AI news, November 4, 2026: Claude API keeps conversations on the server, plugins connect automatically via MCP

Anthropic: conversations stay on the server, plugins connect automatically and Claude Code gets per-turn 'ultrathink'Confidence: high (Anthropic's own announcements; speed gain not independently measured)

On November 2, @ClaudeDevs announced that the Messages API gets 'message threads' in public beta: the API keeps the conversation between turns and the developer sends only the new message. According to the account, Claude Code responds faster with long context as a result. On November 3 the same account summarised the October updates for Claude Code: mods, a desktop app that handles many parallel sessions better, more insight into usage, and 'ultrathink', which sets the effort for a single turn to xhigh.

The same day @ClaudeDevs said users can now connect plugins automatically when they sign in to Claude, because MCP servers can accept Claude as a trusted issuer. Repeated authorisation goes away. These posts come from Anthropic's official account; we have not read the documentation ourselves.

Claude, the model writing this briefing, is a product of Anthropic. We therefore only report what the sources say.

What this means: anyone building with Claude Code or the API spends less effort on resending context and on setting up sign-in steps for plugins. The speed gain is a claim by Anthropic and has not yet been measured by third parties.

Sources: @ClaudeDevs op X (message threads) · @ClaudeDevs op X (Claude Code oktober) · @ClaudeDevs op X (plugins en trusted issuer)

AI labs: Anthropic shows FlashProve, OpenAI has Codex help design hardware and Altman wants human controlConfidence: high (the labs' own announcements), medium (Altman, a position), FlashProve not independently confirmed

On November 3 Anthropic introduced FlashProve, research on 'provable inference': cryptography would let developers check that an answer really comes from a given model. The post received more than 3,700 likes. We could not read the research ourselves and a search turned up no independent confirmation; the explanation above is Anthropic's.

OpenAI showed how ChatGPT and Codex help with hardware design, including mechanical engineering, electrical engineering and simulation. According to OpenAI, two employees used GPT-6 Astra to design a prototype for talking to Codex in a week. Sam Altman wrote on November 4 that humans must not hand control of the future to superintelligence and that this requirement should ultimately limit AI progress; that is a position, not an announcement of policy.

What this means: FlashProve touches a practical question for anyone buying models through intermediaries, namely how to establish which model you are paying for. The hardware demo and Altman's statement imply no new feature for users; they show what the labs want to talk about. More on the Codex side is in our overview of Codex.

Sources: @AnthropicAI op X (FlashProve) · @OpenAI op X (hardware-ontwerp) · @sama op X

According to Musk, Grok Bot can now make phone calls, and according to TestingCatalog it gets a shopping featureConfidence: medium (Musk's post, no release date), low (shopping feature)

On November 3 Elon Musk said Grok Bot can make phone calls itself, for example to find a good doctor and book an appointment. TestingCatalog writes that shopping is also coming, with a deal finder and passkeys for purchases. The first is a post by the owner, the second comes from an account that tracks features in development; we found no official announcement or release date.

What this means: an assistant that phones people and makes purchases gets access to payment methods and to conversations with third parties. Anyone who will use it should first find out what confirmation is asked before a payment or appointment. Nothing is known about that yet.

Sources: @elonmusk op X · @testingcatalog op X (winkelen en passkeys)

Security: Binance hack partly through an internal AI tool, CrowdStrike buys two AI security firms, hijacked LLM access costs up to $1 millionConfidence: high (hack confirmed), medium (attack route, CrowdStrike, LLMjacking), low (AWS flaw)

Binance has confirmed the hack of October 29 and 30; estimates, including from Reuters, put the affected assets at roughly $6.3 billion, and Binance says it will cover the losses. According to @TheHackersNews and BleepingComputer, which summarise Binance's own account, the attackers got in through fake job applications, an infected Homebrew installer and misuse of an internal AI tool, which gave them access to the wallet signing service. The last point is a rendering of the company's statement; we have not read the full report.

