Benchmark Almanac

PostTrainBench

Claude Code + GLM-5.2 34.3% tops its leaderboard Claude Opus 4.8 34.1% leads as a model on its own
GPT-6.2 Astra 74.8% leads since November 2026
In use
Version 1.0

The official leaderboard hasn’t scored a model released after Claude Fable 5 (June 2026). Labs report their own numbers for 8 newer models, such as GPT-6.2 Astra at 74.8%.

CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain.

Known reward hacking (eval tampering, swapping in instruct models). Agent-as-judge added. Dataset link holds trajectories, not tasks.

Released
March 2026
Kind
Machine learning engineering
Level
Research frontier
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Code, Actions
Interaction
Agentic
Environment
Terminal
Run by
Anyone
In views
Agents
Strongest model tested
GPT-6.2 Astra #1 on the capability index
Last in a launch post
October 2026
Last model added
November 2026
Scores refreshed
November 20th, 2026

In the authors’ words

AI agents have become surprisingly proficient at software engineering over the past year, largely due to improvements in reasoning capabilities. This raises a deeper question: can these systems extend their capabilities to automate AI research itself? In this paper, we explore post-training, the critical phase that turns base LLMs into useful assistants. We introduce PostTrainBench to benchmark how well LLM agents can perform post-training autonomously under bounded compute constraints (10 hours on one H100 GPU). We ask frontier agents (e.g., Claude Code with Opus 4.6) to optimize the performance of a base LLM on a particular benchmark (e.g., Qwen3-4B on AIME). Importantly, we do not provide any predefined strategies to the agents and instead give them full autonomy to find necessary information on the web, run experiments, and curate data. We find that frontier agents make substantial progress but generally lag behind instruction-tuned LLMs from leading providers: 23.2% for the best agent vs. 51.1% for official instruction-tuned models. However, agents can exceed instruction-tuned models in targeted scenarios: GPT-5.1 Codex Max achieves 89% on BFCL with Gemma-3-4B vs. 67% for the official model. We also observe several failure modes worth flagging. Agents sometimes engage in reward hacking: training on the test set, downloading existing instruction-tuned checkpoints instead of training their own, and using API keys they find to generate synthetic data without authorization. These behaviors are concerning and highlight the importance of careful sandboxing as these systems become more capable. Overall, we hope PostTrainBench will be useful for tracking progress in AI R&D automation and for studying the risks that come with it. Website and code are available at https://posttrainbench.com/.

Abstract of PostTrainBench: Can LLM Agents Automate LLM Post-Training? (arXiv:2603.08640, CC0)
26
Entries on its own leaderboard, 14 models we track
5
Launch posts list it, from 4 labs
13
System cards report it, from 5 labs

Claude Code + GLM-5.2 tops the leaderboard with 34.3% Claude Opus 4.8 leads with 34.1% among models released by June 2026

Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran on its own with several settings, we keep its best. Every run is in the next chart. Agents other companies built around a model get a pale shade of that model's lab color. Gray marks blends and entries not in our model list, such as fine-tunes and older models. None of these count toward leaders. The best model on its own is Claude Opus 4.8, at 34.1%. Agents built around a model, such as Claude Code + GLM-5.2, are their own entries. We don't count them as the model's score. This view shows only models we track, each at its best.

GPT-6.2 Astra leads with 74.8%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model. Striped bars are claims, and dotted bars are reruns by Epoch AI or another board.

