AIEvals

Benchmarks

PostTrainBench

Whether an agent can post-train a language model by itself: four small base models, one H100 and ten hours to improve them, scored on seven weighted benchmarks.

AI researchPostTrainBench authors (aisa-group), verification by Epoch AIVendor-reported onlyReported by Google DeepMind, Z.ai

How it is run

Agents are given no strategy — they search the web, run experiments and curate data on their own — and the score is the average improvement across the four base models. Version 1.1 is the current results version, with the original v1 run kept beside it, and the study is published as a paper (arXiv 2603.08640) with Epoch AI verifying the results.

What it cannot tell you

A research task, not a coding task: the score depends on which four base models the run used and on how much of a starting point each one leaves, so a strong software agent can still do badly here. Every figure on this page comes from a publisher's own table.

Why there is no independent table

No independent suite in this dataset runs PostTrainBench. The numbers below are the ones model publishers report about their own models, which is a different kind of evidence: each publisher chooses its own harness and effort level, so a row is readable on its own and not against the other rows.

Read this with

Nothing else on this site shares this benchmark's family or category, so these are other evaluations the same publisher reports.

  • LABBench 2Scientific and laboratory problem solving with access to a computational environment.
  • RiemannBenchMathematical reasoning, included in the report’s science and math evaluations.
  • GraphWalks (up to 128k)Breadth-first graph traversal over long inputs, scored with F1.
  • GraphWalks (256k to 1M)Breadth-first graph traversal over very long inputs, scored with F1.

As reported by each publisher

Kept separate per publisher, because the settings are not shared.

Google DeepMind

Gemini 4 Argon: model evaluation · October 2026

RankModelScore
2Gemini 4 Argon45.3%
3GPT-6 Astra44.3%
4Claude Fable 5.140.2%
1Claude Opus 5.549.3%

Ranks compare reported scores in the same published row; ties share a rank.

Page 5 of Google DeepMind’s model evaluation, read 1 October 2026. Gemini runs use the API at the highest thinking setting, generally pass@1 without majority voting or parallel test-time compute; smaller benchmarks average multiple trials. Other models use their highest available published reasoning results. Google quotes Vals AI for knowledge-work and Vibe Code results, Zapier’s private-set AutomationBench leaderboard, Proximal for FrontierSWE v2, Surge AI for RiemannBench and no-tools Chartography, and the CWE-bench v1 leaderboard. Argon’s DeepSWE uses mini-swe; Terminal-Bench 4.0 is self-computed for Argon and quoted from the author board for other models. PostTrainBench v1.1 uses OpenCode with one H100 and a 10-hour budget for all models. Terminal-Bench Science uses a 6x verifier timeout for Argon. LABBench 2 uses terminal, bioinformatics tools, Python, R and internet access for all models. GraphWalks uses BFS F1 on two separate subsets: 650 items up to 128k tokens and 200 items from 256k to 1M. Agent’s Last Exam uses ALE-Claw, a 5-hour window and binary pass rate, with safety filters enabled for Argon. OSWorld 2.0 uses offline partial reward, the best of three single-attempt runs for Argon, safety filters, 1080p, 500 steps, Gemini CUA with parallel batch tools and compaction, screenshot-only observation and the 08.08 patch; Claude’s combined online/offline results are omitted. LVBench runs without tools, using 1 FPS for Gemini, 800 frames for Astra, 300 for Fable and 600 for Opus. Missing cells remain blank. These are the report’s published figures, preserved separately from this site’s independent source snapshots.

storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf

Z.ai

GLM-5.3: Frontier Coding with Emergent Cyber Capabilities — evaluation table · 14 August 2026

RankModelScore
2GLM 5.339.8%
6GLM 5.231.7%
5Kimi K332.0%
4Claude Opus 4.832.9%
1Claude Fable 541.8%
3GPT-5.6 Sol36.2%

Ranks compare reported scores in the same published row; ties share a rank.

The comparison table Z.ai publishes with GLM-5.3, read on the model card itself. Z.ai states that GLM-5.3 uses the same base model as GLM-5.2 and that every gain comes from post-training, so the GLM-5.2 column is the same architecture one training round earlier. Column names were restated in the form this site uses, and Z.ai labels its own Fable column "Fable 5 (w/ fallback)", which is recorded here as Claude Fable 5 and is the only column whose runs used a fallback model. Rows whose scale is not a percentage keep their own unit: ExploitGym is reported as tasks completed with a 2-hour and a 6-hour budget and is split into two rows so both numbers stay readable, and GDPval-AA v2 is a score rather than a percentage. The table is the only source for several benchmarks on this site, and nothing in it is an independent run.

z.ai/blog/glm-5.3

Sources

Every page used for this benchmark, read on 17 October 2026.