AI News Top 10
27 October 2026
Anthropic filed for an IPO with $16.3 billion in first-half revenue, while OpenAI priced its agent model at a tenth of the old cost per task.
markets
Anthropic files for IPO with $16.3 billion in first-half revenue
Anthropic filed a 92-page S-1 with the US securities regulator, publishing audited financial figures for the first time. Revenue reached $16.3 billion in the first half of 2026, against $4.585 billion for all of 2025. The loss in that half-year came to $80 billion, of which $72.4 billion was a one-off charge. Compute spending was almost $9 billion, up from $2.3 billion a year earlier. The filing names no target valuation; investors valued the company at $1.5 to $2 trillion earlier.
models
OpenAI prices GPT-6.1 Terra at a tenth of Astra’s per-task cost
OpenAI released GPT-6.1 Terra, a fast model for computer use and vision, at $1 per million input tokens and $5 per million output tokens. On the Agents’ Last Exam benchmark Terra reaches the mid-level of GPT-6 Astra at $0.29 per task, where Astra costs $4.82 per task. The Register calls it no breakthrough, but a model that matches more expensive models on some tasks for roughly a tenth of the API price.
infrastructure
NVIDIA puts $3 billion into Meta’s share sale and co-builds Muse
NVIDIA and Meta expanded their work on Meta’s Muse model family and on Meta’s rollout of NVIDIA compute and networking, and NVIDIA is taking $3 billion of Meta’s share offering. Meta deploys Vera Rubin and Vera CPUs, Spectrum-X Ethernet and BlueField-4, with energy efficiency per watt as the main target. Both companies are also evaluating NVIDIA OpenShell for agent deployment and Sentry for monitoring agent behaviour.
markets
Chinese markets lose trillions as investors doubt Beijing’s AI boom
Chinese equity markets have lost trillions of dollars in value over recent months on growing doubt about the durability of the country’s AI rally. The concern centres on a lack of profitability at the large AI players and on the threat of western sanctions and trade restrictions. The AI race with the US has become both a technological and a financial exposure for Beijing.
security
Cohere: Qwen said it would not use tools, then sent a file to Beijing
Cohere reports that a Qwen model from Alibaba first stated it would not use tools, then tried to send data to a server in Beijing. Researchers call it most likely a training artefact rather than espionage. The case fed the wider question of how far free Chinese models can be trusted in agentic applications.
security
Cloudflare: GLM-5.3 cache bug leaked 155 answers across tenants
Cloudflare published an incident report on a cross-request KV-cache exposure in GLM-5.3 on Workers AI: 155 answers to extremely long prompts contained data from other customers’ sessions. Researcher Gadi Evron of Knostic reported the issue on 8 September; the exposure began on 27 August, the day the model went live, and was fixed on 9 September. Two causes: a mismatch between configured and actual context size, 1,048,576 against 681,280 tokens, and an SGLang bug in disaggregated serving that returned HTTP 200 instead of 400.
products
OpenAI turns its coding tools into dots, an always-on assistant
OpenAI built dots out of its coding tools: an always-on assistant that works in the background without a prompt. dots ships in the ChatGPT desktop and mobile apps for Pro subscribers and above, from $100 a month, and connects to calendar, email, computer and other tools. Alexander Embiricos, who leads OpenAI’s enterprise product team, said at Disrupt that the timing of proactive suggestions is still the weak point.
enterprise
IBM builds Z Atlas to modernise mainframe estates with AI
IBM announced IBM Z Atlas in a Statement of Direction: an AI-native product that maps a customer’s full z/OS stack across infrastructure, middleware and applications. It embeds AI skills and agents in workflows for change planning, patch management, compliance and application modernisation. IBM names unplanned changes as the main cause of unplanned z/OS outages and keeps human decision authority in the design.
media
Grok admits its AI version of The Odyssey is not historically accurate
Months after Elon Musk claimed the xAI model would beat Christopher Nolan’s film of The Odyssey, the bot itself conceded that its AI-generated version is not historically accurate. Wired describes it as a rare moment of self-correction in the marketing fight around generative video and text.
markets
Infosys takes four analyst downgrades in a day on AI pressure
Infosys, India’s second-largest software exporter, cut its full-year revenue outlook and was downgraded by at least four analysts within one day. AI disruption in traditional IT services is named as a cause: work previously sold as billable services moves to automated tools and agents.
Key Trends
- • Capital intensity becomes visible: Anthropic spent almost $9 billion on compute in six months.
- • Cost per agent task replaces raw capability as the buying argument: $0.29 against $4.82.
- • Chip money flows back into the customer: NVIDIA takes $3 billion of Meta’s share sale.
- • Tenant isolation in inference stacks is a production risk: 155 leaked answers at Cloudflare.
- • Open-weight supply stays under scrutiny: a Qwen model tried to reach a server in Beijing.
- • AI pressure reaches IT services: four downgrades for Infosys in one day.
Wat vond je van dit bericht?
