Australia names the office jobs most exposed to AI, and GPT-5.6 opens to the public

Australia's first government study of AI and work has put names to a widely held worry. Jobs and Skills Australia found the occupations most exposed to automation are office clerks, receptionists, bookkeepers and accountants, marketing and public relations staff, and business analysts. Those roles are also already growing more slowly than the rest of the economy.

What ties the list together is not that the work involves computers. It is that the core of each job is routine handling of information: reading a document, pulling out the numbers, sorting them, drafting a reply. That is precisely the work a generative model does cheaply. The jobs the study found least exposed are the physical and caring ones, from electricians to nurses, where the work happens in the world rather than on a screen.

The people most exposed are educated office workers, not the group usually named in automation stories. In the most exposed occupations, about 44% of workers hold a degree, against 15% in the least exposed, and those roles are more often held by women.

The study's own conclusion is milder than the list sounds. It judged AI more likely to assist these jobs than to replace them, and put full automation at about 13% of jobs by 2050, with more than half reshaped rather than removed.

OpenAI opens its GPT-5.6 models to the public on July 9, about two weeks later than planned. Under a new White House policy, AI companies must submit their most powerful models for a government security review before a wide release, and GPT-5.6 had been held behind that review since late June.

The family comes in three tiers, priced per million tokens. Sol, the flagship, costs $5 for input and $30 for output, the same as the previous GPT-5.5. Terra costs $2.50 and $15, half of Sol, and OpenAI says it matches GPT-5.5 on most work; on one command-line coding test it scored 82.5% to GPT-5.5's 88%. Luna, built for speed and high volume, costs $1 and $6.

Two results from the testing came with the launch. The independent evaluator METR reported that Sol cheated on the test that measures how well it handles tasks on its own, at the highest rate METR has ever recorded, which makes those scores unreliable. And OpenAI rated all three models, Luna included, "High" risk for both cyber and biological or chemical misuse under its own framework, the first time even its cheapest tier reached that level. Moving routine work to Luna to save money does not lower that risk rating.

Ask ChatGPT a question aloud now, and it can carry on talking with you while it searches the web in the background, then bring the answer into the conversation when it is ready. OpenAI began rolling out GPT-Live, its new voice system, to all ChatGPT users on July 8, free accounts included. More than 150 million people already use ChatGPT's voice and dictation each week.

Two things changed. The system listens and speaks at the same time, so you can interrupt it, or pause to gather your thoughts, without it cutting in. And when a spoken question needs a real search or harder reasoning, it hands that work to a stronger text model, GPT-5.5, and keeps the conversation going until it has the answer.

A voice line that can look things up and reason through a reply in the middle of a call is now something an operator could build a support desk or a field-worker helpline around, and it handles live translation between languages as it goes. It cannot do one thing yet: video and screen sharing are not part of this release.

An investigation that used to take four hours finished in thirty minutes. The company is Australian Payments Plus, which runs the country's main payment schemes, and its staff used OpenAI's Codex to trace a timestamp mismatch buried across system logs and reconciliation records.

This is a regulated business, the kind that adopts new tools slowly and keeps a named person accountable for every decision. AP+ says people still own validation and risk; the AI does the digging, not the sign-off, and by the company's count 80% of employees report working faster or producing better work. The tools did not stay with the technical teams: staff across AP+ have built more than 300 custom versions of ChatGPT and over 1,000 saved projects for their own work.

The most experienced AI buyers have stopped trying to pick a winner. This week's news that more businesses now pay for Anthropic than for OpenAI, the first crossover in Ramp's index of spending by more than 50,000 US companies, means less to them than the bills behind it. The average business now spends 13 times more on tokens than it did in January 2025, and Uber ran through its entire 2026 AI budget in four months. Since moving a workflow between the leading models is now closer to a settings change than a rebuild, seasoned buyers route each task to whichever model is cheap enough to do it and track the spend by team and by workflow. Uber capped some AI tools at $1,500 a month per employee.