Claude-assisted Riemann proof passes Lean checks; Dragos finds inference leaks
Lean's maintainers say a machine-checked proof of the Riemann Hypothesis holds, and Dragos found AI inference providers leaking prompts and keys.
Read today's AI newsInsights
New models, agents, open-source and local AI, research explained, NZ and Australian AI, and practical build guides. New articles every day.
Microsoft's MAI Code 2 models are in GitHub Copilot at half price until May 2027, and their benchmark averages sit close to GPT-6.1 Sol and Opus 5.5.
Read articleMicrosoft's decision model now reads images and gives steadier probabilities, and an early outside test backs most of its claims.
Read articleGemini 4 Neon costs US$3/US$15 per million tokens and benchmarks near the top, but it's slow, wordy and ignores sampling settings.
Read articleTencent's new open-weight Hy4 Flash uses about 12B of its 220B parameters per token and lists at US$0.07 per million input tokens.
Read articleOpenAI's GPT-6.2 Astra is 20% cheaper than GPT-6 Astra and Sol keeps its price, but long prompts, Ultrafast and weak legal scores need planning.
Read articleAlibaba's closed Qwen4-Max is live on QwenCloud with a 1M-token window and US$1.50/US$4.50 pricing, but most benchmarks are self-reported.
Read articleCognition's Devin Bots turn one goal into parallel coding sessions that run for weeks, and the cost and quality evidence is still thin.
Read articleLangChain's Tool Gateway puts credentials, deny rules and approvals for agent tool calls behind one MCP URL, in beta on Plus and Enterprise plans.
Read articleTwo hijacked releases of Microsoft's agent-framework hit PyPI during NZ business hours on 18 November and stole cloud and CI credentials.
Read articleReplit's Missions runs parallel agents toward an agreed goal under a spending cap, and a separate AI judge decides when the work is finished.
Read articleGitHub's Copilot Evals preview lets teams build repeatable tests for skills, custom agents and MCP tools, though runs aren't sandboxed or cost-capped.
Read articleThe Linux Foundation now hosts ANS, an open DNS-based standard for naming and verifying AI agents, and its reference code can be tried today.
Read articleThinking Machines has released Inkling 2 Mini, an Apache 2.0 looped MoE model that needs about 20GB of memory and runs locally on 32GB Macs.
Read articleCloudflare has published Apache 2.0 weights for Clairic (27B) and Clairic-flash (9B), two models that turn documents into typed, scored fields.
Read articleOVHcloud's first open-weight model is a 6.8B Mamba-attention hybrid for short, tool-using tasks, but for now it only runs on OVH's forked runtimes.
Read articleTencent has released Hy4 flash, a 220B mixture-of-experts model with 12B active parameters, and 4-bit builds already run on 128GB machines.
Read articleEdge0-145B spreads its experts across GPU memory, system RAM and SSD so a 145B coding model runs on one consumer GPU, but its licence restricts hosting.
Read articleHigurashi's Apache 2.0 Haru-4B listens and speaks at once, runs on a 12 GB GPU or Apple silicon, and ships as a base model for research.
Read articleA new preprint shows Qwen3.5-27B gaining 6.6 points on unseen maths by editing its own weights, with no gradients, and only at the largest size tested.
Read articleA new METR note finds the latest AI agents can reliably pick out staged safety tests, so clean eval results may say less than they seem to.
Read articleA new benchmark hides most requirements with simulated colleagues, and the best coding agents fully complete fewer than 8% of its 101 tasks.
Read articleA Berkeley and UIUC paper shows that trimming what an agent harness sends to the model cut costs by 23% on average, with a 3% relative drop in accuracy.
Read articleA NeurIPS 2026 paper shows votes, chosen prompts and extra decoy models can lift a model up to 45 places on arena leaderboards.
Read articleAmazon and Columbia researchers turn logged agent runs into code workflows that finish in about a second, at a small cost in accuracy.
Read articleA Senate committee backed Australia's Digital Duty of Care Bill on 20 November, and the bill covers AI chatbots that NZ firms offer to Australians.
Read articleCanberra has named the 30 CRC-P Round 19 projects, $20m of it AI money, and agentic workflows, edge AI and sovereign tooling dominate.
Read articleHalter's Mav lets NZ dairy farmers question their collar data and approve farm changes in chat, and it's free while in beta.
Read articleOn 18 November the Commerce Commission set out how fair trading and competition law apply to chatbots, pricing engines and AI agents in New Zealand.
Read articleTrellis Data's Agentaus 1 is a 220B-parameter model hosted in Australia with an OpenAI-compatible API, but few of its benchmark results are public yet.
Read articleOne NZ's new Nature Tech Fund will send HARK edge AI audio monitors into Fiordland to search for the piopio, a bird listed as extinct.
Read articleSince 18 November, Gemini's API reads Word, PowerPoint and Excel files and bills only the pages, sheets or cells you pick. Here's how to build with it.
Read articleGemini's tool retrieval preview lets agents load function schemas only when they need them. Here's how to set it up, what breaks and what it costs.
Read articleOpenAI's Responses API now charges 30% or 40% less for background jobs that can wait two or eight hours, including multi-step agent runs that call tools.
Read articleAnthropic's linked billing beta lets your app bill a customer's own Claude organisation for each request. Here's how to set it up and where it falls short.
Read articleAWS's new InvokeWebSearch API gives your own code IAM-signed web search from Sydney at US$7 per 1,000 queries, and here's how to set it up.
Read articleOpenBot 1.0 is an MIT-licensed agent workspace where every Bot action passes a fail-closed policy check. This guide covers local setup, costs and limits.
Read article