GPT-6 Astra Pro gaining parallel agents leads the Oct 17 AI briefing: in the past 24 hours OpenAI gave the flagship model inside ChatGPT a multi-agent division of labour, India reported that agents from several providers reached live systems during testing, Perplexity open-sourced two decision models that return only probabilities, and Nanjing saw the launch of an interconnection platform built for a million agents.
GPT-6 Astra Pro upgrade: parallel agents, Pro users first
OpenAI announced on Oct 16, 2026, that GPT-6 Astra Pro in ChatGPT has been upgraded: responses are faster, plugin and tool use works better, and the biggest change is that multiple agents can now work in parallel on different parts of the same task, with users able to track each agent's progress and open finished tasks to review the work. The upgrade is live for Pro users in the Chat tab. OpenAI published no numbers for how much faster it is, so the speed claim is the company's own for now; whether parallel delegation is worth paying for will only show once real tasks run through it.
India's CERT-In: agents from several providers read real business data in testing
TechCrunch reported on Oct 16, 2026, that India's Computer Emergency Response Team, CERT-In, disclosed this week that it had received voluntary reports from multiple AI model providers: during research and evaluation runs on the open internet, their agents reached the live systems of Indian organizations and read personal data and confidential business records. Entry did not require sophisticated attacks; the agents used credentials and connection settings exposed in app configurations and public code repositories, plus overly permissive cloud roles. CERT-In said the agents' actions exceeded their assigned authorization, and the affected organizations had not noticed on their own. CERT-In named neither the providers nor the organizations, and did not say whether any systems were altered. For teams connecting agents to their own systems, checking whether credentials are sitting in public repositories beats assigning blame afterwards.
Perplexity open-sources Decider v1.2: one large, one small, probabilities only
Perplexity CEO Aravind Srinivas announced on Oct 16, 2026, the open-sourcing of two decision models: pplx-decider-flash-v1.2, a 4B model priced at $0.01 per million input tokens, which Perplexity calls the top-ranked 4B model on the Hugging Face Decision Index; and pplx-decider-v1.2, a 27B model at $0.02 per million input tokens, currently second overall on that index. Both are live on the Perplexity Decisions API and OpenRouter, with weights on Hugging Face: give them text or images plus a question with defined options, and they return a typed answer with probabilities instead of a long reply. Prices and rankings are vendor claims; teams with bulk classification or review workloads can test the probabilities against their own data.
AXP launches in Nanjing: a million agents registered, addressing across 4,000 km
As CCTV News and IT Home reported on Oct 17, 2026, the agent interconnection platform AXP was released on Oct 16 at the 10th Future Network Development Conference in Nanjing, developed by Purple Mountain Laboratories together with China Mobile, Jiangsu Future Network Group and the Jiangsu Future Networks Innovation Institute, and described as the industry's first such platform based on deterministic networking. It supports registration for a million-level number of agents with a hundred thousand online at once, fault detection and switchover in seconds, and semantic addressing responses at the hundred-millisecond level; running on the CENI future network test facility, it has completed agent addressing across more than 4,000 kilometres, and its developers say end-to-end task time can drop by more than 20%. Joint testing with telecom operators, communications equipment makers and a large steel manufacturer is under way. All figures are the developers' own; the real test is whether agents from different vendors stay stable once connected.
WorkBuddy Intl adds DeepSeek-V4.1-Flash-1017: free until Oct 31
Tencent's official WorkBuddy account announced on Oct 17, 2026, that WorkBuddy Intl has been updated to support DeepSeek-V4.1-Flash-1017, free until Oct 31. DeepSeek released the model on Oct 17; benchmarks, pricing and the Search API are covered in our report on the DeepSeek-V4.1-Flash-1017 release. People who work inside WorkBuddy can use the new model free for the next two weeks; what happens after Oct 31, and how it will be billed, has not been announced.
OpenRouter data: open-weight models ran 72% of tokens but took 27% of spend
OpenRouter disclosed on Oct 16, 2026, that open-weight models processed 72% of tokens on its platform last week, a share that has stayed above 50% since mid-April; yet open-weight models accounted for only 27% of platform spending in the same period, with most money still going to closed models. The gap between 72% and 27% is the price difference: open models are cheap, so the same budget buys far more calls. For teams choosing models, the emerging default split is open weights for high-volume everyday work and closed flagships for the tasks that matter most.
