AI bug-fix rates plateau near 90%
AI just jumped from fixing about half of real code problems to almost all of them on one public test. The next question is whether the test is now too easy.
It used to be a party trick. Now the machine quietly closes tickets a junior engineer would sweat over.
From 1-in-25 to most of them in roughly two years.
- The best AI coding systems now try fixes, run checks, and keep going until the code works.
- The public test uses real problems from open-source software, so the jump still matters.
- But the score is now so high that the test may be running out of room to show the next leap.
- Sep 2025 · 65%
First called: rising, 65% confidence — the curve had clearly turned.
- Mar 2026 · 74%
Raised to accelerating, 74% — gains widened faster than expected.
- Jun 2026 · 80%
Held at 80% — tooling, not just models, now carries the climb.
- Jun 2026 · 72%
Changed to plateauing — June 2026 trackers show the top system at 93.9%, near the ceiling of the public real-bug test.
If you build anything, the boring half of your job is the part getting automated first.
Behind the numbersOpen
Direction inferred from year-over-year results on a public benchmark of real GitHub issues (4.4% → 71.7%, 2024→2025). Projection assumes continued model + tooling gains; a plateau in frontier models would flatten it.