BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

GSO-Bench

Software optimisation tasks measured against expert speed-ups.

As of 24 Oct 2026, Claude Fable 5.1 leads GSO-Bench on BenchLeader with 88.2%, ahead of GPT-6 Astra at 79.4%, across 35 model configurations with a published result.

Published by
GSO-Benchdata via Epoch AI Benchmarking Hub
Category
Coding
Index weight
0.5
Models
35
Data as of
24 Oct 2026
Newest result
3 Sept 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

GSO-Bench asks a model to make real software faster: it is given a repository and a performance target and must submit an optimisation, which is measured against an expert's speed-up.

How it is scored

Percent of tasks where the model's change achieves the expert-level speed-up while keeping the tests green.

What to keep in mind

Performance measurements depend on hardware and noise, and the bar is the expert's result, so most models score low. It rewards systems knowledge that general coding tests miss.

Top 15
  1. 1Claude Fable 5.188.2%
  2. 2GPT-6 Astra79.4%
  3. 3Claude Fable 578.4%
  4. 4GPT-5.6 Sol76.5%
  5. 5Claude Opus 4.847.1%
  6. 6Claude Opus 4.744.1%
  7. 7Claude Opus 4.7 (high)44.1%
  8. 8Claude Opus 4.6 (high)41.2%
  9. 9GPT-5.5 (xhigh)40.2%
  10. 10Claude Sonnet 537.3%
  11. 11Claude Opus 4.633.3%
  12. 12GPT-5.4 (xhigh)31.4%
  13. 13GPT-5.2 (high)27.4%
  14. 14Claude Opus 4.526.5%
  15. 15GPT-5.4 (high)25.5%
35 of 35
#
1Claude Fable 5.188.2%66.42026-09-01
2GPT-6 Astra79.4%68.02026-09-03
3Claude Fable 578.4%67.12026-06-09
4GPT-5.6 Sol76.5%55.92026-07-09
5Claude Opus 4.847.1%60.12026-05-28
6Claude Opus 4.744.1%63.32026-04-16
7Claude Opus 4.7high44.1%61.02026-04-16
8Claude Opus 4.6high41.2%62.72026-02-05
9GPT-5.5xhigh40.2%66.12026-04-23
10Claude Sonnet 537.3%56.12026-06-30
11Claude Opus 4.633.3%62.62026-02-05
12GPT-5.4xhigh31.4%64.72026-03-05
13GPT-5.2high27.4%56.02025-12-11
14Claude Opus 4.526.5%57.92025-11-24
15GPT-5.4high25.5%57.92026-03-05
16Gemini 3.1 Pro22.6%60.62026-02-19
17Gemini 3 Pro18.6%60.42025-11-18
18Claude Sonnet 4.514.7%53.72025-09-29
19GPT-5.113.7%56.82025-11-13
20GPT-5.1high13.7%57.72025-11-13
21Gemini 3 Flash9.8%57.02025-12-17
22o3high8.8%53.92025-04-16
23GPT-5high6.9%57.92025-08-07
24Claude Opus 46.9%52.62025-05-22
25Claude Sonnet 44.9%50.52025-05-22
26Kimi K24.9%50.42025-07-12
27Qwen3 Coder 480B4.9%48.02025-07-31
28Claude 3.5 Sonnet4.6%46.32024-10-22
29Gemini 2.5 Pro3.9%53.92025-06-17
30Claude 3.7 Sonnet3.8%50.02025-02-24
31o4-minihigh3.6%50.12025-04-16
32GLM-4.5-Air2.9%46.62025-07-20
33o3-minihigh1.3%47.22025-01-31
34o3-minilow1.3%40.42025-01-31
35GPT-4o0.0%43.92024-11-20

More coding benchmarks · All LLM benchmarks

Cite as: BenchLeader, “GSO-Bench leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/gso_bench.

GSO-Bench: questions

What does GSO-Bench measure?
GSO-Bench asks a model to make real software faster: it is given a repository and a performance target and must submit an optimisation, which is measured against an expert's speed-up. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads GSO-Bench?
Claude Fable 5.1 leads GSO-Bench with 88.2% as of 24 Oct 2026, ahead of GPT-6 Astra at 79.4%.
How many models have GSO-Bench results?
35 model configurations have a GSO-Bench result on BenchLeader, all taken from GSO-Bench via Epoch AI Benchmarking Hub.
Who runs GSO-Bench and how often is it updated?
GSO-Bench is published by GSO-Bench. BenchLeader re-reads the published results once a day and records the date each result was published.
Does GSO-Bench count toward the BenchLeader Index?
Yes. GSO-Bench contributes to the coding category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.