Podcast: Claude Leads GDPx With 12% in a Tough New AI Benchmark

Monday, 02 November 2026 at 19:14

Claude Opus 5.5 leads the launch results for GDPx, a new benchmark of demanding professional tasks, with a score of 12%. The low number is striking. It is also easy to misinterpret: it does not mean AI can complete only 12% of ordinary work.

Research firm SemiAnalysis published the benchmark on November 2. Its launch leaderboard brings together 206 tasks across 11 specialist fields. The November 2 AI World Today briefing explains what the result measures, alongside Claude’s game-remake restrictions, Apple’s China rollout and the HackerTube project.

Listen to the November 2 episode

Listen and follow AI World Today on Spotify or Apple Podcasts. Browse earlier editions on the podcast page.

The difficult part can be finding the right information

GDPx puts agents into working environments, rather than limiting them to a question and an attached reference. Its domain descriptions include finance tasks in a reconstructed transaction with thousands of files, alongside engineering and research assignments.

Consider a simpler hypothetical task: calculating the annual cost of a supplier. Multiplying a known monthly price by twelve is straightforward. Establishing the correct price from competing contracts, expired discounts and revised order requirements is a different job.

An assistant could perform the arithmetic perfectly and still deliver the wrong answer. The relevant question is whether it used the right evidence and produced something a colleague could check.

That is the distinction worth taking from a demanding work benchmark. A fluent response and a usable result are different things.

Why 12% is not a job-replacement statistic

The published methodology gives each domain equal weight and combines results from three attempts, using task-specific scoring that can include partial credit. Models operate in agent software, with tools and time limits that also affect the outcome.

SemiAnalysis says safety restrictions influenced some results. A refusal can lower a score without establishing that a model lacks the knowledge to attempt the task.

The benchmark therefore provides evidence about performance under specified conditions. It does not establish how much of an employee’s working day can be automated, or which product will save a particular business the most money.

A useful evaluation would compare systems on representative work, define acceptable results in advance and record correction time alongside model costs. A cheaper first draft is not necessarily a cheaper finished job.

Claude’s game restriction is narrower than a ban on making games

In an October 30 clarification, Anthropic’s Addy Osmani said Claude Code on Opus 5.5 had received instructions discouraging recreations of someone else’s specific game. Original games, genre projects and work using the user’s own designs and assets are intended to remain supported.

The distinction matters for anyone following AI World Today’s coverage of the change. A model’s ability to write code and a service’s willingness to fulfil a particular request are separate issues. The clarification does not establish a blanket ban on all modding or reverse engineering.

Apple Intelligence’s China launch begins with selected features

Apple’s October 31 announcement sets developer testing for November 4 and a public release with iOS 27.3 in December. The initial features include writing tools, live translation and photo cleanup, with Alibaba’s Qwen supporting selected experiences.

The newer Siri AI is not part of this phase. The announcement is a reminder to check the particular features and dates behind a product name: the same AI brand can describe different experiences in different markets.

HackerTube offers a smaller, more personal measure of usefulness

Over the weekend, Andrej Karpathy described building HackerTube with Claude, bringing together more than 6,000 YouTube videos recommended in Hacker News discussions. He puts the build at roughly two hours, a self-reported figure rather than an independently timed result.

The site has a narrow purpose: a different way to find something worth watching. Its recommendations reflect one community, not universal taste.

That does not contradict a low score on a difficult professional benchmark. It illustrates why the assignment matters. An AI system can be useful on a particular project without being dependable across every demanding workflow.