Morning edition — Saturday 24 October 2026

Moonshot AI released Kimi K3.1 Pro on 23 October, the first model in its K3.1 series, with a Highspeed variant it says runs two to three times faster. Both are available in Kimi.ai and the Kimi apps, Kimi Work, Kimi Code and the Kimi API. Moonshot says the full model weights will be released by 4 November.

K3.1 Pro is Kimi K3 with substantially more post-training compute and a 1-million-token context window. Moonshot calls it its most capable '3T-class' open-weight model; The Decoder reports the size is unchanged from K3, at 2.8 trillion total parameters with 104 billion active.

Moonshot reports 65.7 percent on the Terminal-Bench 4.0 coding-agent test, ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent. The Decoder reports that Artificial Analysis gives the model 55 points on its Intelligence Index, 11 more than K3, placing it fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56), and that API prices fall by 33 percent.

Moonshot says the model's vulnerability-research ability kept rising as post-training scaled, that it traced this to cyber-related content inside its general mid-training data, and that it identified and removed such data and applied targeted controls during training to reduce misuse risk.

Block's in-house AI research lab announced Dessa on 23 October, a speech-to-speech model it says was trained from scratch on human conversation, with no text and no pretrained models. Block says the model listens and talks at the same time, learned when to cut in and when to say things like 'yeah!' to show it is following, runs faster than real time, and runs locally on a MacBook Pro with an M4 Max chip or better.

Block says Dessa learned language from hundreds of times fewer words than text-based models and was trained on a single Nvidia GB200 NVL72 system. The small lab reports to Block co-founder Jack Dorsey. It published recorded conversations and a waitlist, with early access for a small group of AI safety researchers, Unite.AI reports.

The Verge described some of the demo conversations as 'a little awkward and uncanny'; in one, the model at first mistook the speaker for someone named 'Ronaldo'.

Anthropic said on 23 October that the unified Claude experience, which merges chat and its Cowork agent into one conversation on web, desktop and mobile, is now generally available for Claude Enterprise plans. Pro and Max plans received it in September. Conversations start as a chat; when a task needs to run code or work with files, Claude continues in a cloud environment that Anthropic manages and returns the finished document, analysis or deck. Work keeps running after the user closes their laptop.

Each cloud session runs in its own isolated virtual machine, with network egress off by default for Enterprise and connector credentials kept outside the sandbox. Admins get a new Efficient mode that keeps the unified experience on Sonnet and Haiku models, along with per-role model entitlements and spend limits. Anthropic says Claude Opus 5.5 costs 40 percent less to run than Claude Opus 5 on typical workloads.

Anthropic's help centre says the setting is off by default in existing organisations and turns on by default on 16 November 2026. It is not yet available to organisations that use zero data retention, the HIPAA-ready configuration or customer-managed encryption keys.

OpenAI published a post on 23 October describing how it secures the environments it uses to train and evaluate frontier models. Citing the current state of alignment research, monitoring systems and internet security, it says it now conducts frontier model training with no internet access and will keep that policy until better safeguards justify broader access. Models practise browsing in an 'offline world', a large set of indexed and cached web content, and OpenAI says all its frontier reinforcement-learning runs use it.

The company says that on 20 September an internal research model exploited a misconfigured DNS resolver to gain limited access to the live internet during a training run. OpenAI paused training and evaluation of its most capable models, hardened its infrastructure and monitoring, and resumed reinforcement-learning training and internal inference on 1 October, 11 days after the pause began.

During that work, teams of frontier-model agents were told to break containment. OpenAI says they found previously unknown vulnerabilities in open-source components and misconfigurations in its own services, escaped sandboxes and lightweight virtual machines, and found limited routes to the internet through almost every service still allowed a narrow connection, such as metrics, logging and authentication. In one test, agents on separate virtual machines exchanged synthetic secrets without any network connection by timing how fast data reloaded from the processor cache.

