Opus 5 of Claude is Released, and It's Changing the Way I See AI Construction

Anthropic announced the release of Opus 5 of Claude on July 24, 2026, and having gone through the official press announcement and the initial reactions among developers, as well as the discussion that has been building up on Reddit and other social networks in the few hours since, I felt I needed to get down on paper what truly matters about all of this. Being someone who works at the interface of art and technology day in and day out, such announcements impact the way I work in concrete terms, they change what I can actually build, and how fast I can move from an idea to something real.

Here is the short version. Opus 5 is a genuine step up from Opus 4.8, not just on paper but on the tasks that matter for long, complex creative and technical work. It costs the same as its predecessor, it is meaningfully better at coding and agentic tasks, and it introduces a new control that lets you decide how much effort the model puts into a task before it answers. For anyone building products, brands, or creative systems with AI as a genuine collaborator rather than a toy, this release is worth understanding properly.

What actually got released

Going by Anthropic's own press announcement, Opus 5 is described as a thoughtful, proactive model, one that comes close to the frontier intelligence of Claude's top-tier model, Fable 5, while costing roughly half as much to run. On coding and knowledge-work evaluations such as Frontier-Bench and GDPval-AA, Anthropic states that Opus 5 is now the new state of the art, though it still falls behind their specialized Mythos 5 model when it comes to cybersecurity-specific tasks.

The addition I found most practically interesting is the new effort control. With Opus 5, you can now choose how much thinking the model puts in before it responds, low, medium, high, and a newly introduced "xhigh" tier that sits above the previous ceiling, which effectively lets you trade cost against capability depending on what the task in front of you actually calls for. On CursorBench 3.2, Anthropic reports that at maximum effort, Opus 5 lands within half a percentage point of Fable 5's best score, while costing half as much per task. That is the sort of efficiency gain that genuinely changes how people budget AI into real workflows, rather than treating every request as a fixed, unavoidable cost.

The context window is the other headline change worth noting. Several independent outlets covering the launch, including developer guides published on the same day, reported that Opus 5 now ships with a one million token context window, a substantial jump that matters a great deal for anyone feeding it large codebases, lengthy documents, or entire design systems in a single pass.

Pricing, notably, did not move. It remains at five dollars per million input tokens and twenty five dollars per million output tokens, exactly what Opus 4.8 cost. It has also become the new default model on Claude Max, and it is the strongest model currently available on Claude Pro.

The benchmark results that stood out to me most

A handful of the specific figures in Anthropic's own writeup are worth pulling out on their own, since they point toward something larger than a leaderboard placement. On ARC-AGI 3, an evaluation built specifically around solving problems the model has genuinely never encountered before, Opus 5's score came in at three times that of the next-best model. On Zapier's AutomationBench, which measures whether a model can carry a real business task from start to finish, Opus 5's pass rate landed at roughly 1.5 times the next-best model for the same cost, and even at its lowest effort setting, it still outperformed every other model tested.

What caught my attention in the early-access accounts Anthropic published was less the raw number and more the behavior sitting underneath it. In one instance, Opus 5 was handed a drawing of a machine part and asked to rebuild it as a 3D model, but was deliberately given no direct way to view the image. Rather than fail the task, it wrote its own computer vision pipeline to pull the geometry out of the raw pixels and reconstructed the part from there, something no competing model managed even after five attempts under the same restriction. That is not a clever benchmark trick, it is a model working its way around a missing tool on its own initiative, which feels closer to how a genuine collaborator behaves than how a search engine behaves.

What the community has actually been saying

The run-up to this release had a story of its own. In the weeks before July 24, screenshots and leaked specifications of an "Opus 5" made the rounds across tech forums and social platforms, accompanied by a healthy dose of skepticism. One widely circulated piece of commentary put it plainly: treat any Opus 5 leak as fiction until it turns up in Anthropic's own documentation. That skepticism, as it turned out, was reasonable caution rather than cynicism, since the leaked details, the one million token context window and the new effort tiers among them, ended up matching what actually shipped fairly closely.

Once the release went live, reaction across developer communities and social platforms came in mixed, in a way that tends to be typical for a major model launch. There was genuine enthusiasm from people testing it against coding and agentic tasks, running alongside a separate, and somewhat louder, strain of criticism from people frustrated with how much power continues to concentrate among a small number of frontier AI labs, calling for stronger support of open models and less dependence on any single provider. Both reactions are worth taking seriously in their own right, the technical enthusiasm is backed by genuine benchmark movement, while the concern about concentration reflects a legitimate, ongoing conversation in this industry that no single model release is going to settle.

Early-access partners quoted directly in Anthropic's own announcement, among them teams at Cursor, Devin, Zapier, and Lovable, kept returning to the same point: Opus 5 handles longer, vaguer, harder tasks with noticeably less hand-holding than Opus 4.8 required, and it does so using fewer tokens and fewer turns along the way. That kind of consistency across independent teams, rather than just Anthropic's own internal testing, tends to be a far more trustworthy signal than any single benchmark figure on its own.

Why this matters for anyone building at the edge of art and technology

There is one idea I keep returning to in my own work: the tools ought to disappear so that the intent behind them can come through clearly. A model that needs constant correcting, that loses the thread partway through a long creative or technical task, or that produces something merely plausible rather than something actually reasoned through, gets in the way of that goal. What Anthropic describes in Opus 5, and what early testers keep echoing, is a model that verifies its own work as it goes, catches its own mistakes mid-task, and pushes back with an actual argument when it disagrees, rather than simply complying. One engineer's account of Opus 5 pushing back on a proposed redesign, narrowing its objection down to a single specific issue, and then proposing a compromise that preserved what was good about the idea while fixing the flaw, describes a genuinely different kind of interaction than earlier models tended to offer.

That distinction matters specifically for creative and brand work. A visual system, a product experience, a piece of generative art, each of these demands holding a large amount of context and intent consistently across a long stretch of work, which is exactly the kind of long-horizon task Opus 5 has been built around. The gap between a tool that simply executes instructions and a collaborator that understands why those instructions exist in the first place is the gap this release seems to be closing.

A note on safety and alignment

Anthropic also reported that Opus 5 came out as their most aligned model to date on internal behavioral audits, showing the lowest rates of deceptive behavior and the lowest susceptibility to misuse among their recent models. On cybersecurity in particular, they were explicit that Opus 5 does not advance the frontier of dual-use risk, it remains well behind their specialized Mythos 5 model at turning identified vulnerabilities into working exploits, even though it has become noticeably better at simply finding them. This kind of disclosure, including the places where the model is deliberately not the strongest option available, is worth paying attention to, since it is exactly the part of a launch announcement a company has the least incentive to volunteer on its own.