Benchmark · math

GSM8K

saturated 13 results 11 models

GSM8K (Grade School Math 8K) is a benchmark of grade-school math word problems that each need several steps of arithmetic reasoning to solve; a model is scored on the percentage of problems where its final numeric answer is correct.

Read more
Example
A short word problem like "Natalia sold clips to 48 friends in April and half as many in May; how many clips did she sell in total?" — the model must work through the steps and give the final number.
Scoring
The score is accuracy — the share of problems whose final numeric answer exactly matches the reference answer; intermediate reasoning is not graded, only the final number.
Verification
Verification is automatic and exact-match: the model's final number is parsed out (typically after a delimiter like "####") and compared to the gold answer, so no human judgment is needed.
Why it matters
It became a standard, easy-to-check test of multi-step reasoning, and the well-known chain-of-thought prompting technique was shown to sharply improve results on it, making it a common yardstick for reasoning ability.
Worked example
Task
Maria has 3 boxes of crayons with 24 crayons in each box. She gives 15 crayons to her brother, then buys 2 more full boxes. How many crayons does she have now?
Solution
3 × 24 = 72; 72 − 15 = 57; 2 × 24 = 48; 57 + 48 = 105 → 105
Walkthrough
Multiply to get the starting count, subtract what she gives away, then add the newly bought boxes: 72 − 15 + 48 = 105. GSM8K grades only the final number extracted after the '####' delimiter by exact match, so the reasoning steps are not scored — only 105 counts.
0 25 50 75 100 2023-03-14 2024-11-18 2026-07-26 Grok-1.5 · 90 · 2024-03-28 DUP method · 97.1 · 2024-04-23 Claude 3.5 Sonnet · 91.6 · 2024-06-20 Qwen2-72B · 89.5 · 2024-07-15 Qwen2-72B · 89.5 · 2024-07-15 Falcon3-10B-Instruct · 83.1 · 2024-12-17 Falcon3-10B-Base · 83 · 2024-12-17 Falcon3-7B-Base · 79.1 · 2024-12-17 LLaDA-8B · 76 · 2026-06-25 LLaDA-8B · 76 · 2026-06-29 Gemma-4-12B · 86 · 2026-07-13 Qwen/Qwen2.5-7B-Instruct@a09a35458c702b33eeacc393d103063234e8bc28 · 74.2 · 2026-07-26 GPT-4 · 92 · 2023-03-14
Grok-1.5 DUP method Claude 3.5 Sonnet Qwen2-72B Falcon3-10B-Instruct Falcon3-10B-Base Falcon3-7B-Base LLaDA-8B Gemma-4-12B Qwen/Qwen2.5-7B-Instruct@a09a35458c702b33eeacc393d103063234e8bc28 GPT-4
Timeline
Date Model Score Source
2026-07-26 Qwen/Qwen2.5-7B-Instruct@a09a35458c702b33eeacc393d103063234e8bc28 74.219% CrowdTensor launches volunteer-training Beta and Draft Qwen2.5-7B GSM8K Campaign RFC
2026-07-13 Gemma-4-12B 86.0% Flint compresses reasoning traces to reduce token usage while maintaining accuracy
2026-06-29 LLaDA-8B 76.0% Scheduling Thoughts: Learning the Order of Thought in Diffusion Language Models
2026-06-25 LLaDA-8B 76.0% Self-Aware Scheduling Learns Token Unmasking Order in Diffusion Language Models
2024-12-17 Falcon3-10B-Instruct 83.1% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-12-17 Falcon3-10B-Base 83.0% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-12-17 Falcon3-7B-Base 79.1% Falcon3 family releases five open models with improved science, math, and code capabilities
2024-07-15 Qwen2-72B 89.5% Qwen Team releases Qwen2 series with 72B dense and MoE models
2024-07-15 Qwen2-72B 89.5% Qwen2 Technical Report introduces dense and MoE models up to 72B
2024-06-20 Claude 3.5 Sonnet 91.6% Anthropic releases Claude 3.5 Sonnet addendum with improved coding and vision benchmarks
2024-04-23 DUP method 97.1% Deeply Understanding the Problems method achieves 97.1% on GSM8K
2024-03-28 Grok-1.5 90.0% xAI releases Grok-1.5 with improved reasoning and 128K context
2023-03-14 GPT-4 92.0% OpenAI