Opus 5 beats Anthropic's own Fable 5 — and the computer-use score is the number that matters

The facts, per the note: Anthropic has released Claude Opus 5, and it surpasses Fable 5 across almost all benchmarks, including Frontier Bench (a hard coding evaluation) and OS World (a test of whether a model can actually drive a desktop computer — clicking, typing, navigating apps to complete real tasks).

For context on the bar being cleared: in our own quality index, Fable 5 sits at 90, ahead of GPT-5.6 Sol at 79 and GLM-5.2 at 77. Beating it is not a marketing win over a weak baseline; it is a lab overtaking the strongest model we currently track, its own.

The more interesting detail is OS World. Coding benchmarks measure whether a model can produce correct text. Computer-use benchmarks measure whether it can operate the messy, unforgiving software humans actually use — where a misplaced click fails the whole task. Gains there translate into economic effect far faster than another point on a coding leaderboard, because they turn a model from an advisor into an operator.

Our read: this is real progress, but read it as a within-lab generational step, not a settled repositioning of the field. Vendor-reported benchmark wins are the beginning of the evaluation, not the end — independent replication on non-saturated tests is what will tell us how much of the gap is durable. We will hold our index until we can measure it ourselves. And the pattern worth watching is the one Anthropic keeps reinforcing: the West's frontier labs still set the ceiling, while China's open-weight models close in from below. Better agentic capability is, in the long run, exactly what turns AI from a chat toy into infrastructure that compresses drug discovery timelines and drains the administrative drudgery out of work. In the short run, it also means the automation pressure on routine digital tasks just went up another notch.