Sigmabench Update: GPT 5.2 Codex is now live, ranking third overall.
When running in Codex CLI, GPT 5.2 Codex ties GPT-5.1 Codex-Max on accuracy and consistency but outperforms it by being significantly faster.
As of 15 January 2026
Latest Leaderboard
We compare the latest on , , , and , grouping them by relative performance for each metric.
Alerts
Agent
Sigmascore
Accuracy
Consistency
Speed
Cursor CLI
Composer-1
42.9%
39.1%
51.2%
39.5%
Cursor CLI
Grok Code Fast 1
37.8%
34.4%
48.5%
32.4%
Codex CLI
new
GPT-5.2-Codex
33.9%
45.9%
57.6%
14.8%
Gemini CLI
Gemini 3 Flash Preview
33.4%
38.0%
54.8%
17.9%
Codex CLI
GPT-5.1-Codex-Max
31.8%
44.3%
58.3%
12.5%
Gemini CLI
Gemini 3 Pro Preview
30.3%
36.0%
53.2%
14.6%
Claude Code
Claude 4.5 Opus
30.1%
40.9%
52.4%
12.8%
Codex CLI
GPT-5.1-Codex-Mini
29.3%
40.0%
51.5%
12.2%
Codex CLI
GPT-5.1-Codex
26.3%
40.2%
57.1%
7.9%
Claude Code
Claude Haiku 4.5
24.6%
26.2%
33.0%
17.2%
Gemini CLI
Gemini 2.5 Flash
23.0%
17.9%
37.9%
18.1%
OpenCode
Request
—
—
—
 
Warp
Request
—
—
—
 
Amp
Request
—
—
—
 
Cursor CLI
Composer-1
Accuracy:
39.1%
Consistency:
51.2%
Speed:
39.5%
42.9%
Cursor CLI
Grok Code Fast 1
Accuracy:
34.4%
Consistency:
48.5%
Speed:
32.4%
37.8%
Codex CLI
new
GPT-5.2-Codex
Accuracy:
45.9%
Consistency:
57.6%
Speed:
14.8%
33.9%
Gemini CLI
Gemini 3 Flash Preview
Accuracy:
38.0%
Consistency:
54.8%
Speed:
17.9%
33.4%
Codex CLI
GPT-5.1-Codex-Max
Accuracy:
44.3%
Consistency:
58.3%
Speed:
12.5%
31.8%
Gemini CLI
Gemini 3 Pro Preview
Accuracy:
36.0%
Consistency:
53.2%
Speed:
14.6%
30.3%
Claude Code
Claude 4.5 Opus
Accuracy:
40.9%
Consistency:
52.4%
Speed:
12.8%
30.1%
Codex CLI
GPT-5.1-Codex-Mini
Accuracy:
40.0%
Consistency:
51.5%
Speed:
12.2%
29.3%
Codex CLI
GPT-5.1-Codex
Accuracy:
40.2%
Consistency:
57.1%
Speed:
7.9%
26.3%
Claude Code
Claude Haiku 4.5
Accuracy:
26.2%
Consistency:
33.0%
Speed:
17.2%
24.6%
Gemini CLI
Gemini 2.5 Flash
Accuracy:
17.9%
Consistency:
37.9%
Speed:
18.1%
23.0%
OpenCode
Warp
Amp
Why the best coding agent depends on your codebase.
Across the Sigmabench dataset, coding agent performance varies widely between similar codebases. These differences emerge from the unique characteristics of each codebase — shaped over time by the humans who built it — and cannot be reliably inferred from large-scale features like programming language, problem domain, or codebase size. This makes common decisions hard to answer with confidence:
Which agent should we standardize on?
Are we using the right agent for our codebase?
Would switching agents materially improve accuracy or speed?
Are recent improvements real, or just hype?
Sigmabench lets you run the same benchmark used for the public leaderboard on your own codebase, so you can answer these questions with data — measuring accuracy, consistency, and speed in your real-world development environment.
Benchmarks are read-only and SOC 2-compliant.