Anthropic announced the release of Opus 5 on Friday, the latest iteration of its advanced artificial intelligence model. While previously the company’s flagship offering, Opus 5 now sits just below the more powerful Fable 5, yet it delivers a compelling performance profile that rivals its predecessor, Fable 5, across numerous domains at a significantly reduced cost. This strategic release positions Opus 5 as a highly accessible yet potent tool for a wide array of AI applications, particularly for enterprise and everyday use cases.
The new model retains the cost structure of its predecessor, Opus 4.8, with input tokens priced at $5 per million and output tokens at $25 per million. A key differentiator for Opus 5 is its user-friendly data policy; unlike Fable 5, users are not required to opt into a 30-day data retention policy to leverage its capabilities. This makes Opus 5 the default, most advanced model available to Claude Max subscribers and the top-tier option for Claude Pro subscribers, enhancing accessibility and user experience.
Opus 5: Bridging the Gap to Fable 5
Anthropic emphasizes that Opus 5 represents their most capable Opus model to date. It demonstrates an enhanced ability to autonomously execute tasks over extended periods, incorporating self-correction mechanisms and robust error recovery protocols. This increased autonomy and reliability are crucial for complex, multi-stage projects.
However, Anthropic continues to position Fable 5 as the premier choice for "your most ambitious work" and for projects demanding multi-day autonomous operation. Opus 4, on the other hand, is designated for routine, daily operational tasks. This tiered approach allows users to select the most appropriate model based on the complexity and duration of their AI-driven endeavors.
Despite Fable 5’s designation for highly demanding tasks, benchmarks shared by Anthropic indicate that Opus 5 often outperforms Fable 5 across a broad spectrum of evaluations. This suggests that for many applications, Opus 5 may offer a superior balance of performance and cost.
The recent introduction of Sonnet 5, launched just weeks prior, now occupies a more nuanced position within Anthropic’s model ecosystem. Positioned as an efficient option for everyday tasks, Sonnet 5 finds itself between the cost-effective and rapid Haiku 4.5 and the more powerful, albeit more expensive, Opus and Fable families. The emergence of Opus 5, with its near-Fable 5 capabilities at a lower price point, further refines the competitive landscape for AI models.
Performance Benchmarks: Opus 5’s Competitive Edge
Opus 5 exhibits particular strength in knowledge-intensive tasks and demonstrates significant improvements in coding capabilities compared to Opus 4.8. On the GDPval-AA v2 knowledge work benchmark, Opus 5 achieved a score of 1861, surpassing Fable 5 (1747), GPT-5.6 Sol (1736), and Opus 4.8 (1593). This performance indicates a substantial leap in its ability to process and synthesize complex information.
Furthermore, Opus 5 has topped Zapier’s AutomationBench, a critical evaluation of AI agents performing end-to-end business workflows. This benchmark measures the effectiveness and efficiency of AI in automating real-world business processes. Anthropic reports that Opus 5’s success rate on this benchmark is double that of the next best model at the same cost per task. Even when operating at its lowest effort setting, Opus 5 reportedly surpasses all other models in task completion. This achievement is particularly noteworthy given the rapid advancements in AI agent capabilities, including recent updates to Google’s Gemini models.
On the DeepSearchQA and Humanity’s Last Exam benchmarks, particularly when equipped with tools, Opus 5 also emerges as Anthropic’s most cost-effective and capable model. It even slightly edges out Fable 5 on these evaluations, scoring 64.7 percent compared to Fable 5’s 63.9 percent. However, when tool access is withheld, Fable 5 maintains its lead in these specific scenarios, highlighting the synergistic potential of integrating AI models with external tools.
Fable 5 retains its advantage in a select few areas, including Anthropic’s proprietary legal agent benchmark and DeepSWE, a benchmark for software engineering tasks. Mythos 5 remains the company’s strongest performer for health-related benchmarks, underscoring the specialized strengths within Anthropic’s model family.
In the realm of coding, Opus 5 demonstrates remarkable performance-per-cost efficiency on CursorBench 3.2. Across high, extra-high, and max effort settings, it outpaces all other models. At its maximum effort configuration, Opus 5 achieves a score within 0.5 percent of Fable 5’s peak performance, but at half the cost per task. Sualeh Asif, co-founder of Cursor, commented on Opus 5’s capabilities, stating, "Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost. On CursorBench it’s just under Fable 5 and has many of the same behaviors. We are excited to see how developers use it in Cursor."