The newest model with a score here came out in June 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Claude Code + GLM-5.2 max 34.3%
  2. 2 Claude Opus 4.8 max 34.1%
  3. 3 Claude Fable 5 max 30.7%
  4. 4 Claude Opus 4.7 xhigh 28.6%
  5. 5 GPT-5.5 xhigh, reprompted 28.4%
  6. 6 GPT-5.4 high, reprompted 28.2%
  7. 7 Claude Opus 4.6 1M context 24.8%
  8. 8 OpenCode + Gemini 3.1 Pro 21.6%
  9. 9 GPT-5.2 21.4%
  10. 10 GPT-5.1-Codex-Max 19.7%
  11. 11 Gemini 3 Pro 18.1%
  12. 12 GPT-5.3 Codex high 17.8%
  13. 13 OpenCode + Claude Opus 4.5 17.3%
  14. 14 GPT-5.2 Codex 17.2%
  15. 15 Claude Opus 4.5 17.1%
  16. 16 Claude Sonnet 4.6 16.4%
  17. 17 OpenCode + Gemini 3 Pro 14.9%
  18. 18 OpenCode + GLM-5 13.9%
  19. 19 OpenCode + Kimi K2.5 10.3%
  20. 20 Claude Sonnet 4.5 9.94%
  21. 21 OpenCode + MiniMax-M2.5 9.50%
  22. 22 OpenCode + MiniMax-M2.1 9.33%
  23. 23 OpenCode + GPT-5.1-Codex-Max 7.65%
  24. 24 OpenCode + GLM-4.7 7.48%
  25. 25 Claude Code + Qwen3-Max 7.42%
  26. 26 OpenCode + Kimi K2 Thinking 7.25%

Source: Official PostTrainBench leaderboard

#EntryScoreSettingDateSource
1 Claude Code + GLM-5.2 34.3% max Jun 2026 Official leaderboard
2 Claude Opus 4.8 34.1% max, Claude Code May 2026 Official leaderboard
3 Claude Fable 5 30.7% max, Claude Code Jun 2026 Official leaderboard
4 Claude Opus 4.7 28.6% xhigh, Claude Code Apr 2026 Official leaderboard
5 GPT-5.5 28.4% xhigh, reprompted, Codex CLI Apr 2026 Official leaderboard
6 GPT-5.4 28.2% high, reprompted, Codex CLI Mar 2026 Official leaderboard
7 Claude Opus 4.6 24.8% 1M context, Claude Code Feb 2026 Official leaderboard
8 OpenCode + Gemini 3.1 Pro 21.6% Feb 2026 Official leaderboard
9 GPT-5.2 21.4% Codex CLI Dec 2025 Official leaderboard
10 GPT-5.1-Codex-Max 19.7% Codex CLI Nov 2025 Official leaderboard
11 Gemini 3 Pro 18.1% Gemini CLI Nov 2025 Official leaderboard
12 GPT-5.3 Codex 17.8% high, Codex CLI Feb 2026 Official leaderboard
13 OpenCode + Claude Opus 4.5 17.3% Nov 2025 Official leaderboard
14 GPT-5.2 Codex 17.2% Codex CLI Dec 2025 Official leaderboard
15 Claude Opus 4.5 17.1% Claude Code Nov 2025 Official leaderboard
16 Claude Sonnet 4.6 16.4% Claude Code Feb 2026 Official leaderboard
17 OpenCode + Gemini 3 Pro 14.9% Nov 2025 Official leaderboard
18 OpenCode + GLM-5 13.9% Feb 2026 Official leaderboard
19 OpenCode + Kimi K2.5 10.3% Jan 2026 Official leaderboard
20 Claude Sonnet 4.5 9.94% Claude Code Sep 2025 Official leaderboard
21 OpenCode + MiniMax-M2.5 9.50% Feb 2026 Official leaderboard
22 OpenCode + MiniMax-M2.1 9.33% Dec 2025 Official leaderboard
23 OpenCode + GPT-5.1-Codex-Max 7.65% Nov 2025 Official leaderboard
