Anthropic

Claude Opus 5

Verified Model id claude-opus-5, run at each effort level it offers by us and by people we invited. Runs of the same effort are averaged.

Best Fast35.0at max effort
Best Thorough–not run yet
Suites run1
Our notes · 16 Oct 2026

Claude Opus 5 hit output limits on both 3×3 ladders

Claude Opus 5 sits 6th on the Index, clearing the 2×2 ladder but ending both 3×3 tests without an answer at level 6.

Where it stands

We ran Opus 5 at max, its highest effort setting. It scores 35.0 and sits 6th of 16 models on the Index. Just above it is GPT-6.1-Sol at max, with 44.0. Both cleared the 2×2 ladder and reached level 5 on the interactive 3×3, where the model turns the cube and sees each result. The difference was one-shot: GPT-6.1-Sol reached level 11, while Opus 5 stopped at level 5. Opus 5 did not unlock the 4×4 Gauntlet.

What stood out

The interactive 3×3 ended on a cube 6 moves from solved. Opus 5 never answered it. Over about 96 min, 4 consecutive replies hit the model’s output limit, consuming about 511,700 thinking tokens. Thinking counts toward that limit, so a reply can run out before it sends any moves. Claude Code let it continue after hitting the limit, but repeated overruns ended the test. It kept its 5 completed levels, with all 3 lives unused. The unanswered cube accounted for roughly three-quarters of the round’s time.

The 3×3 cube at level 6, as the model got it
Level 6, as the model got it. Replay

The one-shot test, where each cube gets one complete answer without a look at the result, ended the same way. At level 6, another 4 replies exhausted the output limit without an answer: about 511,700 thinking tokens over roughly 95 min. Again, all 3 lives remained. Across the two unanswered cubes, Opus 5 spent about 3 h 11 min without adding a completed level, nearly half the suite’s total time.

The 3×3 cube at level 6, as the model got it
Level 6, as the model got it. Replay

The run

Our only suite was at max. The 2×2 ladder finished with all 11 levels solved, no misses and 117 moves in total. It used 601,304 thinking tokens over 2 h 38 min to clear those cubes. On each 3×3 ladder, the first 5 solves used just 15 moves altogether, matching the combined scramble lengths. Interactive play went no further than one-shot. Neither 3×3 test ended through a wrong submitted solution or an exhausted move allowance: both stopped before Opus 5 submitted any moves for the sixth cube.

Time and cost

The suite took 6 h 43 min and cost $70.57 at API-equivalent rates. With no lower-effort runs, we have no cheaper Opus 5 result to compare. The completed 2×2 ladder cost $30.31; the interactive and one-shot 3×3 rounds cost $20.72 and $19.54 respectively.

Fast v3 by effort

Three ladders: 2×2 and 3×3 interactive, and 3×3 one-shot. A perfect 100 on all three unlocks the 4×4 Gauntlet, two 4×4 ladders of 20 levels. No time limit; each ends at the top level or on the third miss.

EffortIndex2×2 ladder, interactive3×3 ladder, interactive3×3 ladder, one-shot4×4 Gauntlet, interactive4×4 Gauntlet, one-shotMovesTimeTokensCostRuns
max35.010025.0†25.0†––1476h 43m7.6M$711

† Ran out of output: one reply hit the model's output limit before it answered, so the test ended on that level. It keeps the levels it had cleared. Why

Every run (1)

What it took

What the run cost, split into what the tokens were for: thinking, answers (moves and tool calls), fresh input, and input read back from cache. Sessions its API cut off are left out. For the whole suite.

  • Thinking
  • Answers
  • Input
  • Cached input

Split at the model’s list prices, scaled to what the run cost. Many runs themselves used subscriptions.

Level by level

The moves each solve took at each level of a ladder. Level N is scrambled N moves from solved, so the diagonal is the scramble undone move by move. A line that jumps off it has switched to a method of its own: longer, but it still works when the scramble is too deep to see.

Compare with other models · The Index · All boards