The Epoch Brief - June 26, 2026

2026-06-26 · Source: Epoch AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

Epoch AI has launched MirrorCode, a new long-horizon coding benchmark co-developed with METR, designed to measure autonomous AI's ability to complete large software projects. This benchmark challenges AI models to rebuild 25 real-world programs, including bioinformatics and cryptography, without source code or human assistance. Unlike typical benchmarks capped at \$1-\$10 per task, MirrorCode tasks can involve costs up to \$2,600 for a single run and require AI to work for 19 days. Claude Opus 4.7 currently achieves a 56% solve rate. Additionally, Epoch's data insights reveal that hyperscaler capital expenditures, from companies like Microsoft and Amazon, are projected to surpass their operating cash flows by late 2026, leading many to seek external financing for AI infrastructure. New Gradient Updates also analyze 1,604 Chinese AI job postings and propose an AI R&D taxonomy to track automation in research.

Key takeaway

For AI Scientists and Research Directors evaluating autonomous coding agents, MirrorCode provides a critical new benchmark for real-world software engineering tasks. Your models' performance on long-horizon, high-cost challenges, like those requiring 19 days of inference, will indicate true readiness for complex projects. Consider these metrics when assessing AI's ability to automate R&D, and factor in the increasing external financing trends of hyperscalers when planning infrastructure investments.

Key insights

MirrorCode sets a new standard for evaluating autonomous AI software engineering capabilities.

Principles

• Long-horizon benchmarks reveal true AI capabilities.
• Real-world software tasks require significant inference budgets.
• Hyperscaler AI investments outpace internal cash generation.

Method

MirrorCode evaluates AI by tasking models to rebuild 25 real-world programs without source code or human intervention, allowing for large inference budgets and extended run times.

In practice

• Use MirrorCode to assess AI's long-term coding prowess.
• Monitor hyperscaler capex trends for market shifts.
• Analyze job postings to understand AI lab strategies.

Topics

• MirrorCode
• AI Software Engineering
• AI Benchmarking
• Hyperscaler Capital Expenditure
• AI R&D Automation
• AI Job Market Analysis

Best for: AI Engineer, Machine Learning Engineer, Investor, AI Scientist, Research Scientist, Director of AI/ML

Related on AIssential

• Benchmarking inference at scale: coding agents — Production-scale coding agent inference requires full-stack optimization to manage concurrent long-context loads and prioritize Time To First Token (TTFT).
• An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run — The MirrorCode benchmark reveals AI's current limits in complex code recreation despite impressive partial successes.
• MirrorCode: Evidence that AI can already do some weeks-long coding tasks — AI systems are demonstrating the ability to complete complex, multi-week coding projects autonomously.
• True Positive Weekly #153 — The AI landscape is rapidly evolving across benchmarks, verification, agent skills, and economic integration.
• Finally a good benchmark (DeepSWE) — DeepSWE offers a robust, real-world software engineering benchmark revealing significant performance and cost disparities among LLMs.
• Beyond SWE-Bench Pro - Where do Agents go from Here? — New benchmarks are crucial for advancing AI coding agents beyond simple issue resolution to complex system understanding and improvement.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Epoch AI.