24 OpenCode + GLM-4.7 7.48% Dec 2025 Official leaderboard
25 Claude Code + Qwen3-Max 7.42% Sep 2025 Official leaderboard
26 OpenCode + Kimi K2 Thinking 7.25% Nov 2025 Official leaderboard
#ModelScoreSettingDateSource
1 Claude Opus 4.8 34.1% max, Claude Code May 2026 Official leaderboard
2 Claude Fable 5 30.7% max, Claude Code Jun 2026 Official leaderboard
3 Claude Opus 4.7 28.6% xhigh, Claude Code Apr 2026 Official leaderboard
4 GPT-5.5 28.4% xhigh, reprompted, Codex CLI Apr 2026 Official leaderboard
5 GPT-5.4 28.2% high, reprompted, Codex CLI Mar 2026 Official leaderboard
6 Claude Opus 4.6 24.8% 1M context, Claude Code Feb 2026 Official leaderboard
7 GPT-5.2 21.4% Codex CLI Dec 2025 Official leaderboard
8 GPT-5.1-Codex-Max 19.7% Codex CLI Nov 2025 Official leaderboard
9 Gemini 3 Pro 18.1% Gemini CLI Nov 2025 Official leaderboard
10 GPT-5.3 Codex 17.8% high, Codex CLI Feb 2026 Official leaderboard
11 GPT-5.2 Codex 17.2% Codex CLI Dec 2025 Official leaderboard
12 Claude Opus 4.5 17.1% Claude Code Nov 2025 Official leaderboard
13 Claude Sonnet 4.6 16.4% Claude Code Feb 2026 Official leaderboard
14 Claude Sonnet 4.5 9.94% Claude Code Sep 2025 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-6.2 Astra 74.8% Nov 2026 Lab model card
2 Gemini 4 Argon 45.3% Oct 2026 Lab model card
3 GPT-6.2 Sol 44.9% Nov 2026 Lab model card
4 Muse Spark 1.3 43.6% Oct 2026 Lab model card
5 GLM-5.3 39.8% Aug 2026 Lab model card
6 MiniMax-M3 37.1% Jun 2026 Lab model card
7 Kimi K3 36.6% Jul 2026 Lab model card
8 GLM-5.2 34.3% Jun 2026 Lab model card
9 Claude Opus 4.8 34.1% max, Claude Code May 2026 Official leaderboard
10 Claude Fable 5 30.7% max, Claude Code Jun 2026 Official leaderboard
11 Claude Opus 4.7 28.6% xhigh, Claude Code Apr 2026 Official leaderboard
12 GPT-5.5 28.4% xhigh, reprompted, Codex CLI Apr 2026 Official leaderboard
13 GPT-5.4 28.2% high, reprompted, Codex CLI Mar 2026 Official leaderboard
14 Claude Opus 4.6 24.8% 1M context, Claude Code Feb 2026 Official leaderboard
15 GPT-5.2 21.4% Codex CLI Dec 2025 Official leaderboard
16 GPT-5.1-Codex-Max 19.7% Codex CLI Nov 2025 Official leaderboard
17 Gemini 3 Pro 18.1% Gemini CLI Nov 2025 Official leaderboard
18 GPT-5.3 Codex 17.8% high, Codex CLI Feb 2026 Official leaderboard
19 GPT-5.2 Codex 17.2% Codex CLI Dec 2025 Official leaderboard
20 Claude Opus 4.5 17.1% Claude Code Nov 2025 Official leaderboard
21 Seed2.1 Pro 16.5% Jun 2026 Lab model card
22 Claude Sonnet 4.6 16.4% Claude Code Feb 2026 Official leaderboard
23 Claude Sonnet 4.5 9.94% Claude Code Sep 2025 Official leaderboard

