An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run

2026-06-26 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, quick

Summary

Epoch AI has introduced the MirrorCode benchmark, designed to evaluate AI models' capability to reconstruct entire software programs solely from their functional descriptions, without access to the original source code. Initial results show Claude Opus 4.7 as the top performer, achieving a 56 percent solve rate. This model successfully rebuilt a 16,000-line toolkit in just 14 hours. However, the benchmark also highlights a significant limitation: all tested models consistently fail when confronted with the most complex programming tasks. One notable instance involved an AI model running nonstop for 19 days on a single MirrorCode task, incurring a cost of \$2,600, underscoring the computational intensity and current limitations in tackling intricate code generation challenges.

Key takeaway

For AI Engineers developing code generation tools, you should integrate the MirrorCode benchmark to rigorously assess model capabilities. While Claude Opus 4.7 demonstrates strong performance on moderately sized projects, anticipate that current models will struggle with highly complex programming challenges. Focus your development efforts on improving robustness for intricate tasks and factor in substantial compute costs, like the \$2,600 for a 19-day run, when planning advanced code recreation projects.

Key insights

The MirrorCode benchmark reveals AI's current limits in complex code recreation despite impressive partial successes.

Principles

• AI excels at recreating moderately complex codebases.
• Significant challenges remain for highly complex programming tasks.
• Code recreation benchmarks quantify AI programming limits.

Method

The MirrorCode benchmark evaluates AI by requiring models to recreate complete programs from functional descriptions, without original source code access.

In practice

• Use Claude Opus 4.7 for code recreation tasks.
• Expect failures on highly complex code generation.
• Budget for significant compute costs on complex tasks.

Topics

• MirrorCode Benchmark
• AI Code Generation
• Claude Opus 4.7
• Program Reconstruction
• AI Model Evaluation
• Computational Costs

Best for: Research Scientist, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• OPUS 4.8!!! (also maybe GPT5.6??) — Opus 4.8 enhances AI agentic capabilities and cost-efficiency through parallel processing and optimized inference.
• SWE-bench February 2026 leaderboard update — Independent SWE-bench Verified results show Claude 4.5 Opus leading in coding problem resolution.
• Claude Opus 4.7: Most Powerful Coding Model Ever! Beats EVERYTHING! (Fully Tested) — Claude Opus 4.7 significantly advances coding, web development, and vision, but at a higher token cost.
• If You’re Still Hitting the Claude Code 5-Hour Wall, You’re Doing It Wrong — The Claude Code 5-hour rate limit is caused by four inefficient habits: routing, delegation, caching, and hygiene, not usage volume.
• Claude Opus 4.6 vs GPT-5.3 Codex: How I shipped 93,000 lines of code in 5 days — Pairing generative AI models with specialized review models optimizes complex software development workflows.
• Claude Opus 4.8: Capabilities and Reactions — Comprehensive AI model evaluation requires diverse benchmarks, user feedback calibration, and understanding model "welfare" for accurate assessment.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.