Claude Opus 5: Performance and Agentic Coding Impressions

Performance and Benchmarks

Claude Opus 5 demonstrates competitive performance against Fable 5, particularly in agentic coding and front-end design tasks. On the RKGI 3 benchmark, the model achieved a 30% score, representing a significant jump over previous state-of-the-art models. Third-party evaluations, including Cursor and Frontier Code, place Opus 5 on par with or exceeding Fable 5, depending on the configured effort level. Anthropic has maintained the pricing structure of Opus 4.8 while delivering these intelligence gains, though the model exhibits slower reasoning times during complex tasks compared to its predecessor.

Agentic Capabilities and Creative Demos

The model shows increased agency, successfully executing complex tasks like reconstructing 3D geometry from raw pixel data without explicit visual input. In creative testing, the model generated a Pokémon encyclopedia website that avoided standard template-based layouts, and a crowd animation forming text with functional camera controls. The model also produced a detailed ISS tracker, providing real-time positioning, orbital footprints, and predicted paths. While the model is highly capable, the author noted potential inconsistencies in benchmark reporting, specifically citing a chart where a lower percentage value was visually represented as higher than a greater one.

Operational Notes

Anthropic intentionally reduced the cybersecurity capabilities of Opus 5 to facilitate a faster release schedule. The model is currently available via claude.ai and Claude Code, the latter supporting a 1 million token context window. While the model is token-efficient, the increased reasoning effort required for complex agentic tasks can lead to higher latency.