Labs report scores for 9 models in their cards. Switch the scores above to rank them together.

Performance vs. runtime

Each point is one agent: the average time it spent post-training (of a 10-hour budget) against its average benchmark performance, with the Pareto frontier. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0% 10% 20% 30% 40% 0246810 Average time spent (h), lower is better Average benchmark performance Claude Code + GLM-5.2 (max)Claude Opus 4.7 (xhigh)GPT-5.5 (xhigh)OpenCode + Gemini 3.1 ProGPT-5.4 (high)GPT-5.3 Codex (high)GPT-5.3 Codex (med)OpenCode + GPT-5.1-Codex-Max 0% 10% 20% 30% 40% 0246810 Average time spent (h), lower is better Average benchmark performance Claude Code + GLM-5.2 (max)Claude Opus 4.7 (xhigh)GPT-5.5 (xhigh)OpenCode + Gemini 3.1 ProGPT-5.4 (high)GPT-5.3 Codex (high)GPT-5.3 Codex (med)OpenCode + GPT-5.1-Codex-Max
ModelSettingAverage benchmark performanceAverage time spent
Claude Code + GLM-5.2 max, Claude Code 34.3% 8.35 h
Claude Opus 4.8 max, Claude Code 34.1% 8.93 h
Claude Opus 4.8 high, Claude Code 33.8% 9.27 h
Claude Opus 4.7 xhigh, Claude Code 28.6% 7.57 h
GPT-5.5 xhigh, reprompted, Codex CLI 28.4% 9.24 h
GPT-5.4 high, reprompted, Codex CLI 28.2% 8.32 h
GPT-5.5 xhigh, Codex CLI 25% 4.98 h
Claude Opus 4.6 1M context, Claude Code 24.8% 8.8 h
Claude Opus 4.6 Claude Code 23.2% 9.66 h
OpenCode + Gemini 3.1 Pro OpenCode 21.6% 4.05 h
GPT-5.2 Codex CLI 21.4% 6.08 h
GPT-5.4 high, Codex CLI 20.2% 1.77 h
GPT-5.1-Codex-Max Codex CLI 19.7% 4.05 h
Gemini 3 Pro Gemini CLI 18.1% 6.6 h
GPT-5.3 Codex high, Codex CLI 17.8% 1.65 h
OpenCode + Claude Opus 4.5 OpenCode 17.3% 6.88 h
GPT-5.2 Codex Codex CLI 17.2% 2.43 h
Claude Opus 4.5 Claude Code 17.1% 7.88 h
Claude Sonnet 4.6 Claude Code 16.4% 6.83 h
OpenCode + Gemini 3 Pro OpenCode 14.9% 6.06 h
OpenCode + GLM-5 OpenCode 13.9% 3.56 h
GPT-5.3 Codex med, Codex CLI 13.8% 0.89 h
OpenCode + Kimi K2.5 OpenCode 10.3% 2.55 h
OpenCode + MiniMax-M2.5 OpenCode 9.5% 2.98 h
OpenCode + MiniMax-M2.1 OpenCode 9.3% 3.73 h
OpenCode + GPT-5.1-Codex-Max OpenCode 7.7% 0.55 h
OpenCode + GLM-4.7 OpenCode 7.5% 2.69 h
Claude Code + Qwen3-Max Claude Code 7.4% 2.1 h
OpenCode + Kimi K2 Thinking OpenCode 7.3% 1.37 h

Performance vs. runtime

Each point is one agent: the average time it spent post-training (of a 10-hour budget) against its average benchmark performance, with the Pareto frontier. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

20% 30% 40% 0246810 Average time spent (h), lower is better Average benchmark performance Claude Opus 4.8 (max)Claude Opus 4.7 (xhigh)GPT-5.5 (xhigh)GPT-5.4 (high)GPT-5.3 Codex (high)GPT-5.3 Codex (med) 20% 30% 40% 0246810 Average time spent (h), lower is better Average benchmark performance Claude Opus 4.8 (max)Claude Opus 4.7 (xhigh)GPT-5.5 (xhigh)GPT-5.4 (high)GPT-5.3 Codex (high)GPT-5.3 Codex (med)
ModelSettingAverage benchmark performanceAverage time spent
Claude Opus 4.8 max, Claude Code 34.1% 8.93 h
Claude Opus 4.8 high, Claude Code 33.8% 9.27 h
Claude Opus 4.7 xhigh, Claude Code 28.6% 7.57 h
GPT-5.5 xhigh, reprompted, Codex CLI 28.4% 9.24 h
GPT-5.4 high, reprompted, Codex CLI 28.2% 8.32 h
GPT-5.5 xhigh, Codex CLI 25% 4.98 h
Claude Opus 4.6 1M context, Claude Code 24.8% 8.8 h
Claude Opus 4.6 Claude Code 23.2% 9.66 h
GPT-5.2 Codex CLI 21.4% 6.08 h
GPT-5.4 high, Codex CLI 20.2% 1.77 h
GPT-5.1-Codex-Max Codex CLI 19.7% 4.05 h
Gemini 3 Pro Gemini CLI 18.1% 6.6 h
GPT-5.3 Codex high, Codex CLI 17.8% 1.65 h
GPT-5.2 Codex Codex CLI 17.2% 2.43 h
Claude Opus 4.5 Claude Code 17.1% 7.88 h
Claude Sonnet 4.6 Claude Code 16.4% 6.83 h
GPT-5.3 Codex med, Codex CLI 13.8% 0.89 h

