Anthropic released Claude Opus 5 on July 24, 2026, positioning it as a model meant for everyday use that comes close to top-tier frontier intelligence at a fraction of the cost. Pricing stayed at the Opus tier — reported around $5 input / $25 output per million tokens — roughly half what the flagship Fable tier costs.
The scores
The benchmark story is strong. Opus 5 reportedly earned a perfect 42/42 on IMO 2026, with every generated proof machine-verified — echoing the broader trend of models producing checkable math. On coding it posted about 96% on SWE-bench Verified and ~79% on the harder SWE-bench Pro, with a 1M-token context window across the API, apps, and Claude Code.
The interesting move isn't a new top score — it's frontier-class capability at half the top-tier price. The frontier is getting cheaper, fast.
Why "everyday frontier" matters
For most of the last two years, near-best capability meant paying top-tier prices, so teams rationed it. A model that lands close to the frontier at half the cost changes the calculus: you can route far more traffic to a very strong model without the bill exploding. That's arguably more impactful for real products than a marginal capability record.
Confirmed vs. reported
The release and pricing tier are confirmed; exact benchmark figures come from Anthropic's announcement and early coverage, and independent replication (especially on SWE-bench Pro and IMO verification) is how they'll firm up. Benchmarks also never fully capture real-world reliability.
The takeaway
Opus 5 is a clean example of 2026's defining dynamic: the race shifted from "who has the single best model" to "who delivers frontier-grade capability affordably and reliably." Cheaper frontier intelligence is the story — and it's good news for anyone building on these APIs.
