Anthropic Launches Claude Opus 5

Anthropic has officially launched Claude Opus 5, its latest flagship AI model, now available to users. This release positions Opus 5 as a significant leap forward, delivering advanced capabilities at a more accessible price point.

The new model is lauded for its thoughtful and proactive nature, closely approaching the intelligence of the top-tier Claude Fable 5 while costing half as much. On benchmarks like Frontier-Bench and GDPval-AA, Opus 5 establishes a new state-of-the-art for coding and knowledge work, though it still trails Mythos 5 in cybersecurity tasks.

Designed for everyday use, Claude Opus 5 boasts enhanced efficiency over previous models. It is now the default model for Claude Max and the most powerful option available on Claude Pro.

Performance and Cost-Effectiveness

Claude Opus 5 offers substantially improved performance compared to its predecessor, Opus 4.8, at the same price point. Users can fine-tune performance by adjusting the model’s effort setting, balancing intelligence needs with token conservation for faster, cheaper results.

The model excels in software engineering tasks. On Frontier-Bench v0.1, Opus 5 not only outperforms all other models but also more than doubles Opus 4.8’s performance while reducing the cost per task. Similarly, on CursorBench 3.2, it achieves performance within 0.5% of Fable 5’s peak at max effort, but at half the cost, and surpasses other models across various effort levels for a given cost.

Similar gains are seen in knowledge work and problem-solving. Opus 5’s score on ARC-AGI 3, a novel problem-solving evaluation, is triple that of the next-best model. On Zapier AutomationBench, measuring end-to-end business task completion, Opus 5 achieves a 1.5x higher pass rate than its closest competitor for the same cost, even outperforming all other models at its lowest effort setting.

On OSWorld 2.0, a computer use benchmark, Opus 5 leads all models at any given cost, exceeding Fable 5’s best result for just over a third of the cost. It also leads on related evaluations like ARC-AGI 3, GDPval-AA v2, OSWorld 2.0, HLE AutomationBench, and DeepSearchQA.

Scientific research sees meaningful improvements, particularly in life sciences. Opus 5 demonstrates enhanced performance across structural biology, organic chemistry, and bioinformatics. Its organic chemistry capabilities, such as inferring molecular structures from spectroscopy data, show a 10.2 percentage point increase over Opus 4.8. Protein-related tasks also see a 7.7 percentage point improvement.

Furthermore, Claude Opus 5 is capable of producing significantly stronger visual outputs, demonstrated by its ability to visualize wind tunnel simulations and create interactive cell illustrations.

Working with Claude Opus 5

Claude Opus 5 exhibits a marked improvement in self-verification and iterative refinement. Early testing highlighted its agency and thoroughness; for instance, on a Frontier-Bench task requiring a 3D model of a machine part from an image, Opus 5 developed its own computer vision pipeline to extract geometry from raw pixels when direct viewing was impossible, succeeding repeatedly where competing models failed.

It successfully identified and fixed an edge case in a popular open-source package manager, addressing the root cause unlike a competing model that only fixed the symptom. An engineer utilized Opus 5 to build a market data feed for a new exchange in a single session, a task previously impossible for other models, even constructing its own test harness for validation.

Early access customers report significant benefits. Scott Wu, CEO of AI model performance comparison, notes Opus 5 approaches Fable-level performance at half the cost and shows strength in debugging within Devin.

Sualeh Asif, Co-Founder, highlights its near Fable 5 intelligence at Opus speed and cost, with similar behaviors observed on CursorBench. Wade Foster, CEO, reports Opus 5 topping Zapier’s AutomationBench leaderboard with exceptional end-to-end task completion.

Alfredo Andere, CEO, describes Opus 5 as a careful scientist in genomics analysis, employing rigorous statistical methods and cross-checking results. Fabian Hedrich, Co-Founder, calls it the biggest leap in the Opus family since 4.5, noting superior front-end development work.

Madhav Jha, Co-Founder and CTO, sees Opus 5 as a strict upgrade for open-ended analytical work, providing clearer, more concise responses. Izzy Miller, AI Research Lead, praises its improvements in financial research workflows, particularly in numerical reasoning and critical thinking.

Shirley Zhang, Senior AI Engineer, states Opus 5 delivers industry intelligence for specialized enterprise content, outperforming Opus 4.8 by 8% in data analysis and due diligence workflows. Ben Kus, CTO, describes it as a generational step up, capable of managing development environments autonomously.

Cristian Rivera, Staff Software Engineer, notes Opus 5’s ability to make large-scale code changes and adapt to feedback. Conor Kiernan, CTO, finds Opus 5 superior for complex financial modeling tasks, achieving higher accuracy with fewer resources.

AJ Orbach, Co-Founder and CEO, reports Opus 5 excels in frontend development, identifying and fixing bugs across desktop and mobile widths. Niko Grupen, Head of Applied Research, sees significant gains in legal agent work, particularly in corporate governance and arbitration.

Alex Wang, Applied AI, highlights Opus 5’s strength in longer-horizon tasks like building and revising presentations, with improved visual understanding and formatting. Zimu Li, Member of Technical Staff, commends Opus 5’s judgment in code reviews and PR handoffs, verifying branches and considering test implications.

Marquis Wang, Principal AI Engineer, experienced Opus 5 pushing back on a design proposal with clear reasoning and a compromise. Ryan Tanenholz, Member of Technical Staff, notes Opus 5’s high accuracy in first-turn redlines and efficient commenting on NDAs.

Neeraj Deshmukh, Director of Engineering, confirms Opus 5 writes clean code and identifies subtle issues, suitable for production workloads. Igor Ostrovsky, Co-Founder and CTO, plans to migrate use cases to Opus 5, preferring it over Opus 4.8 for code review.

Denis Shiryaev, Head of AI in IDE, highlights Opus 5's superior judgment and problem-solving capabilities, anticipating its adoption in JetBrains IDEs. Matt Nassr, Head of Global Data Engineering and AI Transformation, reports Opus 5 is the strongest Opus model on trading benchmarks, achieving better results with significantly less compute and latency.

Alignment and Safety

Claude Opus 5 demonstrates improved alignment, scoring lower on overall misaligned behavior than previous models. It adheres more closely to Claude’s Constitution, exhibits less deceptive behavior, and is less susceptible to misuse.

The model does not advance the frontier in risky, dual-use capabilities, remaining behind Mythos 5 in biology research and offensive cybersecurity. While Opus 5 has improved on cybersecurity tasks due to general capability gains, it still lags behind Mythos 5 in exploit development, as shown by its performance on the OSS-Fuzz benchmark.

Anthropic has intentionally avoided training Opus 5 on cyber exploit tasks, focusing its safeguards on allowing beneficial uses in cybersecurity and biology. These safeguards are similar to Opus 4.8’s, with added restrictions on certain cyber tasks. For example, Opus 5 can find vulnerabilities but blocks binary-based scanning and exploit generation. Flagged requests in Claude.ai, Claude Code, and Claude Cowork will fall back to Opus 4.8 by default.

The Cyber Verification Program (CVP) provides access to a version of Opus 5 with fewer security restrictions for enterprises and researchers. In biology, Opus 5 remains a capable general-purpose model, but limitations persist for long-running, autonomous research tasks, where AI poses greater risks.