PostTrainBench v1.2 (PostTrainBench)
We show this table for reference; we do not rank on it.
An AI agent gets ten hours and one H100 GPU to post-train a small base model; the score is how well the trained models do on six benchmarks.
Benchmark score on PostTrainBench — November 13, 2026
We compile the PostTrainBench rows from benchmark-owner or independent runs and provider self-reports. Claude Opus 5.5 leads the table at 43.8%, followed by GPT-6 Astra (41.9%) and Gemini 4 Carbon (41.0%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
Run by PostTrainBenchGPT-6 Astra
OpenAI
Run by PostTrainBenchGemini 4 Carbon
21 modelsCodingCurrentDisplay onlyUpdated November 13, 2026
Current releasePostTrainBench v1.2 · Released · 4 base models × 6 benchmarks · Official leaderboard · Not comparable with v1.1 or v1
Benchmark score table (21 models)
Of 21 results: 18 benchmark-run, 3 lab-reported.
Ranks on this table count the 18 benchmark-run and independent results. A lab-reported result keeps its place by score, without a rank.
- Lab-reported
- 3
- Independent
- 0
- Benchmark-run
- 18
Showing all 21 results, in score order.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5AnthropicRun by PostTrainBench | 43.8% | Run by PostTrainBench | Not reported | Closed |
| 2 | GPT-6 AstraOpenAIRun by PostTrainBench | 41.9% | Run by PostTrainBench | Not reported | Closed |
| 3 | Gemini 4 CarbonGoogleRun by PostTrainBench | 41.0% | Run by PostTrainBench | Not reported | Closed |
| 4 | Gemini 4 ArgonGoogleRun by PostTrainBench | 40.4% | Run by PostTrainBench | Not reported | Closed |
| 5 | GPT-6.1 SolOpenAIRun by PostTrainBench | 38.8% | Run by PostTrainBench | Not reported | Closed |
| 6 | GLM-5.3-FlashZ.AIRun by PostTrainBench | 36.8% | Run by PostTrainBench | Not reported | Open |
| No rank on this table | Kimi K3.1 ProMoonshot AIReported by Moonshot AI | 36.5% | Reported by Moonshot AI | Not reported | Open |
| 7 | Claude Opus 5AnthropicRun by PostTrainBench | 35.7% | Run by PostTrainBench | Not reported | Closed |
| No rank on this table | Hy4TencentReported by Tencent | 35.2% | Reported by Tencent | Not reported | Pending |
| 8 | GLM-5.3Z.AIRun by PostTrainBench | 34.5% | Run by PostTrainBench | Not reported | Open |
| No rank on this table | Inkling 2 SmallThinking Machines LabReported by Thinking Machines Lab | 31.9% | Reported by Thinking Machines Lab | Not reported | Closed |
| 9 | Kimi K3Moonshot AIRun by PostTrainBench | 31.5% | Run by PostTrainBench | Not reported | Open |
| 10 | GPT-5.6 SolOpenAIRun by PostTrainBench | 31.0% | Run by PostTrainBench | Not reported | Closed |
| 11 | Claude Opus 4.8AnthropicRun by PostTrainBench | 31.0% | Run by PostTrainBench | Not reported | Closed |
| 12 | GLM-5.2Z.AIRun by PostTrainBench | 30.4% | Run by PostTrainBench | Not reported | Open |
| 13 | Claude Opus 4.7AnthropicRun by PostTrainBench | 24.3% | Run by PostTrainBench | Not reported | Closed |
| 14 | Muse WonderMetaRun by PostTrainBench | 24.2% | Run by PostTrainBench | Not reported | Closed |
| 15 | GPT-5.5OpenAIRun by PostTrainBench | 23.9% | Run by PostTrainBench | Not reported | Closed |
| 16 | Grok 4.5xAIRun by PostTrainBench | 22.6% | Run by PostTrainBench | Not reported | Closed |
| 17 | Gemini 3.1 ProGoogleRun by PostTrainBench | 18.1% | Run by PostTrainBench | Not reported | Closed |
| 18 | GPT-5.4OpenAIRun by PostTrainBench | 17.8% | Run by PostTrainBench | Not reported | Closed |
A dash in the rank column marks a lab-reported result. It keeps its place by score, without a rank on this table.
