Last verified 2026-11-08

AI researchVendor-reportedPrices verified

PostTrainBench leaderboard

As of November 8, 2026, Gemini 4 Argon (Google DeepMind) leads PostTrainBench at 45.3%, ahead of GLM-5.3 at 39.8% and MiniMax M3.1 Flash at 39.3%. 6 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.

See the ranking →All benchmarks →Model prices →

Top score

45.3%

Gemini 4 Argon

Systems ranked

one setting

Independent rows

0/6

all vendor-reported

Cheapest near top

$1.21/1M

Gemini 4 Argon · within 5 pts

With API prices

6/6

rates from our price index

Newest score

Nov 2

2026 · newest result

What is PostTrainBench? PostTrainBench tests whether an AI agent can post-train a small base language model on its own, with 10 hours on one H100 GPU, to score as high as possible on a target benchmark. Made by Ben Rank, Maksym Andriushchenko and co-authors, March 2026.

§ 01 / LEADERBOARD

#	Model	Score	Price / 1M	Context	Source
1	Gemini 4 Argon Google DeepMind	45.3% as reported	$2 / $10	1M	Vendor-reported Google DeepMind 2026-10-13
2	GLM-5.3 Zhipu (Z.ai / GLM)	39.8% as reported	$1.4 / $4.4	1M	Vendor-reported Z.ai (Zhipu) 2026-08-25
3	MiniMax M3.1 Flash MiniMax	39.3% v1.1	$0.15 / $0.6	1M	Vendor-reported MiniMax 2026-10-19
4	Kimi K3 Moonshot (Kimi)	36.6% max effort	$3 / $15	1M	Vendor-reported Moonshot AI 2026-07-16
5	Kimi K3.1 Pro Moonshot (Kimi)	36.5% max effort, v1.2	$2 / $10	1M	Vendor-reported Moonshot AI 2026-11-02
6	GLM-5.2 Zhipu (Z.ai / GLM)	31.7% as reported	$1.4 / $4.4	1M	Vendor-reported Z.ai (Zhipu) 2026-08-25

PostTrainBench: score against price

The dashed line joins the models that no cheaper model beats: MiniMax M3.1 Flash (39.3%, $0.0842 per 1M tokens), GLM-5.3 (39.8%, $0.78 per 1M tokens) and Gemini 4 Argon (45.3%, $1.21 per 1M tokens). Anything below the line costs more for a lower score. Price is the effective rate per 1M tokens on current list prices: 92% input, 8% output, 82% of input served from cache.

§ 03 / PROVENANCE

Can you trust PostTrainBench scores?

No independent leaderboard publishes PostTrainBench results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.

!The suite has changed between versions (benchmarks added, dropped and reweighted), so scores from different versions are not comparable.
!Runs flagged for contamination are scored as the untrained base model.
!The 10-hour, one-GPU limit is the point of the test: it says nothing about post-training at lab scale.

§ 04 / HOW IT WORKS

How PostTrainBench works

The task

The agent gets a base model (such as Qwen3-4B), a target benchmark (such as AIME) and a fixed compute budget. It finds data and methods on its own, runs the training and hands back a model.

The score

The trained models are evaluated on their target benchmarks, and the leaderboard averages the result across base models and benchmarks.

Test set

Four base models × the target benchmarks of the current version, as published by the authors.

Made by

Ben Rank, Maksym Andriushchenko and co-authors, first published March 2026.

Official sources

posttrainbench.com
Paper: arxiv.org/abs/2603.08640
Definition read on 2026-11-08

§ 05 / SOURCES

Sources and freshness

6 published scores from 5 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.

Vendor-reported Z.ai (Zhipu) 2 scores · newest 2026-08-25
Vendor-reported Google DeepMind 1 score · newest 2026-10-13
Vendor-reported MiniMax 1 score · newest 2026-10-19
Vendor-reported Moonshot AI 1 score · newest 2026-11-02
Vendor-reported Moonshot AI 1 score · newest 2026-07-16

PostTrainBench FAQ

Answers derived from the 6 published scores on this page and the current price index.

Q · 01

Which AI model has the highest PostTrainBench score?

Gemini 4 Argon with 45.3% (as reported), as reported by Google DeepMind on October 13, 2026. Next: GLM-5.3 at 39.8% and MiniMax M3.1 Flash at 39.3%.

Q · 02

What is the cheapest model near the top of PostTrainBench?

The leader itself. No priced model within 5 points of Gemini 4 Argon is cheaper: it runs at $1.21 per 1M tokens on our effective blend ($2 input, $10 output).

Q · 03

Are these PostTrainBench scores independently verified?

Not yet. Every score here is the model maker's own report. Each row names the lab and links to the page where it published the number.

Q · 04

What does PostTrainBench measure?

PostTrainBench tests whether an AI agent can post-train a small base language model on its own, with 10 hours on one H100 GPU, to score as high as possible on a target benchmark. The agent gets a base model (such as Qwen3-4B), a target benchmark (such as AIME) and a fixed compute budget. It finds data and methods on its own, runs the training and hands back a model.

Q · 05

How is the PostTrainBench score calculated?

The trained models are evaluated on their target benchmarks, and the leaderboard averages the result across base models and benchmarks. The test set has Four base models × the target benchmarks of the current version.

Q · 06

How current is this PostTrainBench leaderboard?

The newest score was published on November 2, 2026. Every row shows the date its source published it, and we last read all of them on November 8, 2026. Prices come from our model price index, where each rate carries the date it was read on the vendor's page.