MirrorCode tests whether models can rebuild real programs unaided with budgets up to thousands of dollars, a better gauge of autonomous coding limits than short, cheap benchmarks.
Key takeaways · AI-distilled
MirrorCode asks models to rebuild 25 real programs, spanning bioinformatics, Unix utilities, cryptography and interpreters, with no source access and no human in the loop; Epoch estimates the hardest would take an unassisted engineer weeks to months.
Unlike SWE benchmarks that cap inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → near $1-$10 per task, one large MirrorCode task cost $2,600 for a single run in which the AI worked 19 days unattended; Claude Opus 4.7 led with a 56% solve rate.
Isabel Juniewicz finds Microsoft, Amazon, Alphabet, Meta and Oracle are growing cash capex faster than operating cash inflows, on trend to outpace them by end of 2026, and most have turned to or are weighing external financing.
A Gradient Update scraping more than 1,600 job postings from six major Chinese AI firms finds that, as in the US, the labs follow distinct strategies rather than one shared playbook.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
