Anthropic Just Shipped Claude Opus 5 — Half the Price, Double the Score
Anthropic dropped a new flagship model today: Claude Opus 5. The headline isn't just "smarter" — it's smarter and cheaper, which is the combination that actually changes how people use these things day to day.
Here's what shipped, and why it matters if you're building or running AI agents.
The Numbers That Matter
Opus 5 reportedly performs at a level close to Claude Fable 5 — Anthropic's most capable model — but at half the price. On Frontier-Bench v0.1, it more than doubles the score of Opus 4.8, its immediate predecessor. On ARC-AGI 3, a benchmark designed to test genuine reasoning rather than memorized patterns, it scores roughly three times higher than the next-best model.
That ARC-AGI 3 gap is the more interesting number of the two. Benchmark gains on task-specific evals happen every few months across every lab. A 3x lead on a reasoning benchmark specifically built to resist gaming is a different kind of signal — it suggests the model is actually better at figuring things out, not just better at pattern-matching known problem types.
Default Now, Not an Option
Anthropic isn't treating this as a side release. Opus 5 is now the default model for Claude Max and the strongest model available on Claude Pro. That's a meaningful rollout decision — it means Anthropic is confident enough in cost and reliability to put it in front of every paying user immediately, not gate it behind an opt-in toggle.
For a lab that's historically been conservative about default-model swaps, that's worth noting on its own.
Why the Price Cut Matters More Than the Score
Model releases get judged on benchmarks, but the price halving is arguably the bigger deal for anyone actually building with these models. Intelligence gains are nice in the abstract; a model that's both markedly smarter and half the cost changes the math on what's worth running through the top-tier model versus a cheaper one.
That matters most for agentic workloads — anything that calls a model dozens or hundreds of times to complete a multi-step task, rather than a single one-off chat response. Cutting the per-call cost in half while improving reasoning quality means more of those workloads become economical to run on the frontier model instead of a cheaper fallback.
The Pattern Behind the Release
This isn't happening in isolation. The same week, Anthropic also published research with Andon Labs on evaluating how well AI models can autonomously operate physical hardware — in this case, drones performing tracking tasks. Read together, the direction is clear: Anthropic is pushing hard on models that can reason well and act autonomously in the world, not just answer questions in a chat window.
That's the same bet every serious AI lab is making right now — the frontier isn't chat quality anymore, it's how reliably a model can drive a multi-step task to completion without a human checking every intermediate step.
What This Means If You Use OpenClaw
OpenClaw runs on exactly this kind of agentic reasoning — an agent that plans, calls tools, checks its own work, and keeps going without a human approving every step. A cheaper, sharper reasoning model at the frontier is good news for that whole category: better reasoning means fewer dead-end plans and wasted tool calls, and a lower per-call cost means agents can afford to think longer before acting.
The gap between "a chatbot that answers well" and "an agent that gets things done" has always come down to reliability over long task chains. Model releases like this one are what closes that gap, one generation at a time.
If you want to see what a capable reasoning model looks like when it's wired up to actually do work — not just chat — that's what ClawWorld tutorials are built around.