Opus 5 tops Fable 5: at the frontier, Anthropic's toughest rival is now its own previous model

Anthropic has released Claude Opus 5. According to the launch, the model surpasses Fable 5 on almost every benchmark, with two named explicitly: Frontier Bench, a coding evaluation, and OS World, a computer-use test in which the model has to drive an actual desktop — click, navigate windows, handle files — rather than just emit text.

The context makes that claim heavier than it looks. Fable 5 is not a rival product: it is Anthropic's own top model, and the highest-scoring system in our Zendoric quality index at 90 points, ahead of GPT-5.6 Sol at 79. So the reference point for this launch isn't OpenAI, and it isn't the fast-closing Chinese open-weight pack (GLM, Qwen, DeepSeek, Kimi). It's the company's previous release. That is what a compressed frontier looks like: leadership is now defended against yourself, in cycles measured in months rather than generations.

Of the two benchmarks cited, OS World is the one we would watch. Coding gains are real but incremental at this point; controlling a computer is the actual bottleneck for agents doing office work, because most administrative labor is not writing code — it's moving between applications, forms and files. Progress there feeds directly into the thesis we have been tracking: back-office and administrative roles are the most exposed, while judgment, relationships and physical presence hold up. If computer use keeps improving at this rate, the exposure gets broader before any of the upside arrives.

A necessary caveat: these are vendor-reported results from a launch note, with no independent verification yet, no published pricing or latency, and no data on the metric that actually discriminates at the top — expert unguided cybersecurity tasks, where saturated benchmarks stop telling you anything. "Beats our previous model on almost every benchmark" is the most standard sentence in this industry. It becomes evidence when someone outside the company reproduces it.

Our reading: this is a genuine but ordinary step on a very steep curve, and the honest way to read it is as compounding rather than as a breakthrough. Short term, faster and more capable agents mean more pressure on exactly the jobs that were already under pressure, and more capability handed to whoever wants to misuse it. Long term, models that can operate real systems — not just describe them — are the precondition for AI that actually does the work of curing disease, extending healthy life and generating abundance, instead of merely advising on it. We will update our index when the numbers can be checked, not when they are announced.