The 32 Lean projects of LeanLean October 2026, and what 5 models made of them. Each agent had 12 hours per project to make the codebase as small as it could, in Lean tokens, while the project still builds, uses only the permitted axioms and keeps every protected theorem's statement.
P01-DedekindZetaZeroFree/
problem.json benchmark entry: upstream repository and commit, license, protected theorems, ...
base/ baseline source tree
base-graph.json declaration dependency graph of base/
opus-55-high/
repo/ submitted source tree
diff.patch git diff from base/ to repo/
graph.json declaration dependency graph of repo/
result.json status, tokens, compression, cost and time
...
Problems are numbered largest first by baseline tokens. Source trees leave out hidden paths (.lake, .gitignore, agent scratch files), as in scoring. Graphs are created by lean-strip.
A failed submission is scored at its latest saved checkpoint that passes final verification, and repo/ holds that checkpoint (result.json has a verification_note). A failed status means no checkpoint passed: the submission gets no credit and has no graph.json.
Each problem has a baseline commit, which adds the problem and seeds every model's repo/ with a copy of base/. One commit per model follows, turning that copy into the submission, so the model's change is exactly that commit's diff:
git log --oneline
git show <commit> -- P01-DedekindZetaZeroFree/opus-55-high/repo
git log -p --reverse -- 'P*/kimi-k31-pro/repo'diff.patch holds the same diff with paths relative to the project root, so git apply diff.patch in a copy of base/ reproduces repo/ (outside any git checkout, where git apply would skip paths outside the current directory).
| Folder | Model | Harness |
|---|---|---|
gpt-61-sol-max |
GPT-6.1 Sol (max) | Codex 0.159.0 |
opus-55-high |
Claude-Opus-5.5 (high) | Claude Code 2.1.280 |
gemini-4-argon |
Gemini 4 Argon | Antigravity 1.3.5 |
kimi-k31-pro |
Kimi K3.1 Pro | Kimi Code 2.3.1 |
deepseek-v41-flash-dsh |
DeepSeek-V4.1-Flash-1017 (Max) | DeepSeek Harness 0.2.1-rc.5 |
Credited compression in percent (token reduction against the baseline, zero for failed submissions).
| Problem | Baseline tokens | gpt-61-sol-max |
opus-55-high |
gemini-4-argon |
kimi-k31-pro |
deepseek-v41-flash-dsh |
|---|---|---|---|---|---|---|
| P01-DedekindZetaZeroFree | 10,053,215 | 13.3 | 13.9 | 0.0 | 7.0 | 4.7 |
| P02-zeta-density-39-over-17-lean | 8,602,200 | 7.1 | 9.8 | 7.3 | 0.0 (failed) | 0.8 |
| P03-erdos-408 | 6,268,302 | 10.1 | 4.9 | 0.0 | 0.9 | 1.1 |
| P04-LinearTimeTriangulation-ChazelleAlgorithm-Lean4 | 4,338,771 | 12.8 | 27.3 | 1.6 | 1.7 | 1.9 |
| P05-DOZZ | 2,992,942 | 15.7 | 12.8 | 11.8 | 0.0 (failed) | 1.2 |
| P06-hypoelliptic-aleksandrov | 2,642,100 | 17.3 | 14.3 | 10.3 | 2.3 | 2.1 |
| P07-PottsDiscontinuity-lean | 1,054,327 | 24.2 | 28.7 | 16.6 | 1.9 | 0.4 |
| P08-vaught-conjecture-palomar | 1,043,844 | 40.5 | 32.4 | 35.8 | 28.7 | 21.8 |
| P09-whitehead | 975,198 | 20.4 | 25.6 | 13.4 | 4.1 | 2.9 |
| P10-polya-neumann-simply-connected-formalization-Palomar-version-v2 | 903,891 | 22.0 | 17.5 | 20.3 | 2.8 | 0.0 |
| P11-ccv-palomar | 854,028 | 12.5 | 39.0 | 16.3 | 1.9 | 3.3 |
| P12-ImprovedExponents | 672,860 | 27.5 | 18.5 | 18.0 | 8.0 | 3.8 |
| P13-MIPStarRE-QPBT | 597,369 | 12.0 | 24.8 | 15.2 | 3.4 | 3.6 |
| P14-ccdn-dimension-growth | 552,362 | 19.2 | 24.2 | 25.9 | 2.3 | 2.3 |
| P15-gs_Lean_formalization | 303,700 | 68.2 | 65.3 | 56.2 | 23.6 | 22.7 |
| P16-compact-albert-recognition | 300,239 | 64.6 | 30.2 | 43.7 | 5.8 | 4.0 |
| P17-lean4-adr11 | 175,441 | 85.0 | 59.7 | 61.7 | 7.5 | 13.9 |
| P18-lean-dual-analytic-ifs | 146,557 | 66.6 | 44.8 | 46.3 | 3.5 | 6.0 |
| P19-omega_le_nine_quarters_all_rings | 145,127 | 77.0 | 69.6 | 52.3 | 10.1 | 1.2 |
| P20-palomar-p046-critical-window-shape-law | 101,498 | 59.9 | 48.9 | 33.0 | 3.0 | 5.1 |
| P21-radian-ramsey | 81,623 | 45.8 | 64.3 | 41.2 | 5.3 | 3.9 |
| P22-linniks-constant-3.99 | 80,635 | 59.9 | 53.0 | 36.7 | 7.2 | 6.7 |
| P23-cell-attachment-lean | 80,254 | 93.4 | 93.2 | 86.6 | 9.9 | 5.0 |
| P24-bergman-weak-type | 77,810 | 67.2 | 63.0 | 41.5 | 15.9 | 14.3 |
| P25-normalizer-lean | 34,716 | 82.3 | 73.4 | 65.6 | 20.9 | 17.2 |
| P26-hk-heterogeneous-lean | 27,736 | 78.1 | 63.8 | 68.8 | 30.5 | 28.2 |
| P27-triangle-group-growth | 25,916 | 83.0 | 76.4 | 60.3 | 25.2 | 14.6 |
| P28-semigroup-betti | 19,921 | 87.0 | 81.9 | 75.6 | 16.8 | 15.7 |
| P29-goldbach-lean | 17,726 | 90.2 | 83.4 | 74.8 | 26.9 | 21.5 |
| P30-alcs | 15,133 | 80.2 | 79.9 | 60.5 | 25.7 | 16.5 |
| P31-kkp-connected-spectral-counterexamples | 14,983 | 86.0 | 87.2 | 84.2 | 46.7 | 47.7 |
| P32-galvin-dinitz | 10,465 | 85.4 | 83.0 | 69.6 | 30.1 | 18.8 |
| Mean | 50.5 | 47.3 | 39.1 | 11.9 | 9.8 |
Each project keeps its upstream license (Apache-2.0, MIT or AGPL-3.0; see license in problem.json and the license file in base/). Please abide by the license of each project when using its code.