Anthropic has released its new model Opus 4.8, which has been tested in early-access form by a reporter. The model excels at creating greenfield prototypes and one-shot features quickly, but struggles with edge cases in existing codebases and hallucinations. In business strategy work, Opus 4.8 is comparable to its predecessor Opus 4.7, but the reporter still prefers Opus 4.7 for data-heavy tasks like roadmap planning.
Anthropic has also released new features alongside Opus 4.8, including dynamic workflows with parallel subagents and effort control in Claude.ai and Cowork. The model is priced at a benchmark performance level, although exact pricing details are not provided.
The reporter conducted extensive testing of Opus 4.8 on various tasks, including building a prototyping tool and creating games for a 9-year-old. While the model shows promise in certain areas, it still struggles with the "last 10%" problem and hallucinations, where it produces inaccurate or irrelevant output.
