Anthropic的Claude Fable 5在最新的MirrorCode编程排行榜上以64%的绝对成功率位居首位，远超GPT-5.6 Sol等竞争模型，后者成功率仅为前者的三分之一[1]。这一成绩表明该模型已具备从零开始构建完整软件项目的能力，而非仅依赖对训练语料的记忆。
在语言适配方面，Claude Fable 5在高资源语言Go中的解出率达到64%，在冷门语言Ada中的解出率为61%，仅下降3个百分点[1]。这一表现尤为突出，因为Python语料库的规模约为Ada的230倍，而模型成绩却仅下降3%[1]。MirrorCode评测包含25个目标程序，涵盖Unix工具、解释器、数据查询等多个领域，要求模型在隔离环境中通过反复调试达到100%通关率[1]。
MirrorCode的评测规则极为严苛：可见测试与隐藏测试的完成率必须同时达到100%才算通过一个项目[1]。为应对这一难度，评测单次预算已提升至100亿Token，最长允许模型连续运行7天，单次任务最高成本达2600美元[1]。根据Epoch的估计，人类工程师完成同一任务需要2到17周[1]。
Anthropic's Claude Fable 5 has claimed the top position on the MirrorCode programming leaderboard with an absolute success rate of 64%, significantly outpacing competitors including GPT-5.6 Sol, which achieved approximately one-third of that performance [1]. The benchmark demonstrates the model's capability to build complete software projects from scratch rather than relying solely on memorized training data [1].
The model's performance across different programming languages reveals its robust generalization abilities [1]. In Go, a resource-rich language, Claude Fable 5 achieved a 64% success rate, while in Ada, a less common language with approximately 230 times less training data than Python, it maintained a 61% success rate—a decline of just 3 percentage points [1]. MirrorCode's evaluation framework requires models to achieve a 100% pass rate through iterative debugging in isolated environments, with a single task budget of 10 billion tokens and a maximum continuous runtime of 7 days [1]. The benchmark encompasses 25 target programs spanning domains such as Unix utilities, interpreters, and data querying systems [1].
The computational demands reflect the complexity of the tasks, with some individual attempts reaching costs of up to $2,600 and models running continuously for as long as 19 days [1]. Anthropic's research indicates that human engineers would typically require two to seventeen weeks to complete equivalent tasks [1]. The company has also demonstrated practical applications of its AI agents in real-world development scenarios, including using 16 Claude agents to build a C compiler capable of compiling the Linux 6.9 kernel [1].
