提前实测Opus 5：3D细节封神，然而有一个致命缺陷
Merged summary
TL;DR - Unverified early tests of Anthropic’s rumored Opus 5 suggest strong 3D/UI code generation and a possible 1M-token context window, but also extreme token consumption and weaknesses in basic visual recognition. Its reported release delay and capabilities remain based on leaks and community claims.
- Demos reportedly showed unusually detailed 3D scenes, games, and weather UI generated in one pass.
- Three conversations allegedly exhausted the tester’s usage allowance, indicating high inference cost.
- The model failed to identify a small insect in an image, suggesting uneven multimodal performance.
- Leaks describe configurable reasoning effort, stronger agent/tool use, and fallback to Opus 4.8 when confidence is low.
Sources (1)
提前实测Opus 5：3D细节封神，然而有一个致命缺陷
TL;DR - Unverified early tests of Anthropic’s rumored Opus 5 suggest strong 3D/UI code generation and a possible 1M-token context window, but also extreme token consumption and weaknesses in basic visual recognition. Its reported release delay and capabilities remain based on leaks and community claims.
- Demos reportedly showed unusually detailed 3D scenes, games, and weather UI generated in one pass.
- Three conversations allegedly exhausted the tester’s usage allowance, indicating high inference cost.
- The model failed to identify a small insect in an image, suggesting uneven multimodal performance.
- Leaks describe configurable reasoning effort, stronger agent/tool use, and fallback to Opus 4.8 when confidence is low.