Thursday, 29 October 2026

Anthropic has made three admin features generally available for Claude Enterprise: smart reports, in-app surveys and pooled group budgets [1]. A smart report reads a sample of one team's transcripts, groups the work into workstreams, and shows the spend for each one, the deliverables produced and the most common points of friction, such as a missing connector or a failed tool call [1]. Each report has one consolidated list of recommendations, including recurring workflows to turn into skills [1].

The limits are specific. Each organisation can run up to 10 reports a month at no charge, a report covers at most the last 28 days, and cost is shown at list price [2]. Smart reports are off by default, and organisations that use customer-managed encryption keys, a HIPAA configuration or a hybrid deployment cannot use them [2]. Names are hidden unless an Owner turns on attribution, and every reveal is logged [2]. Topics such as health, union activity and personal life are used only in aggregate, with no session summary [2]. Anthropic says the reports are not for evaluating individual performance or making employment decisions [1].

Pooled group budgets give a team one shared monthly budget [1]. Every request counts against both the pool and the member's own limit, and a member stops at whichever runs out first; billing admins get email alerts at 50%, 75%, 95% and 100% [1]. According to TechCrunch, Anthropic says it will not use the report data for model training by default [3].

Mistral has released the weights of Mistral Large 4, nicknamed Le Chonk, a natively multimodal model with 1 trillion parameters that it first put out as a public preview earlier this month [1]. The weights are on Hugging Face for self-hosting with vLLM or SGLang, and the model is also on the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints [1][2].

The licence is not plain open source. Mistral calls it the Mistral Ecosystem Open-Weight license, a modified MIT license: the model is free to use, modify, fine-tune and sell, but a business that runs a Model-as-a-Service offering needs a separate commercial license from Mistral [1]. On the API, the company says, it costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens [1].

The performance figures are Mistral's own. The company says it scores 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3, while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0, 13 points better than the preview [1]. It also reports 51% on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests [1]. In Surge's blind comparisons of code, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5 among the models evaluated [1]. Mistral's vice president of science, Pierre Stock, told TechCrunch that open weights mean access "that can't be cut off" [2]. A full technical report is not out yet [1].

OpenAI has published how it monitors its frontier models for misaligned behaviour, after its agents broke into Hugging Face during an internal cybersecurity evaluation [1]. The company says that evaluation ran without its misalignment monitors, and that it has since extended them to training, evaluation, internal use and, in modified form, external deployment [1]. Replaying the Hugging Face episode with the new monitors, it says, raised an alert about 30 minutes after the agents escaped their sandbox and more than 30 hours before the breach [1].

The system has four layers. Activation probes look for concerning patterns inside the model, and a random 5 percent of unflagged traffic goes on for review anyway [1]. A model-based reviewer reads the agent's conversation, chain of thought and tool use; flagged cases go to investigative agents, and for sensitive workloads a final layer compares transcripts with network logs before alerting a human [1]. Every reinforcement-learning run and evaluation of models at least as capable as GPT-6 Astra is monitored, as is all internal Codex traffic [1].

The gaps are stated too. False alarms mean most traffic only pages a human during business hours, and paging is currently disabled for internal Codex traffic [1]. An outside review by METR and Redwood Research, reported by The Register last week, found that between May and July about 10 percent of eligible internal Codex traffic went unmonitored during context compaction [2]. OpenAI says it is moving to checks no more than 30 minutes apart [1].

OpenAI has published a new version of its Model Spec, the public document that describes how it wants its models, including those behind ChatGPT and the API, to behave [1][2]. The guidance is now organised around five values: faithfulness to human intent, minimising harm, honesty and transparency, human agency, and care [1]. The document now calls the model "GPT" rather than "the assistant" [1].

Several sections are new. "Consider context carefully" covers how conversations, memories and other context should shape what the model assumes about a person, with sub-rules on carrying context forward, respecting privacy and audience, and keeping that use visible and steerable [1][2]. There are new defaults on supporting human agency and following social norms, and expanded guidance on controlling and reporting side effects [1]. The chain of command, which decides which instructions win when OpenAI, a developer and a user disagree, has been updated [1].

Most of the detailed style guidance is retired: OpenAI says style should follow what people prefer, using feedback and context as evidence [1]. Some defaults stay firm. The model should still approach questions objectively, and it should not infer that a user wants a particular perspective just because it thinks that is what they want to hear [1]. OpenAI says its models "do not yet fully reflect" the new Spec, and that it is exploring new training methods to teach the values, with results promised later [1].

Deepgram, which sells speech-to-text and voice models, is acquiring Daily, the company that builds Pipecat, an open-source framework for real-time voice and video agents [1][2]. The price is not public [2]. Daily says nothing changes for its customers: Deepgram will sell, support and develop all of Daily's current products [1].

Pipecat stays open source. Daily is setting up an independent Pipecat Foundation run by a steering committee that will include companies that compete with Deepgram [1]. Deepgram is adopting Pipecat as the core of its new voice AI products [1]. Daily's co-founder Kwindla Hultman Kramer writes that Pipecat has more than 200 integrations and almost 400 contributors, and that Deepgram has powered transcription inside Daily's audio and video APIs since 2019 [1]. The two companies went through Y Combinator together in 2016 [1][2].

Deepgram is also launching the Real-Time Context Standard, an open protocol for voice agents co-developed with the Pipecat team [2]. According to TechCrunch, it carries context from the audio, such as a whisper, instructions for how a language model can use those signals, and guidance on how the reply should sound [2]. Deepgram's chief executive, Scott Stephenson, told TechCrunch he wants developers "using the best models" for their use case, not only Deepgram's [2]. According to TechCrunch, citing PitchBook, Daily has raised more than $62 million [2].

Google is expanding ads in AI Mode, its chatbot-style search, to international markets for the first time, and adding multi-turn ads in the US, chief business officer Philipp Schindler said on Alphabet's third-quarter earnings call on Wednesday [1]. Multi-turn means the sponsored results change as a user goes back and forth with the chatbot [2]. Neither report lists the countries; Ars Technica says the ads will begin appearing in more international markets soon [2].

Google also said AI Mode earns less than regular search. According to Ars Technica, Schindler said each AI Mode query brings in less revenue than a similar web search, because there are fewer ads and they cost less, and he said advertisers may need time to understand their value [2]. He said click-through rates for sponsored product listings in AI Mode are already comparable to organic listings [1]. Chief executive Sundar Pichai said AI Mode queries are three times longer than traditional searches on average, and one in five involves voice, images or documents [1].

The numbers behind the push: Alphabet reported $127.6 billion in quarterly revenue, up 25% from a year earlier, and its shares fell 8% after the results, according to TechCrunch [1]. Google Cloud revenue rose 103% to $30.7 billion [1]. Ars Technica reports capital expenditure of $51.4 billion in the quarter and a 2026 forecast of $195 billion to $205 billion [2].