Best coding models:
the independent leaderboard

DeepSWE v1.1 Best, mini-swe-agent harness, public 113-task split, board updated 2026-08-13. Effort is the lab's own setting on that board and is printed rather than normalized. Cost, output tokens, and steps come from the same row as the pass rate, which is why this is the board we lead with: a percentage with no cost attached is not a purchasing decision.

19 rows: every card in this cluster carrying a DeepSWE official chip, and no others. Rails draw the printed pass rate on a fixed 100-point track, so bar length is the score itself and not a distance from the leader. A blank cell means the board did not print that field for that row, not that the value is zero. Board source.

This is not a continuation of the table above. The vals.ai hero index is a separate methodology over a separate task mix, so a model can place well here and poorly on DeepSWE without either number being wrong. Ranks in this table are only comparable to other ranks in this table.

22 rows. vals.ai prints an index percent per card; where it also printed an interval, that interval lives on the model card, not here. Board source.

A third board again, and the least coding-specific of the three: the Intelligence Index is a published integer over nine evaluations, most of which are not software engineering. It is here because it is independent and broadly available, not because it answers the coding question. Artificial Analysis publishes no interval, so equal integers are genuinely not separated and are shown sharing a rank.

123 rows. The integer is the whole chip: per-evaluation percents printed on an AA model page sit in that card's named-benches table and are never averaged into anything. Board source.

Every row on this page comes from a chip that already exists on a Continuum model card, and every chip carries a source URL and a fetch date. Nothing here is computed by us, averaged by us, or estimated. If a board did not publish a number for a model, that model has no row.

What DeepSWE measures. The official DeepSWE board runs mini-swe-agent, a deliberately minimal bash-only agent, against 113 public software-engineering tasks. The reward is binary: the patch either makes the repository's tests pass or it does not. There is no partial credit, no human grader, and no model acting as judge. Because the harness is fixed and published, the board can also report what a run cost, which is where the dollars per task, output tokens, and step count in the first table come from. Effort is the lab's own setting on that board. A row at max and a row at high are not the same experiment, so we print the label instead of hiding it.

Why the intervals matter more than the order. 113 tasks is a small set. The board publishes an interval around each pass rate, and on this data most of the top rows overlap. When two intervals overlap, the board cannot tell you which model is better, only that both sit in the same band. Rather than sell a clean one, two, three, we mark every row that overlaps another and name the ranks it cannot be separated from. Read the bands, not the ordering.

Why we do not blend boards. The three tables above are three different experiments. DeepSWE is a coding-agent pass rate on a fixed harness. The Artificial Analysis Intelligence Index is a published integer over nine evaluations. The vals.ai hero index is a third methodology on a third task mix. Averaging them produces a number nobody measured and nobody can reproduce. We stack the boards. We do not merge them, and we do not offer a combined score.

Why usage rankings are not on this page. OpenRouter This Week ranks models by tokens routed. That is popularity: it tracks price, free tiers, the default model in somebody else's tool, and which model a large customer wired up last month. A cheap model serving a high-volume classification job outranks every frontier coder on it. We publish that board on the models index because it is useful market data, and we keep it off this page because it is not evidence of quality.

Why vendor self-reports are excluded. A lab that picks its own harness, its own effort setting, its own subset, and its own retry policy is not running the same test as the board. Vendor figures appear on individual cards, labeled vendor-reported, and they never enter a ranking here. The same rule removes boards we could not fetch cleanly: Scale SWE-bench Pro and Terminal-Bench 2.1 are omitted cluster-wide rather than borrowed from a scraper.