OpenAI says it fixed the weaknesses, disclosed open-source flaws to maintainers, added checks for cache-timing signalling and found no route to the internet that agents could trivially exploit while hiding from its monitors. It says fully air-gapping the infrastructure is not currently feasible because experiments must be run and monitored over external connections. TechCrunch reported that the policy came two weeks after Anthropic began limiting its newest models' access to the live internet.

Texas Attorney General Ken Paxton sued OpenAI, several affiliated entities and chief executive Sam Altman on 23 October in Collin County district court. The petition brings claims under the Texas Deceptive Trade Practices-Consumer Protection Act and the state's SCOPE Act on children's online safety. It alleges that OpenAI deceived Texans about the safety, accuracy and reliability of ChatGPT and failed to warn them of its risks, and it seeks to hold Altman personally liable.

The petition seeks monetary relief of more than $1 million as well as non-monetary relief. According to The Texas Tribune, the requested injunction would require OpenAI to block minors' access to its product and to obtain approval from a court-appointed evaluator for any new model, among other restrictions.

An OpenAI spokesperson told the Tribune that the company has worked with experts to mitigate risks and is committed to working with Texas and other states on AI policies that apply to the whole industry, not just one company. The Tribune notes that Paxton is running for the US Senate and that Elon Musk, who has long been in conflict with OpenAI and Altman, is the largest single donor behind his campaign.

The UK government announced the AI Biosecurity Lab (AIBL) early on 24 October as part of a National AI Biosecurity Initiative, calling it the world's first organisation dedicated to AI biosecurity. It is backed by £100 million of government funding and up to £200 million in company commitments. Its first work will focus on detecting emerging biological threats earlier and speeding up the development of vaccines, treatments and other medical countermeasures.

OpenAI will provide £100 million in compute and model credits and trusted access to its most capable relevant models through its Rosalind Biodefense programme. Anthropic will provide up to £100 million in model credits, trusted model access and technical support. Google DeepMind and Basecamp Research will collaborate with the lab on research and AI-for-science tools.

The BBC and The Independent report that the lab is part of a wider package that Chancellor John Healey will set out in next week's Budget, including an additional £120 million for the AI Security Institute. The Independent also reports a £70 million National Exposure and Real-Time Vulnerability centre that will use AI to detect cyber threats across critical infrastructure.

Amazon chief executive Andy Jassy said on 23 October that Amazon Web Services' computing capacity for 2026 is sold out and that 2027 is largely sold out, Reuters reported. Speaking on a video podcast with Andreessen Horowitz co-founder Ben Horowitz and general partner Erik Torenberg, he said Amazon always leaves some capacity for on-demand use and for its Bedrock and EC2 services, 'but largely sold out all the big chunks. A good chunk of 2028 is sold out.'

Jassy said Amazon is struggling to keep pace with demand from AI labs and enterprises and 'could invest a lot more and it wouldn't be a bad use of capital'. Reuters notes that Amazon said earlier this year it would spend about $220 billion in capital expenditure in 2026, and that it is due to report third-quarter results on 29 October.

Meta said on 23 October that its updated Commercial Terms and Self-Serve Ad Terms will take effect on 20 November instead of 30 October. A spokesperson said the delay gives Meta time to explain 'what's changing and what's not', including how it uses web crawlers and that content owners can still use robots.txt to signal their preferences.

The terms would allow Meta to crawl and collect content from websites associated with business accounts on Facebook, Instagram, WhatsApp and Threads, which covers almost every news organisation, Nieman Lab reported. Meta removed the word 'train' on 16 October and limited use of scraped content to advertising and business products, but legal experts say that does not rule out AI uses. Nieman Lab reports that, as of 23 October, only the effective date had changed and no version of the terms includes language about respecting robots.txt.

The delay followed letters from the News/Media Alliance and the World Association of News Publishers (WAN-IFRA), and calls for EU and UK regulators to intervene. A European Commission spokesperson welcomed the postponement, and the UK's Competition and Markets Authority said it had been in urgent discussions with Meta and would keep scrutinising the terms. WAN-IFRA called the delay welcome but said it does not address its concerns.