LLM Stats (PostTrainBench Lite): leaderboard

OpenAI's PostTrainBench Lite evaluation (adapted from PostTrainBench), as tracked by LLM Stats: agents get five hours on one H100 GPU to improve a pretrained base model (Qwen3-4B-Base, Qwen3-1.7B-Base or SmolLM3-3B-Base) on AIME 2025, BFCL, GSM8K or HumanEval (12 tasks, four attempts each). Scored in percent as mean share of the performance gap closed toward the developer's instruction-tuned reference (higher is better). LLM Stats combines GPT-6.2 Astra (including three replacement HumanEval attempts) and GPT-6.2 Sol from the GPT-6.2 system card (Table 5) with GPT-5.6 Terra, Sol and Luna from the GPT-5.6 launch post.

Metric: Score (%). Source: llm-stats.com. Saturation forecast: Around April 2027. 5 models tracked. Updated 19 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-6.2 Astra74.79#4
2GPT-5.6 Terra51.5#72
3GPT-5.6 Sol50.3#36
4GPT-6.2 Sol44.89#12
5GPT-5.6 Luna29.6#133

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #5 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=llm-stats-posttrainbench-lite · How It Works · Data refreshed daily, snapshot 2026-11-20.