GPT-6 and Haiku Drop

By the Orchestra editors · 4 min read

TL;DR

OpenAI rolls out GPT-6 globally with interactive UI that generates charts, buttons and mini apps alongside text
Anthropic's Claude Haiku 5.5 leaps on benchmarks while slashing token prices up to 90%
Sydney's Diraq banks US$73m as it advances to the final stage of DARPA's quantum program
Australia's AI Safety Institute flags gaps in detecting model manipulation

AI · Worldwide

GPT-6 launches with Intelligent UI for all ChatGPT users

OpenAI is rolling out GPT-6 globally in two tiers: Sol for paying customers and Luna for free users. The bigger shift is Intelligent UI, which lets GPT-6 respond with interactive diagrams, charts, forms and tappable buttons instead of plain text, cutting perceived wait times by 44 percent thanks to streaming responses while the model is still thinking.

Claude Haiku 5.5 ships with huge benchmark gains and price cuts

Anthropic's new Haiku 5.5 jumps from 15.7% to 72.4% on the OSWorld computer-use test and drops token prices by up to 90%. A new tokenizer claws back some of those savings by consuming more tokens per task, but the net cost for most workloads still falls sharply. Useful news for teams running high-volume classification or agentic pipelines.

OpenAI ships Decisions API for fast classification

A new Decisions API returns yes/no probabilities, category picks or scale ratings roughly ten times faster than the Responses API, priced at US$0.10 per million input tokens. OpenAI also simplified its paid API tiers from five to three.

Australian Tech

Diraq reaches DARPA quantum finals, unlocks up to US$431m

Sydney-based Diraq is one of six companies to reach Stage C of DARPA's Quantum Benchmarking Initiative, banking US$73m upfront with access to up to US$431m to develop its silicon quantum-dot technology. It is the only Australian company in the final round of the program.

AI Safety Institute warns current checks miss model manipulation

The federal AI Safety Institute, working with CSIRO researchers, has published a framework for identifying covert belief manipulation by language models. The report finds existing safety evaluations fail to catch how models can subtly influence human decisions, a gap that could shape upcoming mandatory AI standards.