5 topics covered

Listen to today's briefing
0:00--:--

Australia Central Bank: AI Hitting Young Workers

What happened: Australia's central bank reports early signs that AI is hitting entry-level young workers in exposed jobs such as software development and customer service.

Key details:

  • Since ChatGPT launched, job postings and employment among 15- to 24-year-olds in those fields have dropped more sharply than among older workers, per The Decoder.
  • Author Mike Major cautions against jumping to conclusions, noting business cycle shifts and remote work trends could also explain the numbers.

Why it matters: It adds official-institution evidence to the debate on AI's early labor-market effects, while the authors flag alternative explanations.

Practical takeaway: Treat the findings as early signals, and look for follow-up data that separates AI effects from cyclical and remote-work factors.

OpenAI Dissolves Standalone Robotics Team

What happened: OpenAI is dissolving its standalone robotics team for the second time and redistributing the researchers across other teams.

Key details:

  • The Decoder reports the company frames the move as a way to accelerate its robotics efforts.
  • According to the article, general-purpose models like GPT-6 Astra can apparently control robots directly, which makes a dedicated unit unnecessary.
  • The article states the move is not because OpenAI lacks data.

Why it matters: It signals OpenAI's bet that general-purpose models, rather than specialized robotics groups, will handle robot control.

Practical takeaway: Watch for robotics capabilities appearing in general OpenAI model releases rather than in a separate robotics product line.

Meituan LongCat-2.5 Agent Model Trained on Chinese Hardware

What happened: Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents.

Key details:

  • Described as trained entirely on Chinese hardware.
  • Scores 88.4 points on Terminal Bench, just behind Claude Opus 5, according to The Decoder.
  • The sample deployment runs on 194 Huawei Ascend chips, and Meituan makes no mention of Nvidia GPUs.

Why it matters: The release is a large open-weight agent model whose published deployment path relies on domestic Chinese accelerators rather than Nvidia hardware.

Practical takeaway: If you track open-weight agent benchmarks, check LongCat-2.5's Terminal Bench results against Claude Opus 5 when evaluating options.

Gemini 4 Carbon Tops MathArena Benchmark

What happened: Google's Gemini 4 Carbon and Anthropic's Fable 5.2 have displaced OpenAI's GPT-6.1 Sol at the top of the MathArena AI math benchmark.

Key details:

  • Both Gemini 4 Carbon and Fable 5.2 cost significantly more to run than GPT-6.1 Sol.
  • The Decoder headline puts Gemini 4 Carbon at roughly 15 times the cost of GPT-6.1 Sol.
  • In testing, Gemini ignored the time limit and almost never gave up on a task.

Why it matters: The leaderboard shift shows a top math-reasoning position changing hands, but it comes with a substantially higher cost per result.

Practical takeaway: Weigh MathArena accuracy against per-task cost and time-limit behavior before choosing a model for math-heavy workloads.

Claude Opus 5.5 Default and Prompting Guide

What happened: Anthropic made Claude Opus 5.5 the default model on all paid plans and published a prompting guide.

Key details:

  • The guide advises dropping the phrase "think hard" and starting at Medium effort, per The Decoder.
  • The guide's stated premise is that outdated prompting habits are slowing down Opus 5.5.

Why it matters: Users and developers who rely on older prompt patterns may get weaker results from the new default model unless they adjust how they set effort levels.

Practical takeaway: If you use Claude on a paid plan, try Medium effort as the starting setting and remove "think hard" from prompts.