- Lab-reported
- Benchmark-run
- Independent
PostTrainBench (lab runs)
Moonshot (Kimi K3, official Harbor implementation with Claude Code), Z.AI (GLM-5.3, three Claude Code runs) and Tencent (Hy4 Preview) report these figures in launch posts. They are not rows of an official PostTrainBench board, and the version is not stated, so BenchLM keeps them apart from the v1, v1.1 and v1.2 keys.
| Source rank | Source model or configuration | Score | Who ran it | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|---|
| 1 | GLM-5.3Z.AI | 39.8% | Reported by Z.AI | Model makers' launch postsOriginal result | Not reported | Open |
| 2 | Kimi K3Moonshot AI | 36.6% | Reported by Moonshot AI | Model makers' launch postsOriginal result | Not reported | Open |
| 3 | Hy4 previewTencent | 35.6% | Reported by Tencent | Model makers' launch postsOriginal result | Not reported | Open |
PostTrainBench version history
Every PostTrainBench version, newest first. Each version changes the rules, the audits or the benchmark set, and older runs were rescored or removed, so scores are not comparable across versions; ranks are the fairer way to follow a model. The lab-run table holds launch-post results that name no version.
| Version | Released | Setup | What changed | Leader | Models |
|---|---|---|---|---|---|
| PostTrainBench v1.2Current | Sep 30, 2026 | 4 base models × 6 benchmarks | Drops BFCL and reweights the other six benchmarks, fixes HumanEval scoring, stops evaluation scripts from running the submitted model’s code, averages final evaluations over up to five seeds, and decides contamination by a majority of three judges. | LeaderClaude Opus 5.5 43.8Official board | 18 models |
| PostTrainBench v1.1 | Jul 28, 2026 | 4 base models × 7 benchmarks | Same budget, base models and benchmarks as v1, with stricter rules and audits for test-set contamination, external API use, benchmark lookup and submitting a different model. Older runs were rerun or removed. | LeaderGPT-5.6 Sol 36.2Official board | 10 models |
| PostTrainBench v1 | Feb 22, 2026 | 4 base models × 7 benchmarks | The first seven-benchmark board (AIME 2025, Arena Hard Writing, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval). The source keeps it as an archived version. | LeaderGLM-5.2 34.3Official board | 13 models |
Rankings across PostTrainBench versions
Each cell is the model’s rank among catalog models on that version’s official board, using its highest published effort, with its weighted score below. Fable rows that borrow another model’s scores for refused runs, and the Locus agent system, are on the full board but not ranked here. A dash means the model has no row on that version.
| Model | PostTrainBench v1.218 models | PostTrainBench v1.110 models | PostTrainBench v113 models |
|---|---|---|---|
| Claude Opus 5.5 | #143.8 | Not on this board | Not on this board |
| GPT-6 Astra | #241.9 | Not on this board | Not on this board |
| Gemini 4 Carbon | #341.0 | Not on this board | Not on this board |
| Gemini 4 Argon | #440.4 | Not on this board | Not on this board |
| GPT-6.1 Sol | #538.8 | Not on this board | Not on this board |
| GLM-5.3-Flash | #636.8 | Not on this board | Not on this board |
| Claude Opus 5 | #735.7 | #235.0 | Not on this board |
| GLM-5.3 | #834.5 | Not on this board | Not on this board |
| Kimi K3 | #931.5 | #432.0 | Not on this board |
| GPT-5.6 Sol | #1031.0 | #136.2 | Not on this board |
| Claude Opus 4.8 | #1131 | #332.9 | #234.1 |
| GLM-5.2 | #1230.4 | #531.7 | #134.3 |
| Claude Opus 4.7 | #1324.3 | #628.6 | #328.6 |
| Muse Wonder | #1424.2 | Not on this board | Not on this board |
| GPT-5.5 | #1523.9 | #727.2 | #425.0 |
| Grok 4.5 | #1622.6 | #823.4 | Not on this board |
| Gemini 3.1 Pro | #1718.1 | #922.0 | #621.6 |
| GPT-5.4 | #1817.8 | #1019 | #820.2 |
| Claude Opus 4.6 | Not on this board | Not on this board | #523.2 |
| GPT-5.2 | Not on this board | Not on this board | #721.4 |
| GPT-5.1-Codex-Max | Not on this board | Not on this board | #919.7 |
| Gemini 3 Pro | Not on this board | Not on this board | #1018.1 |
| GPT-5.3 Codex | Not on this board | Not on this board | #1117.8 |
| GPT-5.2-Codex | Not on this board | Not on this board | #1217.2 |
| Claude Opus 4.5 | Not on this board | Not on this board | #1317.1 |
Previous releaseReleased · 4 base models × 7 benchmarks · Official leaderboard · Not comparable with v1.2
PostTrainBench v1.1
Version 1.1 audits contamination, external API use, prior-run lookup, and model identity. Flagged runs receive the base-model score. Native CLI leaderboard runs and Google OpenCode runs use different harnesses; each row records its source and setup.
