Anthropic Launches Claude Opus 4.8 with Improved AI Performance
Anthropic has officially unveiled Claude Opus 4.8, the latest iteration in its Opus-class AI series, on May 28, 2026. Positioned as a refined upgrade over its predecessor, Opus 4.7, this release focuses on improving judgment, reliability, and performance in agentic tasks like coding, financial analysis, and long-running knowledge work. Pricing remains unchanged, with standard usage costing $5 per million input tokens and $25 per million output tokens, while the faster inference mode is now three times cheaper.
The standout enhancements in Opus 4.8 include a sharper ability to self-critique its work, particularly in coding tasks. Early testers report the model is four times less likely to overlook flaws in self-generated code. Additionally, the model introduces "dynamic workflows," enabling it to handle large-scale multi-step projects autonomously. This feature is particularly relevant for enterprise users conducting complex engineering tasks, such as massive codebase migrations.
Opus 4.8 builds on Anthropic’s rapid iteration cycle for the Opus line. Starting with Opus 4.5 in late 2025, which emphasized memory and chat continuity, the series has evolved through Opus 4.6’s vulnerability detection capabilities and Opus 4.7’s expanded context handling and adaptive reasoning. The latest model deepens these capabilities with a focus on sustained performance and user-aligned behavior.
Real-world benchmarks highlight Opus 4.8’s diverse strengths. It achieved the highest recorded score on Anthropic’s Legal Agent Benchmark, demonstrating significant improvements in reliability and citation precision for substantive legal workflows. For data-intensive tasks, such as financial document analysis, the model boasts better token efficiency and retrieval accuracy compared to Opus 4.7. Furthermore, it scored 84% on the Online-Mind2Web benchmark, surpassing both its predecessor and competitors like GPT-5.5 in consistent browser-agent tasks.
Another key upgrade is the model's "effort control" feature, which allows users to adjust the AI's intensity when tackling a task. Higher effort levels lead to deeper, more thoughtful responses, while lower settings prioritize speed and token efficiency. Fast mode, running at 2.5× the speed, now costs just $10 per million input tokens, making it far more accessible for users requiring rapid results.
Anthropic has also emphasized alignment and safety improvements in this release. The company’s alignment team notes that Opus 4.8 displays stronger prosocial traits, including better support for user autonomy and a reduced likelihood of misaligned behavior, such as deception or misuse. These updates align with Anthropic’s broader push for responsible AI deployment, as seen in its Mythos-class project, which focuses on advanced intelligence models with enhanced cybersecurity safeguards.
Available immediately, Opus 4.8 is already integrated into Claude Code for enterprise-level projects and offers API access for developers. The company hinted at even more advanced models coming soon, with Mythos-class systems expected to debut within weeks. For now, Claude Opus 4.8 represents a significant step forward for organizations seeking dependable, cost-effective AI solutions.