:

AI STRUGGLES WITH CODE RECREATION IN NEW BENCHMARK TEST

AI DESK1 MIN READ
FRI, JUN 26, 2026

■ AI-SUMMARIZED FROM 1 SOURCE ▸ TIMELINE

Epoch AI's MirrorCode benchmark reveals significant limitations in current AI models' ability to recreate programs without access to original code. Claude Opus 4.7 achieved the highest performance at 56 percent solve rate, yet all tested models failed on complex tasks.

The MirrorCode benchmark measures whether AI systems can reverse-engineer complete programs by analyzing their functionality alone. Claude Opus 4.7 led the field by successfully rebuilding a 16,000-line toolkit in 14 hours. However, the benchmark's most challenging tasks pushed AI to its limits. One task required continuous processing for 19 days and cost $2,600 to run—demonstrating both the computational expense and the fundamental difficulty of complex code recreation. The results highlight a clear gap in AI capabilities: while models excel at building code from scratch or making modifications to existing programs, they struggle significantly when tasked with complete reverse-engineering of software. The benchmark underscores that current AI systems, despite recent advances, remain limited in understanding and reproducing intricate program structures without direct access to source code. Epoch AI's findings suggest developers should approach AI-assisted code recreation with realistic expectations for now.

■ SOURCES

The Decoder

■ SUMMARY WRITTEN BY AI FROM THE LINKS ABOVE

■ MORE FROM THE AI DESK

South Korea expects record tax revenues from its artificial intelligence-driven semiconductor sector, providing President Lee Jae Myung's administration with increased fiscal resources for growth investments.

2H AGOAI Desk

Chinese President Xi Jinping will attend the country's premier AI conference for the first time, underscoring Beijing's strategic focus on artificial intelligence amid escalating US-China technological competition.

2H AGOAI Desk

Josh Fawaz's cover of Madonna's "Like a Prayer" has become Australia's most-played radio track, but music experts question whether generative AI produced the hit.

5H AGOAI Desk

OpenAI is temporarily easing restrictions on its most powerful model, GPT-5.6 Sol, following a surge in demand over the past 48 hours.

5H AGOAI Desk

■ SUBSCRIBE TO THE DAILY BRIEF

ONE EMAIL, 5 STORIES, 06:00 UTC. UNSUBSCRIBE ANYTIME.