Best composite score
78.4%
GPT-5.1 Codex-Max · agentic coding
vs. previous index+4.8 pts
Practical intelligence index / 2026
The benchmark for models that do the whole job.
A clear view of model performance across coding, reasoning and agentic work — with cost and time kept in the frame.
Lowest cost / task
$0.09
DeepSeek-V4-Pro · median run
efficiency signaltop 12%
Runs completed
1,184
Across 9 task families
Leaderboard
Model configurations ranked by the current composite index
#ModelProviderFocusScoreCostTrend
01
OAI
OpenAIcoding78.4%$4.20+4.802
ANT
Anthropicagentic74.1%$11.84+2.103
G
Googlereasoning71.8%$2.18+3.504
DS
DeepSeekcoding68.9%$0.09+1.705
QW
Qwencoding65.4%$0.42+1.206
xAI
xAIagentic63.2%$2.00+0.8Showing 6 of 42 configurations
What makes a useful benchmark?
A score should reveal how a model behaves when the obvious answer is not the finished work.
Outcome over syntax
We score observable behavior and verified results, not similarity to a reference patch.
Context is part of the task
Repositories, documents, tools and changing constraints are part of the environment.
Quality has a price tag
Every run keeps time, tokens and cost visible beside the pass rate.
Run history
Recent suite activity across the indexed configurations
Latest batch
#0148
113 tasks · completed 12m ago
verified passes76.2%
Median latency
17min
Across all agent runs
vs. last week−8.4%
Next refresh
04:12h
Scheduled index update
data pipelinehealthy