More effort adds a median 3 points

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Claude Opus 4.8 Claude Code
high
34.1%
GPT-5.5 Codex CLI
xhigh xhigh, repr…
28.4%
GPT-5.4 Codex CLI
high high, repro…
28.2%
Claude Opus 4.6 Claude Code
1M context
24.8%
GPT-5.3 Codex Codex CLI
med high
17.8%

Source: Official PostTrainBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Code + GLM-5.2 max Claude Code 34.3%
Claude Opus 4.8 max Claude Code 34.1%
Claude Opus 4.8 high Claude Code 33.8%
Claude Fable 5 max Claude Code 30.7%
Claude Opus 4.7 xhigh Claude Code 28.6%
GPT-5.5 xhigh, reprompted Codex CLI 28.4%
GPT-5.4 high, reprompted Codex CLI 28.2%
GPT-5.5 xhigh Codex CLI 25%
Claude Opus 4.6 1M context Claude Code 24.8%
Claude Opus 4.6 Claude Code 23.2%
OpenCode + Gemini 3.1 Pro OpenCode 21.6%
GPT-5.2 Codex CLI 21.4%
GPT-5.4 high Codex CLI 20.2%
GPT-5.1-Codex-Max Codex CLI 19.7%
Gemini 3 Pro Gemini CLI 18.1%
GPT-5.3 Codex high Codex CLI 17.8%
OpenCode + Claude Opus 4.5 OpenCode 17.3%
GPT-5.2 Codex Codex CLI 17.2%
Claude Opus 4.5 Claude Code 17.1%
Claude Sonnet 4.6 Claude Code 16.4%
OpenCode + Gemini 3 Pro OpenCode 14.9%
OpenCode + GLM-5 OpenCode 13.9%
GPT-5.3 Codex med Codex CLI 13.8%
OpenCode + Kimi K2.5 OpenCode 10.3%
Claude Sonnet 4.5 Claude Code 9.94%
OpenCode + MiniMax-M2.5 OpenCode 9.50%
OpenCode + MiniMax-M2.1 OpenCode 9.33%
OpenCode + GPT-5.1-Codex-Max OpenCode 7.65%
OpenCode + GLM-4.7 OpenCode 7.48%
Claude Code + Qwen3-Max Claude Code 7.42%
OpenCode + Kimi K2 Thinking OpenCode 7.25%

More effort adds a median 3 points

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Claude Opus 4.8 Claude Code
high
34.1%
GPT-5.5 Codex CLI
xhigh xhigh, repr…
28.4%
GPT-5.4 Codex CLI
high high, repro…
28.2%
Claude Opus 4.6 Claude Code
1M context
24.8%
GPT-5.3 Codex Codex CLI
med high
17.8%

Source: Official PostTrainBench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Opus 4.8 max Claude Code 34.1%
Claude Opus 4.8 high Claude Code 33.8%
Claude Fable 5 max Claude Code 30.7%
Claude Opus 4.7 xhigh Claude Code 28.6%
GPT-5.5 xhigh, reprompted Codex CLI 28.4%
GPT-5.4 high, reprompted Codex CLI 28.2%
GPT-5.5 xhigh Codex CLI 25%
Claude Opus 4.6 1M context Claude Code 24.8%
Claude Opus 4.6 Claude Code 23.2%
GPT-5.2 Codex CLI 21.4%
GPT-5.4 high Codex CLI 20.2%
GPT-5.1-Codex-Max Codex CLI 19.7%
Gemini 3 Pro Gemini CLI 18.1%
GPT-5.3 Codex high Codex CLI 17.8%
GPT-5.2 Codex Codex CLI 17.2%
Claude Opus 4.5 Claude Code 17.1%
Claude Sonnet 4.6 Claude Code 16.4%
GPT-5.3 Codex med Codex CLI 13.8%
Claude Sonnet 4.5 Claude Code 9.94%

