Coding · sub-raceDominant lead

Frontier SWE race

Very hard, long software-engineering tasks in real codebases, graded by tests.

Warming up the track…

Distance on the track = score gap. Ticks on the inside edge mark 5.0 pts steps behind the leader.

Oct 20, 2026GPT-6 Astra leads

2025Apr ’25Jul ’25Oct ’252026Apr ’26Jul ’26Oct ’26

Score history

Top 6 today over the leader's frontier · click to jump to a week

#1 frontierGPT-6 AstraOpus 5.5Sonnet 5.5Fable 5.1Gemini 4 ArgonHaiku 5.5

Longest: GPT-6 Astra, 8 weeks (from Sep 2026).

Sibling races

Shows which models can carry a large engineering change to completion, not just a single function.

FrontierSWE score from Epoch AI's Benchmarking Hub. One result per model, placed at the model's release date (as measured).

Unit: percentHigher is betterPhoto finish under 0.5 ptsDominant over 3.0 pts

FrontierSWE · Epoch AI

As measured: one result per model, placed at its release date.

19 tracked models with a result (21 entries on the source board) · newest data October 7, 2026.

Partial coverage: the source does not report every tracked model, so treat the order as indicative.