Anthropic released Claude Opus 5 on July 24. The short version: it is now the default model on Claude Max and the strongest model available on Claude Pro, it costs $5 per million input tokens and $25 per million output tokens, and Anthropic claims it approaches Fable 5's coding performance at half the price. There is also a fast mode that runs about 2.5x quicker at double the base rate.

Two months ago Fable 5 was the expensive, powerful thing you reached for on hard problems. The pattern of the last few releases keeps repeating: what was the frontier becomes the daily driver, and the price drops.

What the benchmarks are claiming

Anthropic reports that Opus 5 more than doubles Opus 4.8's score on their Frontier-Bench evaluation while costing less, scores roughly three times higher than competitors on ARC-AGI 3 — a test of solving genuinely novel problems rather than recalling familiar ones — and outperforms every other model at any given cost on OSWorld, which measures a model actually operating a computer. It also beat Opus 4.8 on every life-sciences evaluation they ran.

Treat vendor benchmarks as directional. What they consistently signal is where a model has been tuned, and the tuning here is clear: novel problem solving and hands-on computer work.

The part that matters for an SMB

Two things. First, if you are on Pro or Max, you got an upgrade without doing anything — the model behind your daily work changed. Anthropic describes Opus 5 as proactive, which in practice tends to mean it takes more initiative on multi-step tasks. That is worth knowing before it does something on your behalf you expected it to ask about.

Second, the computer-use scores are the real story for anyone considering agent workflows. If you tried handing an AI a task that required it to actually click through your software and gave up because it fumbled, that capability is the one that moved most.

The honest caveat

A better default model can change output quality on prompts you had already tuned — usually for the better, occasionally not, especially for anything where you had worked around an old model's quirks. If you have automations running unattended, spot-check their output this week rather than assuming an upgrade is free. And Anthropic notes Opus 5 remains behind Mythos 5 on some security-specific tasks, so "newest" still is not "best at everything."

The takeaway

Take your hardest recurring task — the one you previously decided needed the most expensive model — and rerun it on Opus 5 this week. The odds are good you have been overpaying for a gap that just closed.