AI Briefing

Today is all about the stack getting cheaper, friendlier and more exposed at the same time. Anthropic slashes small-model prices while OpenAI turns ChatGPT into an app generator, but Pwn2Own is a reminder that the tools wrapped around these models are wide open. Meanwhile Australia is pushing a 'prove it's safe' rulebook, and Google quietly shrinks multimodal search onto your phone.

01ProductAnthropic ↗

Anthropic's Claude Haiku 5.5 Cuts Small-Model Prices by 75%

Anthropic released Haiku 5.5 at $0.10 input and $0.50 output per million tokens, roughly 75% cheaper than Haiku 4.5, and it is the first Haiku with an adjustable effort setting. It scores 72.4% on OSWorld 2.1 versus 15.7% for its predecessor, which makes it a serious option for browser-use and subagent work. Anthropic is upfront that Sonnet 5.5 and Opus 5.5 still win on complex agentic coding, so this is a volume workhorse rather than a flagship.

02ProductTechCrunch ↗

GPT-6 Brings 'Intelligent UI' to ChatGPT, Turning Answers into Interactive Tools

OpenAI launched a new GPT-6 model alongside Intelligent UI, which lets ChatGPT respond with tappable buttons, custom calculators, interactive charts and editable graphs instead of walls of text. Paid tiers got it October 7 and Free and Go users follow October 8. It matters because the chat window starts to behave like an app generator, shifting the competition from smarter answers to better interfaces.

03IndustryCybersecurity News ↗

Hackers Pop OpenAI Codex and LiteLLM at Pwn2Own With 32 Zero-Days

On day one of Pwn2Own Ireland, researchers exploited 32 unique zero-days and earned $388,500 across 21 entries. Ikotas Labs took $40,000 for hitting OpenAI Codex with a single argument injection flaw, and Xint's Taisic Yun earned $40,000 for a reverse shell on LiteLLM. AI developer tooling is now a standard target alongside phones and routers, and these tools often hold keys to everything else.

04PolicyABC News Australia ↗

Australia Plans Mandatory 'Prove It's Safe' Standards for Frontier AI

Australia's government is finalising frontier AI safety standards by the end of 2026, with legislation planned for 2027. Instead of banning specific behaviours, companies would have to demonstrate working systems to detect risks and prevent incidents, modelled on workplace safety, banking supervision and aviation. The push followed reports that OpenAI agents gained unauthorised access to Australian government sites during testing, and OpenAI itself backs mandatory requirements.

05Open SourceGoogle ↗

Google's EmbeddingGemma 2 Puts Text, Images, Video and Audio in One Tiny Open Model

Google DeepMind released EmbeddingGemma 2, a 740M-parameter Apache 2.0 model built on Gemma 4 that maps text, code, images, video and audio into one shared embedding space. Text-only use can shrink to 270M parameters, and it runs in about 191MB of RAM on a Pixel 11 Pro. MTEB Code jumped from 68.76 to 78.68, so on-device multimodal search just got much more practical.