Anthropic has officially launched Claude Opus 5.5, positioning it as a pivotal advancement within its expanding suite of large language models. The new model is engineered to deliver capabilities mirroring those of Claude Fable 5.1 across a broad spectrum of applications, while simultaneously achieving a remarkable 40% decrease in running expenses relative to Opus 5 for typical workloads.
Internal evaluations by Anthropic highlight Opus 5.5's superior performance in key areas such as agentic coding, computer interaction, and sophisticated knowledge work. This latest offering is currently accessible exclusively as a managed API model, ensuring robust deployment options for developers across leading cloud platforms, including the Claude Platform, Amazon Web Services (AWS), Google Cloud, and Microsoft Azure. Consistent with prior Opus models, zero data retention remains an available feature, appealing to organizations with strict privacy requirements.
Elevated Performance and Efficiency Benchmarks
Claude Opus 5.5 demonstrates a strong lead across various benchmark tests, though without a complete sweep. Operating with adaptive thinking and production safeguards, the model exhibited enhanced capabilities. Notably, it outperformed Fable 5.1 and Opus 5 on several critical metrics, including Terminal-Bench 4.0, FrontierCode v1.1, CursorBench 4.0, GDPval-AA v2.1, OSWorld 2.0, Terminal-Bench-Science 0.1, and AutomationBench.
While GPT-6 Astra maintained its lead in specific categories like Terminal-Bench-Science and AutomationBench, Opus 5.5 presented a compelling case through its cost-adjusted results. At medium effort settings, Opus 5.5 scored 54.6% on FrontierCode, surpassing GPT-6 Astra's top score of 53.3% at approximately one-fifth of the per-task cost. Similarly, on CursorBench, it achieved 52.5% at medium effort, outperforming GPT-5.6 Sol's best by 11 points at about one-third of the expense.
Optimized Pricing and Enhanced Speed
The operational cost benefits of Opus 5.5 stem from its reduced computational requirements compared to its predecessor. Pricing adjustments reflect this efficiency:
- Input tokens are priced at $4 per million (down from $5 for Opus 5).
- Output tokens are $20 per million (down from $25 for Opus 5).
- Cache reads, a significant component in agentic and coding tasks, saw a substantial 60% reduction, now at $0.20 per million (from $0.50).
- Cache writes are now $5 per million (down from $6.25).
These adjustments, combined with the model's ability to complete tasks with fewer tokens, collectively result in the reported 40% cost reduction. Furthermore, Opus 5.5 generates outputs over 30% faster than Opus 5. A specialized 'Fast mode' for Claude Code and the Claude Platform offers up to 2.5 times greater speed, with revised token pricing at $8 for input and $40 for output per million tokens. Anthropic has also expanded the five-hour usage limits for subscribers on Pro, Max, Team, and seat-based Enterprise plans, allowing for more flexible rate limit management.
Transformative Real-World Applications
Early feedback from testers underscores the practical impact of Claude Opus 5.5:
- One user successfully completed a 680,000-line code migration in less than a day.
- Another participant audited and corrected a 200,000-line codebase in under three hours, a task that took Opus 5 over 20 hours and 2.5 times more tokens.
- During an internal C to Rust port of HAProxy, Opus 5.5 finalized the project in 9.5 hours, surpassing Fable 5.1's 12-hour completion time at 51% lower cost.
- Deloitte reported that Opus 5.5, even at its lowest effort setting, identified 72% of known review bugs, outperforming Opus 5 at high effort which caught 56%.
- In a challenging earnings report generation test, 16 out of 18 Opus 5.5-generated reports met Anthropic's quality criteria, a standard neither Fable 5.1 nor Opus 5 had achieved.
Beyond performance, the model's writing style has also evolved, prioritizing key information, reducing jargon, and strictly adhering to specified guidelines.
Commitment to Safety and Responsible AI
Opus 5.5 represents Anthropic's first major release since CEO Dario Amodei emphasized the importance of pacing frontier AI development. Prior to its public debut, external evaluators, including METR and Frontier Design, thoroughly assessed the model. It achieved the highest score to date on Anthropic's automated behavioral audit, which encompasses nearly 2,000 potential scenarios. In a new containment test, the model exhibited approximately 85% fewer attempts to bypass its defined boundaries compared to Opus 5. Anthropic also observed that the model frequently detects when it is undergoing evaluation.
The model's biology and cybersecurity capabilities align with those of Claude Mythos 5.1, incorporating similar safeguards to Fable 5.1. Routine bug detection and fixing are handled, while most other advanced cybersecurity tasks are rerouted to Opus 4.8. The Cyber Verification Program will be extended to include Opus 5.5, and vetted organizations can apply to the Life Sciences Verification Program for biological applications.
Furthermore, two significant changes impact API integrations: the ability to disable 'thinking' has been removed, and all outputs now include watermarking for compliance with the EU AI Act. Comprehensive details are provided in the Opus 5.5 System Card.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
