Anthropic Ships Opus 5 at Opus 4 Prices. Agents Just Got Cheap.

On 24 July 2026, Anthropic released Claude Opus 5 at the same price as its predecessor, Opus 4.8, while more than doubling coding performance and topping knowledge-work benchmarks. The headline for growth teams is not the intelligence. It is the maths. Frontier-tier reasoning is now cheap enough to run long, multi-step agents at scale, which changes what your marketing operation can automate this quarter.

Claude Opus 5, released 24 July 2026. Credit: Anthropic.

What actually happened

Opus 5 is available today and is the new default on Claude Max. On the Frontier-Bench coding evaluation, it beats every other model and more than doubles Opus 4.8 at a lower cost per task. On GDPval-AA knowledge-work tests and OSWorld 2.0 computer use, it sets the new state of the art while costing a fraction of rival frontier models.

In my opinion the standout stat sits in agentic work. On Zapier AutomationBench, which measures whether a model can complete a business task start to finish, Opus 5's pass rate is around 1.5 times the next-best model for the same cost. Even at its lowest effort setting, it passes more tasks than any competitor.

Why price-to-performance is the real story

Every growth team I speak to has a backlog of jobs that were technically possible with AI but too expensive to justify running continuously. That is the constraint Opus 5 breaks. When frontier reasoning drops to the price of last year's mid-tier, the cost case for always-on agents finally clears.

From my observation, three workflows move from pilot to production overnight:

1. Automated content audits that crawl your site, score pages against intent, and flag decay without a human triggering each run.
2. GEO monitoring that queries AI answer engines daily to track whether your brand is cited, and where the citations drift.
3. Multi-step research workflows that gather sources, verify claims, and draft briefs across a whole content calendar.

These jobs were viable as one-offs. Running them nightly across hundreds of URLs is where the token bill used to kill the idea.

How 'effort' changes your deployment strategy

The mechanism worth learning is the effort setting. Opus 5 lets you dial intelligence up or down per task, so you optimise for reasoning depth or conserve tokens for speed and cost. This turns model selection into a two-axis decision rather than one.

In practice that means you stop asking "which model" and start asking "which model at what effort". Route a bulk metadata sweep to low effort. Reserve max effort for the ambiguous, high-stakes task, like reverse-engineering why a cornerstone page lost visibility. I think this is the shift most teams will underprice: the same model can be your cheap workhorse and your specialist, depending on the dial.

The catch worth naming

Cheaper long-running agents are also longer-running agents, and autonomy has a failure mode. This could quietly erode quality if you deploy without oversight, because the more steps an agent takes, the more its behaviour drifts from the original brief. I have written before on why AI agents drift the longer they run, and Opus 5 makes that risk cheaper to scale, not smaller.

The concrete action: pick one recurring, expensive workflow, run it on Opus 5 at graded effort, and put a human checkpoint at the output stage. If you are still deciding where AI belongs in your operation, start with how AI becomes the default operating layer and build outward from there.