4 models have held number one on the official leaderboard since its release in March 2026 8 models have held number one with claims included since its release in March 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

Anthropic OpenAI MiniMax Z.ai Google DeepMind
Own leaderboard Replaced by 1.1, Jul 2026
Lab claims 4 leaders

Sources: Official PostTrainBench leaderboard and labs’ model cards

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6.2 Astra Gemini 4 Argon GLM-5.3 Claude Opus 4.8 Claude Opus 4.7 GPT-5.2 Claude Sonnet 4.5 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6.2 Astra Gemini 4 Argon

Sources: Official PostTrainBench leaderboard and labs’ model cards

4 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

MiniMax
Z.ai
Moonshot AI
Google DeepMind
DateLaunch postScoreAs the post writes it
Oct 13, 2026 Gemini 4 Argon 45.3 PostTrainBench (ML engineering)
Aug 14, 2026 GLM-5.3 39.8 PostTrainBench
Jul 16, 2026 Kimi K3 36.6 PostTrainBench
Jun 16, 2026 GLM-5.2 34.3 PostTrainBench
Jun 1, 2026 MiniMax-M3 37.1 PostTrainBench

13 system cards report it, from 5 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
Nov 17, 2026 GPT-6.2 Sol system card 44.9 PostTrainBench Lite, mean share of performance gap closed, 12 tasks, includes three zero-score attempts (Table 5) AI R&D risk
Nov 17, 2026 GPT-6.2 Astra system card 74.8 PostTrainBench Lite, mean share of performance gap closed, 12 tasks (Table 5) AI R&D risk
Oct 23, 2026 Muse Spark 1.3 system card 43.6 PostTrainBench, seven constituent benchmarks, Muse Spark 1.3 (Helpful) (Table 1, Table 5). Version not stated AI R&D risk
Oct 13, 2026 Gemini 4 Argon model card 45.3 PostTrainBench, ML engineering (results table, p.5) Capabilities
Sep 3, 2026 GPT-6 Astra system card PostTrainBench Lite, mean reward vs output tokens, chart without printed numbers (Figure 55) AI R&D risk
Jul 27, 2026 Kimi K3 technical report 36.6 PostTrainBench, official Harbor implementation, mean of 3 runs on H20 GPUs (Table 2) AI R&D risk
Jul 9, 2026 GPT-5.6 Terra system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jul 9, 2026 GPT-5.6 Sol system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jul 9, 2026 GPT-5.6 Luna system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jun 25, 2026 GPT-5.6 Terra system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jun 25, 2026 GPT-5.6 Sol system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jun 25, 2026 GPT-5.6 Luna system card PostTrainBench Lite, mean reward by latency, chart without printed numbers (Figure 45, 9.1.3.4) AI R&D risk
Jun 23, 2026 Seed2.1 Pro model card 16.5 PostTrainBench (Table 6) AI R&D risk

Versions

We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.

Anthropic OpenAI
PostTrainBench 1.2 6 leaders
PostTrainBench Replaced by 1.1, Jul 2026
PostTrainBench 1.1 Replaced by 1.2, Sep 2026

Source: Official leaderboards

More in AI research

All 164 in AI research
  • PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics. Qwen 3.8 Max 93%
  • MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
  • MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
  • PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
  • KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup. MiniMax-M3 28.8%
  • OpenAI PRs Only in system cards Agents replicate internal OpenAI pull requests from a pre-PR branch and a prompt. Graded by hidden unit tests as pass@1. GPT-5.2 55%

Data audit

All 10 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
  • Pass Independent runs Scores from the official leaderboard (26 entries), launch posts (5 scores) and system cards (6 scores).
  • Pass Models scored 23 models we track, plus 12 other entries on its leaderboard.
  • Pass Newest score The newest is from November 2026, for GPT-6.2 Astra.
  • Note Frontier models 2 of the top 10 have a score, which is fine for a retired test. The strongest tested is GPT-6.2 Astra (#1).
  • Pass Headroom The best is 74.8%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on posttrainbench.com: 31 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check