Summary: Anthropic released Claude Opus 5 on July 24, 2026 as a daily-driver model that approaches Fable 5’s frontier intelligence at half the price. The review highlights strong cost-adjusted coding, knowledge-work, automation, and computer-use results, along with judgment and self-verification in long tasks. It still trails Mythos 5 on cybersecurity, so teams should test effort levels on ambiguous workloads.
What’s New: A Daily-Driver That Gets Close to Frontier Intelligence
Claude Opus 5 arrived from Anthropic on July 24, 2026, framed as a “thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.” For people building products, the appeal isn’t a bigger spec sheet—it’s a way to reach near-top-tier capability without paying top-tier rates.
Anthropic’s announcement puts Opus 5 at the new state of the art for coding and knowledge work evaluations such as Frontier-Bench and GDPval-AA, though it still trails Mythos 5 on cybersecurity tasks. The model is built for everyday use, runs more efficiently than other models, and now serves as the default on Claude Max and the strongest model on Claude Pro.
Performance and Cost: Beyond Simply Being Cheaper
According to Anthropic’s data, Opus 5 delivers greatly improved performance at the same cost as its predecessor, Opus 4.8. The model’s effort setting lets customers either push for maximum intelligence or conserve tokens for results that come faster and cheaper.
On Frontier-Bench v0.1, Opus 5 beats every other model, more than doubling Opus 4.8’s performance while costing less per task. On CursorBench 3.2 at max effort, it lands within 0.5% of Fable 5’s peak score—at half the cost per task. Across the high, xhigh, and max effort settings, it also returns more performance for a given cost than any other model.
Knowledge work and problem-solving follow the same pattern:
- On ARC-AGI 3, which probes novel problem-solving, Opus 5 scores three times as high as the next-best model.
- On Zapier AutomationBench, which measures end-to-end business task completion, Opus 5 passes at roughly 1.5× the rate of the next-best model for the same cost per task. Even at its lowest effort setting, it clears more tasks than any other model.
- On OSWorld 2.0, the computer use benchmark, no other model keeps up with Opus 5 at any given cost, and its result edges past Fable 5’s best while spending just over a third of the cost.
Scientific research improves too: across all life sciences evaluations—structural biology, organic chemistry, and bioinformatics among them—Opus 5 is a meaningful step up from Opus 4.8. It scores 10.2 percentage points higher on organic chemistry tasks (such as inferring molecular structures from spectroscopy data) and 7.7 percentage points higher on protein-related tasks (such as predicting how sequence variations affect function).
Token efficiency underpins much of the cost story. A legal tech company reported that Opus 5 held similar quality to Opus 4.8 at max reasoning while generating 26% fewer tokens on average.
Where It Actually Shines: Judgment and Self-Verification
The word “judgment” came up independently from several early-access customers. One engineer recounted a rearchitecting session in which Opus 5 pushed back on a proposed design and held its ground when the engineer insisted—not by digging in, but by laying out what was valuable in the idea, reducing its objection to a single design question, and offering a compromise that preserved the good part while repairing the flaw. Behavior like that is what lets developers hand work over with less oversight.
Verifying its own work and iterating carefully is another area where Opus 5 stands out. Anthropic and users have collected several illustrations:
- On a Frontier-Bench task, Opus 5 was handed a drawing of a machine part and asked to write code rebuilding it as a 3D FreeCAD model—with no way to view the drawing directly. It wrote its own computer vision pipeline, pulled geometry out of raw pixels, and reconstructed the part successfully, repeatedly. No competing model managed it after five attempts.
- Facing a real bug in a popular open-source package manager, Opus 5 located the root cause and repaired an edge case the community’s patch had missed. A competing model touched only the surface symptom and reported the bug resolved.
- An engineer at a trading firm had Opus 5 build a market data feed for a new exchange in a single session, a task previous models could not complete at all. With no live feed to validate against, Opus 5 stood up its own test harness to confirm that its code parsed the exchange’s data correctly.
Customers described the same thoroughness elsewhere. In a frontend benchmark, Opus 5 opened pages in a browser at desktop and phone widths, caught a product hidden below the mobile fold and an off-screen checkout button, and fixed both before returning the work. A monitoring agent built on Opus 5 flagged a potential anomaly, went back to production to re-check its assumption, determined the signal was benign, wrote the correction into its memory, and retired its monitoring queries by itself.
In Practice: From Code to Long-Horizon Work
For product builders, the payoff isn’t “stronger” so much as “more reliable.” On their hardest agentic coding tasks, Lovable measured Opus 5 at 22% above Opus 4.7, with far less variance run to run. That consistency matters enormously for products that depend on AI generating consistent results.
Long-horizon work is another strength. One customer put Opus 5 in charge of dev environments as a kind of chief of staff: the monitor was its own build, every box got driven by it, and a human was pulled in only when a judgment call was needed. Another pointed it at financial research and saw accuracy average 9 percentage points higher, with a third fewer turns and tool calls and 60% less time.
Agentic workflows get larger changes out of it, too. One customer watched it rework a codebase at scale, take in feedback as the agentic workflow ran, and lay out its reasoning more clearly than any model they had used—taking on work they would normally have chopped into much smaller pieces.
Code review gets the same treatment. Diffs come back clean and tight, with no dead code left behind, and the model is a sharper spotter of hazards hiding in subtle issues that are specific to a codebase. Branches get verified, templates get checked, and test implications get reasoned through before any PR is handed off—the rushed errors of earlier models don’t show up.
Where It Falls Short
Opus 5 is not an across-the-board upgrade. Anthropic acknowledges it still trails Mythos 5 on cybersecurity, particularly at exploiting vulnerabilities. On OSS-Fuzz, Opus 5 comes close to Mythos 5 at identifying vulnerabilities but is considerably less successful at developing exploits. Nor does it advance the frontier in risky dual-use capabilities such as biology research.
Because official benchmarks come from specific evaluations, real-world results shift by task type. Product builders should test inside their own workflows, paying attention to the effort setting—Anthropic’s different effort levels exist to balance intelligence against token cost.
Safeguards differ as well. Where Fable 5’s cyber classifiers are tighter, Opus 5’s permit vulnerability finding; what stays off limits is binary-based scanning, penetration testing, and exploit generation. Anything that does get flagged defaults to a fallback on Opus 4.8, and the API supports automatic fallbacks.
The Takeaway: Test It on Hard, Vague Tasks
Pricing sits exactly where Opus 4.8 sits: $5 per million input tokens, $25 per million output tokens. Opus 5 is available today on every platform, and developers can start with claude-opus-5 on the Claude API. A Fast mode runs at about 2.5× default speed for twice the base price.
The most direct step for developers: try Opus 5 inside integrated environments such as Cursor, Devin, and Kiro, then put it head to head with Opus 4.8 on difficult, ambiguous tasks—multiple customers noted that this is where Opus 5 improves most.
If cost efficiency matters most to you, the low effort setting may be enough for daily tasks; for complex, multi-step projects, raise the effort level. Whether Opus 5’s judgment turns into a competitive advantage for your product ultimately depends on how you design workflows to take advantage of its self-verification and long-horizon planning strengths.