What Anthropic released
On July 24, 2026 Anthropic released Claude Opus 5. In one line: the flagship model gets close to Fable 5 on plenty of tasks, but at half the cost per task, and without raising the list price.
The numbers that matter, right away. Same price as Opus 4.8: $5 per million input tokens, $25 on output. State-of-the-art on Frontier-Bench and on GDPval-AA. On ARC-AGI 3 — the test on problems never seen before — it scores roughly three times the model right behind it. It's already the default on Claude Max and the most powerful model available on Pro.
But the news that really moves the needle is something else, and it's a feature, not a benchmark. It's called the effort dial. Below you'll find what it means in practice, when you still need a bigger model, and how to review your setup.
The effort dial: the real news
Until now, with a flagship model you always paid the same: maximum power on every request, even when the request was trivial. Opus 5 introduces a dial. You decide how much to make the model "think": high effort when the task is complex, low effort when you just need a quick answer.
At reduced effort Opus 5 keeps most of its quality while consuming far fewer tokens, so it costs less and runs faster. Translated for anyone putting AI into production: you no longer pay Ferrari money to go buy bread. It's a direct cost lever, not a nerdy detail. In the Opus 5 vs Opus 4.8 comparison it's the point that weighs most.
Pricing and Fast mode
The list price doesn't change: $5/$25 per million tokens, like Opus 4.8. What changes is the effective cost per task, which drops thanks to the effort dial and a model that wastes less to reach the result.
There's also a Fast mode: it runs at roughly 2.5 times the default speed, at double the base price. It's for when latency matters more than cost — think of a copilot inside a flow where the user is waiting for the answer on screen. If you want to estimate real spend on your volumes, the Claude pricing simulator gives you a ballpark in two minutes.
Get updates on Claude and AI for business
One email when there's something worth reading. No spam.
Evaluating Claude for your organization? Learn what it costs or which plan fits best
How much better it is (and where it isn't)
The jump shows on the ground that matters most today, the agentic one: planning, using tools, completing tasks autonomously. On CursorBench 3.2, at max effort, Opus 5 gets within half a point of Fable 5 while paying half the cost per task. On OSWorld 2.0 it beats every model at the same cost. On Zapier AutomationBench it reaches roughly 1.5 times the success rate of the second best.
Anthropic also describes it as "the most aligned Opus": the lowest score on its behavioral audit, the hardest to push toward misuse. Intellectual honesty, though: it stays behind Mythos 5 on biological research and offensive cybersecurity. It's not the model for everything. It's the flagship model for business work.
The two betas for developers
Along with the model, Anthropic shipped two beta features on the API that matter to anyone building agents. The first: you can change the tools available to Claude mid-conversation without invalidating the prompt cache. In practice an agent can gain or lose capabilities on the fly without paying for the context again — a concrete saving on long flows.
The second: automatic fallbacks. If a request is blocked by the safety classifiers, instead of failing outright it gets routed to another available model. Fewer interruptions in production. These are details, but they're the details that make the difference between a demo and an agent that actually runs. We go deeper in the Opus 5 for coding guide.
What changes in practice for your company
If you already run Opus 4.8, the move is almost mandatory: same price, more capability, more control over cost. Two things are worth reviewing.
First, model routing: with the effort dial you can keep Opus 5 on more tasks than before, lowering the effort where you'd previously have used a smaller model. Second, your agents: if you have any in production, Opus 5 at high effort closes tasks that used to fail, and at low effort saves you money on the simple ones. In both cases it pays to re-measure, not to take the setup from two months ago for granted. If you want to understand where Opus 5 really shifts the cost/quality equation, we wrote a dedicated guide on Opus 5 for business.
At Maverick AI this is exactly what we work on: picking the right model for each task and putting agents into production without burning budget. If you're weighing how to integrate Opus 5 into your processes, let's talk.