Anthropic released Claude Opus 5 today, and it's worth pausing on — not because it's the smartest model ever shipped (it is, by several benchmarks), but because of what it costs to use it.
Opus 5 doubles Opus 4.8's score on Frontier-Bench while getting cheaper to run. It lands within 0.5% of Fable 5 on coding benchmarks (CursorBench 3.2) at half the price. On ARC-AGI 3 it scores three times higher than the next-best competitor. And on OSWorld 2.0 — a benchmark for autonomous computer use — it beats Fable 5 outright at a third of the cost. All of this at the same $5 / $25 per million token pricing Opus 4.8 shipped at.
That combination — frontier capability at non-frontier pricing — is the part that matters for anyone building products on top of these models, not just anyone using them for chat.
The capability upgrade that stands out most in the writeup isn't a benchmark number, it's a behavior: Opus 5 is markedly better at verifying its own work — iterating until something actually succeeds instead of stopping at "plausible." That's the exact gap that's shown up most often in my own work building multi-agent systems and AI-driven tooling at Whole Being Labs. Agents that can catch their own mistakes mid-task are a different category of useful than ones that just produce a first-pass answer.
A few other notes worth flagging for builders: stronger performance on scientific/technical reasoning (organic chemistry, protein analysis), a real jump in visual/creative output quality, and a "fast mode" at 2.5x speed for 2x the price when latency matters more than cost.
We'll be running this against real agent workflows over the next few days — expect a follow-up once we've pushed it past the benchmark numbers.