PostTrainBench
Integrity rank #48 of 65 · 11 models scored · top score 41.79 · Claude Fable 5
Autonomous agentic post-training/fine-tuning skill under a fixed compute budget
Strongest on harness comparability, weakest on contamination resistance. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Fable 5 | 41.79 | unverifiedT2 | 2026-06-09 |
| 2 | GPT-5.6 Sol | 36.23 | unverifiedT2 | 2026-07-09 |
| 3 | Claude Opus 5 | 35.04 | unverifiedT2 | 2026-07-24 |
| 4 | Claude Opus 4.8 | 33.84 | unverifiedT2 | 2026-05-28 |
| 5 | Kimi K3 | 31.96 | unverifiedT2 | 2026-07-16 |
| 6 | GLM-5.2 | 31.7 | unverifiedT2 | 2026-06-16 |
| 7 | Claude Opus 4.7 | 28.56 | unverifiedT2 | 2026-04-16 |
| 8 | GPT-5.5 | 27.23 | unverifiedT2 | 2026-04-23 |
| 9 | Grok 4.5 | 23.45 | unverifiedT2 | 2026-07-08 |
| 10 | Gemini 3.1 Pro | 21.99 | unverifiedT2 | 2026-02-19 |
| 11 | GPT-5.4 | 19 | unverifiedT2 | 2026-03-05 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 11 models, and the gap between its best and worst performers spans 22.79 points. That margin is wide enough to support real separation between models, so the ordering of ranks is meaningful rather than an artifact of near-identical scores.
3 cited facts
This benchmark's Benchmark Integrity Index is 74, with a letter grade of B and a rank of 46 among 62 benchmarks in the universe, describing its health as a discriminator of frontier models under a disclosed harness. The weakest component of its integrity breakdown is contamination, which narrows how much the top scores can be trusted. This score is not a statement of any model's capability.
7 cited facts
This benchmark is not saturated, with a top score of 41.79 and a remaining headroom of 58.21 points to the ceiling. Only one model clusters near the top, so there is no group of leaders within a couple of points. That means the field still has genuine room to separate at the top, unlike a saturated benchmark where ranking differences near the top would be close to noise.
5 cited facts
Because the harness is consistent, scores are directly comparable and can be compared head-to-head; the task remains highly sensitive to the agent scaffold used. The test set is public, so high scores deserve more scrutiny for possible contamination given this benchmark's public nature and overfitting risk, and results should be read as task performance under a disclosed harness rather than deployed capability.
4 cited facts
This benchmark measures autonomous agentic post-training and fine-tuning skill under a fixed compute budget, giving the agent free choice of data and post-training technique within that budget. It was built around several small base models, with the agent handed a single accelerator and a fixed time allotment, and the resulting models evaluated on a set of downstream benchmarks; a human baseline is not documented. The design rests on the assumption that improvement on those fixed downstream benchmarks under a fixed compute budget proxies general post-training competence, and that the agent's chosen methods generalize. The sharpest caveat is that a small-model, single-GPU, limited-time setting may not reflect frontier post-training, that the target benchmarks are public and partly gameable, and that outcomes are heavily scaffold-dependent.
6 cited facts
PostTrainBench: PostTrainBench tests a CLI coding agent's ability to autonomously post-train small base LLMs: given a 1-4B base model, a single H100, and a 10-hour window, the agent must improve the model across 7 downstream benchmarks using post-training techniques of its own choosing. Best submission approx 34.1% average.
Claude Fable 5 leads PostTrainBench at 41.79. The full leaderboard above lists every recorded measurement, not just the headline number.
11 models have recorded scores on PostTrainBench, spanning a score spread of 22.79.
tensor.news grades PostTrainBench B for integrity (score 74/100), ranking #48 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — PostTrainBench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is public, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on PostTrainBench are reasonably apples-to-apples.
Every PostTrainBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites posttrainbench.com.