GSM8K Leaderboard — AI Model Scores

What this benchmark measures

• multi-step arithmetic reasoning
• chain-of-thought problem solving

Relevant to:

• reasoning baseline screening
• chain-of-thought regression checks

Failure modes it exercises:

• arithmetic slips across multi-step calculations

Method and its limits

8.5K grade-school maths word problems solved with chain-of-thought prompting and scored by exact match on the final numeric answer.

• Exact-match on a final number gives no credit for a sound method with one slip, and no penalty for a right answer reached by faulty reasoning.
• The paper's framing is that conceptually simple problems defeated the largest models of 2021; frontier models now score near the ceiling, so the benchmark no longer separates them.

Dataset

Dataset

GSM8K

Type

grade-school maths word problems

Freshness

aging

8.5K problems as registered in the primary paper. The date records paper submission for this dataset release, not a knowledge cutoff.

How to read this score

Accuracy is interpretable only alongside the prompting protocol. Near the ceiling the remaining errors are concentrated in a small set of items, so point differences are not meaningful.

Similarity to real tasks: Low — Established from the primary methodology: GSM8K is 8.5K linguistically diverse grade-school maths word problems requiring multi-step reasoning to a single numeric answer. It is a controlled probe of arithmetic chain-of-thought and does not reproduce production work, where quantities arrive from documents or tools, intermediate results must be shown and audited, and there is rarely one indisputable number.

Data contamination risk: High — GSM8K has been public since 2021 and is among the most heavily used maths datasets in instruction tuning. No run-specific overlap audit is registered.

Benchmark gaming risk: High — Prompt style, number of chain-of-thought exemplars, self-consistency sampling, and answer-extraction rules all move the reported figure substantially, and none is registered per score here.

What you still need to test yourself

• Test arithmetic over your own documents and tool outputs, where the numbers are not supplied in the prompt.
• Evaluate whether working is auditable, which exact-match scoring ignores.

This benchmark supports decisions about:

• Use as a floor check that multi-step arithmetic reasoning is intact.
• Detect regressions after fine-tuning rather than rank frontier models.

Limitations

• Saturated: it no longer separates capable models.
• Not evidence of reasoning quality on problems without a single numeric answer.

Primary source reviewed 2026-10-12. GSM8K is treated as an aging 2021 benchmark that is effectively saturated for frontier models.

What this evidence does not cover

• No model scores are registered on this page, so it supports no comparison.

Sources

• Training Verifiers to Solve Math Word Problems Cobbe, Kosaraju, Bavarian, Chen, Jun, Kaiser, Plappert, Tworek, Hilton, Nakano, Hesse, Schulman · accessed 2026-10-12Supports the 8.5K grade-school dataset release and the authors' finding that the largest models of the time failed despite the problems' conceptual simplicity; it does not support any model score, and no score is registered on this page.

How this benchmark is scored

Primary source: https://arxiv.org/abs/2110.14168

Published results

Each score reflects the protocol and date of its own source run. Results from different harnesses are not directly comparable.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to GSM8K Leaderboard — AI Model Scores.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path