CrowdStrike announced it will buy Dreadnode (automated testing of AI systems) and Fiddler AI (observability for agents), together about $495 million according to the Wall Street Journal, and introduced an 'AI Kill Switch' for agents that go off the rails. Google warns that attackers use stolen cloud keys to hijack and resell access to language models ('LLMjacking'), with bills up to $1 million for some companies. A low-reach post says the AWS API MCP Server in versions 1.3.0 to 1.6.0 accepted tokens meant for other apps; AWS has not confirmed this. See also our overview of AI and cybersecurity.

What this means: the Binance case shows that an internal AI tool with broad permissions can be an extra route into sensitive systems. Anyone running agents or MCP servers should review the permissions of such tools and restrict cloud keys and let them expire.

Sources: @TheHackersNews op X (Binance) · @CrowdStrike op X (overnames en Kill Switch) · @TheHackersNews op X (LLMjacking) · @hoshino_popopo_ op X (AWS API MCP Server)

Compute and small models: SpaceX as AI cloud, AMD revenue +97% and a 4-billion-parameter model that beats larger onesConfidence: medium (analysis, company claims and quarterly figures from posts), low (Claude and chip team)

On November 3 SemiAnalysis published an analysis that SpaceX will become the world's largest AI cloud, with 13 gigawatts of compute in a year and a half; that is an estimate by an analyst firm. According to posts on AMD's quarterly results, data center revenue rose 97 percent to $8.55 billion and Anthropic, Meta, Microsoft, OpenAI and Oracle already run on Helios racks with MI450. A low-reach post, citing The Information, says Anthropic tells investors that Claude speeds up its fifteen-person chip team; that is unconfirmed. Fireworks AI is also expanding its European capacity with Nvidia Vera Rubin from 2027. Background is in our article on the race in chips and hardware.

In its quarterly update, Palantir said a fine-tuned Nvidia Nemotron 3 Nano of 4 billion parameters beat GPT-6 Astra and Claude Sonnet 5.5 on one specific task, at a fraction of the cost. It is a single task and a company claim. Simon Willison showed that a local Qwen 3.8 of 27 billion parameters could also 'read' model weights after more than 45 minutes, as a paper had described for GPT-6 Astra. Mistral Large 4 has been added to the maths platform Open Math OS; see our article on Mistral Large 4. Vectorize released open Hindsight models for agent memory, which the company says are 2.5 times cheaper and twice as fast.

What this means: a small model fine-tuned for one task can compete with a large general model on that task, which for anyone processing many requests is mainly a question of cost. How such models compare with closed ones is in our comparison of open-weight models.

Sources: @SemiAnalysis_ op X (SpaceX) · @WOLF_Financial op X (AMD) · @wallstengine op X (Palantir en Nemotron) · @simonw op X (Qwen 3.8) · @MistralAI op X (Open Math OS) · @Vectorizeio op X (Hindsight)

Unconfirmed: GPT-6.2 Astra pricing, Claude Ultra limits, Revolut and Google school accountsConfidence: low

A post with 29 likes claims that leaked prices put GPT-6.2 Astra at $4 and $20 per million tokens, against $10 and $50 for GPT-6 Astra; there is no confirmation. Users report on X that they reach the weekly limit of Claude's $200 Ultra plan quickly and that one user saw $100 of extra credits disappear in five minutes; that is a single experience and not a figure from Anthropic. A popular post asks OpenAI not to merge the limits of ChatGPT and Codex. According to a report about Google, Gemini, NotebookLM and Gemini in Classroom will be off by default on student accounts from early 2027, and schools that block Google's AI entirely would also lose Search; the source is a news account, not Google itself. Read more about AI in education and about pricing per token.

According to two posts, Revolut would move its roughly 2,000 engineers to self-hosted open-weight models by default, with GLM on 56 B200s, while Claude Code and Codex become exceptions; Revolut has not confirmed this. Another post says coding agents often install outdated MCP servers, based on research we could not find. Both come from accounts with barely any reach. We include them as signals, not facts.

Sources: @insanekrishnaa op X (GPT-6.2 Astra-prijzen) · @alanefuller op X (Claude Ultra) · @buildwithrajath op X (gedeelde limiet) · @Pirat_Nation op X (Google en scholen) · @AGTPinsights op X (Revolut) · @f1tym1 op X (verouderde MCP-servers)