it to an outside server [1]. "Essentially, we could perform any action on behalf of the victim's ChatGPT account," they wrote [1]. OpenAI fixed the flaw on October 26 and rated it P1, its most serious category, according to [2].

The chain joined features that were safe on their own. A graph in the shared chat loaded the MathLive math library into the main ChatGPT page, and MathLive accepted keyboard-layout messages from any sender, with raw HTML allowed in them [1]. An HTML Preview block, itself sandboxed on another address, sent that message, and the HTML built an unsandboxed frame that carried the chatgpt.com identity [1]. ChatGPT's content security policy blocked inline scripts but allowed scripts from YouTube, and a YouTube oEmbed callback supplied the code [1]. In the most reliable version the victim clicked Preview once; in another, opening the link was enough [1].

Parameter reported the bug through OpenAI's bug bounty program on October 18. OpenAI reproduced it, raised it from P2 to P1 and paid $5,000 [1]. Parameter published its writeup on October 28, after the fix [2]. It does not report any real-world use, and users have nothing to install or reset, according to [2].

Apple will switch on part of Apple Intelligence in China with iOS 27.3, according to its China website as reported by MacRumors [1]. The first iOS 27.3 beta is due next week, which The Verge puts on November 4th, and the full release is scheduled for December [1][2].

The China launch covers the older features: Writing Tools, Genmoji, Clean Up, Notification Summaries, Priority Notifications, Visual Intelligence and Live Translation with AirPods [1]. Siri AI and Image Playground are not included [1]. The features are limited to the iPhone 15 Pro and newer until Apple receives approval to launch them on the iPad, Mac and other devices in China [1].

Apple also says it has worked with Alibaba to add support for its Qwen chatbot in some experiences, according to The Verge, which compares the arrangement with the ChatGPT integration iOS added in 2024 [2]. Apple Intelligence was approved in China in July, and Reuters reported in August that Apple trained its own model specifically for China, according to [2].

For a business with staff or customers in China, the split is the point: the same iPhone offers a smaller set of AI features there, and a different chatbot partner, than it does elsewhere [1][2]. Device policies, training material and support scripts written for one region will not match the other.

Anthropic has deprecated fast mode for Claude Opus 4.8 and will remove it on November 30, 2026, according to its API release notes [1]. After that date, requests to claude-opus-4-8 with speed: "fast" return an error [1][2]. To keep using fast mode, the company says, migrate to Claude Opus 5.5 [1].

Fast mode runs the same model with a faster inference configuration. The company says it delivers up to 2.5x higher output tokens per second at premium pricing, with no change to intelligence or capabilities [2]. The gain is in output tokens per second, not time to first token [2]. Fast mode is available on Claude Opus 5.5 through the Claude API, Claude Models on Google Cloud and Microsoft Foundry, and in research preview on Claude Opus 5 and Claude Opus 4.8, which need the fast-mode-2026-02-01 beta header [2]. Opus 5 fast mode requires an account manager or the waitlist [2].

On Opus 5.5, fast mode has its own rate limit, separate from standard Opus limits, with a default in the Start, Build and Scale usage tiers, so no access request is needed [2]. Organizations in the Evaluation tier have to move up a tier first [2]. Going over the limit returns a 429 error with a retry-after header [2]. The usage object in each response has a speed field that shows whether the request actually ran fast or standard [2].

Teams that call Opus 4.8 in fast mode from scripts or agents have a month to change the model ID, check that prompts behave the same on Opus 5.5 and confirm what the switch costs at fast-mode prices.