| Source rank | Source model or configuration | Score | Who ran it | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|---|
| No rank on this table | Claude Opus 5.5Anthropic | 49.3% | Lab-reported, from another publisher’s document | PostTrainBench team · Original result | Not reported | Closed |
| No rank on this table | Gemini 4 ArgonGoogle | 45.3% | Reported by Google | PostTrainBench team · Original result | Not reported | Closed |
| No rank on this table | GPT-6 AstraOpenAI | 44.3% | Lab-reported, from another publisher’s document | PostTrainBench team · Original result | Not reported | Closed |
| No rank on this table | Claude Fable 5.1Anthropic | 40.2% | Lab-reported, from another publisher’s document | PostTrainBench team · Original result | Not reported | Closed |
| No rank on this table | MiniMax M3.1 FlashMiniMax | 39.3% | Reported by MiniMax | PostTrainBench team · Original result | Not reported | Closed |
| 1 | GPT-5.6 SolOpenAI | 36.2% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 2 | Claude Opus 5Anthropic | 35.0% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 3 | Claude Opus 4.8Anthropic | 32.9% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 4 | Kimi K3Moonshot AI | 32.0% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Open |
| 5 | GLM-5.2Z.AI | 31.7% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Open |
| 6 | Claude Opus 4.7Anthropic | 28.6% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 7 | GPT-5.5OpenAI | 27.2% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 8 | Grok 4.5xAI | 23.4% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 9 | Gemini 3.1 ProGoogle | 22.0% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
| 10 | GPT-5.4OpenAI | 19.0% | Run by PostTrainBench v1.1 | PostTrainBench team · Original result | Not reported | Closed |
Older releaseReleased · 4 base models × 7 benchmarks · Official leaderboard · Not comparable with v1.1 or v1.2
PostTrainBench v1
Version 1 (published February 22, 2026, superseded by v1.1 on July 28, 2026) scores agents that post-train Qwen3 1.7B and 4B, SmolLM3 3B and Gemma 3 4B in ten hours on one H100, across AIME 2025, Arena Hard Writing, BFCL, GPQA Main, GSM8K, HealthBench and HumanEval. v1.1 reran or removed runs under stricter contamination and API rules, so v1 scores are not comparable with later versions.
| Source rank | Source model or configuration | Score | Who ran it | Source setup | Parameters (B) | Open / closed |
|---|---|---|---|---|---|---|
| 1 | GLM-5.2Z.AI | 34.3% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Open |
| 2 | Claude Opus 4.8Anthropic | 34.1% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 3 | Claude Opus 4.7Anthropic | 28.6% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 4 | GPT-5.5OpenAI | 25.0% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 5 | Claude Opus 4.6Anthropic | 23.2% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 6 | Gemini 3.1 ProGoogle | 21.6% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 7 | GPT-5.2OpenAI | 21.4% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 8 | GPT-5.4OpenAI | 20.2% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 9 | GPT-5.1-Codex-MaxOpenAI | 19.7% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 10 | Gemini 3 ProGoogle | 18.1% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 11 | GPT-5.3 CodexOpenAI | 17.8% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 12 | GPT-5.2-CodexOpenAI | 17.2% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
| 13 | Claude Opus 4.5Anthropic | 17.1% | Run by PostTrainBench v1 | Ben Rank, Hardik Bhatnagar, Matthias Bethge, Maksym Andriushchenko and the PostTrainBench team · Original result | Not reported | Closed |
Among the reported PostTrainBench rows, Claude Opus 5.5 is first at 43.8%. The third row is 2.8 points behind. The broader top-10 range is 9.3 points, so many of the published results sit in a relatively narrow band.
21 models have been evaluated on PostTrainBench. The benchmark falls in the Coding category. PostTrainBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About PostTrainBench
Year
2026
Tasks
Post-training four base models on six benchmarks
Format
Weighted aggregate across four base models and six benchmarks
Difficulty
Autonomous ML engineering
Version 1.2 (September 30, 2026) drops BFCL, reweights AIME 2025, Arena Hard Writing, GPQA Main, GSM8K, HealthBench and HumanEval, fixes HumanEval scoring, averages final evaluations over up to five seeds and decides contamination by a majority of three judges. Each agent post-trains Qwen3 1.7B and 4B, SmolLM3 3B and Gemma 3 4B; runs flagged for contamination, external API use or a swapped model receive the base-model score. BenchLM stores each catalog model's official aggregate at its highest published setting, plus lab-reported v1.2 runs, which name their own setup.
Freshness and provenance
Version
PostTrainBench v1.2
Refresh cadence
Rolling
Staleness state
Current
Question availability
Public tasks, code and run traces
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does PostTrainBench measure?
An AI agent gets ten hours and one H100 GPU to post-train a small base model; the score is how well the trained models do on six benchmarks.
Which model scores highest on PostTrainBench?
Claude Opus 5.5 by Anthropic currently leads with a score of 43.8% on PostTrainBench.
How many models are evaluated on PostTrainBench?
21 AI models have published results on PostTrainBench in the BenchLM catalog.
Compare top models on PostTrainBench
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.