Users can continue to select from various effort settings—low, medium, high, extra-high, and max—across Anthropic’s products. These settings dynamically adjust the model’s intelligence and token consumption, allowing for tailored resource allocation.
Safety and Alignment: A Core Focus
Anthropic also highlights Opus 5 as its most aligned model to date, based on internal automated behavioral audits. Alongside Fable 5, it exhibits a significantly reduced susceptibility to misuse or manipulation. The company has intentionally refrained from training Opus 5 on cybersecurity exploitation tasks. Despite this, Opus 5 demonstrates performance comparable to Mythos 5 in identifying vulnerabilities in open-source code, while crucially lagging in the ability to exploit them. This carefully calibrated approach aligns with Anthropic’s commitment to responsible AI development.
To mitigate potential risks, Opus 5 is equipped with safety classifiers designed to screen requests for a specific set of cyber-related tasks, including exploit generation and penetration testing. These classifiers are scoped more narrowly than those in Fable 5, leading Anthropic to anticipate approximately 85% fewer interventions. This targeted approach aims to permit legitimate defensive cybersecurity work, such as vulnerability discovery, while blocking potentially malicious activities like "binary-based vulnerability scanning," a technique Anthropic associates with malicious actors.
When a classifier flags a request within Claude.ai, Claude Code, or Cowork, the system defaults to Opus 4.8. API users have the option to enable similar fallback mechanisms. Anthropic’s Cyber Verification Program offers select enterprises and researchers a version of Opus 5 with fewer restrictions, facilitating advanced security research under controlled conditions.
The emphasis on cost-effectiveness is a recurring theme in Anthropic’s product releases. As businesses increasingly scrutinize operational expenses, the perceived value and pricing of AI models become paramount. This strategy acknowledges the current market sensitivity to token costs and model pricing.
Niko Grupen, head of applied research at Harvey, a legal AI company, noted significant gains with Opus 5 in practice areas such as corporate governance and arbitration. Grupen further observed, "We were also impressed with Opus 5’s ability to maintain quality at lower reasoning levels, achieving similar performance while generating 26% fewer tokens on average compared to Opus 4.8 at max reasoning." This indicates substantial efficiency gains for established users migrating from earlier versions.
Enhanced Speed and Platform Updates
Mirroring Opus 4.8, Opus 5 will offer a "fast mode" accessible through the Claude Platform for developers and via extra usage credits within Claude Code. This mode accelerates output token generation by 2.5 times, at double the cost.
Two additional beta platform updates are also being introduced. Automatic fallbacks allow API requests flagged by Opus 5 or Fable 5’s safety classifiers to be rerouted to an alternative model within the same request, preventing errors and ensuring workflow continuity. Developers can now also modify the tools Claude can utilize mid-conversation without invalidating the prompt cache. This feature enables more dynamic and adaptable agent workflows, where each stage of an agent’s operation only has access to the necessary tools.
Opus 5 is currently available on all of Anthropic’s paid subscription tiers and on the Claude Platform under the identifier claude-opus-5.
Historical Context and Market Positioning
The Opus model family has a notable history of advancement. Opus 4, released in May 2025, laid the groundwork for subsequent iterations. Opus 4.5, launched in November 2025, and Opus 4.6 in February 2026, solidified the family’s reputation as a leading choice for coding tasks. While Opus 4.7, released in April 2026, was met with mixed reactions, Anthropic quickly followed with the well-received Opus 4.8 just six weeks later, demonstrating a rapid development cycle.
With the introduction of the Mythos and Fable series, Anthropic established a new tier of flagship models. However, the company anticipates Opus 5 will serve as the go-to model for daily operations, particularly within enterprise environments.
This release occurs amidst intense competition. OpenAI recently launched its GPT-5.6 family, including the Sol flagship model, which is priced slightly higher than Opus 5 at $5 per million input tokens and $30 per million output tokens. Anthropic’s benchmark comparisons show Opus 5 outperforming GPT-5.6 Sol on most metrics, including knowledge work, computer use, and business workflows, with GPT-5.6 Sol holding a lead only on the agentic coding test DeepSWE v1.1. Notably, Google’s Gemini models were not included in Anthropic’s comparative benchmarks, suggesting Anthropic’s assessment of their relative performance.
The strategic positioning of Opus 5 underscores Anthropic’s ongoing effort to balance cutting-edge AI capabilities with practical accessibility and cost-effectiveness, aiming to empower a broader range of users and applications.
