0

FrontierCode

Frontier

Frontier software-engineering benchmark from Cognition; the Diamond split is the hardest tier.

Domain
coding
Published
Jun 2026
Notable for
Hard agentic-coding benchmark used at Claude Fable 5 launch.

Cite

Notes

Only stored in your browser.

Attribution

Leaderboard scores
AnthropicOpenAI
Attribution policy →

Top score 53.3% by GPT-6 Astra - 4 models reporting (4 frontier)

Score history

4
0%25%50%75%100%Apr 26Jun 26Aug 26GPT-5.5Claude Opus 4.8Claude Fable 5GPT-6 Astra

Top models

4
FrontierCodeBar chart with 4 bars. Highest value: GPT-6 Astra at 53.3.
4 models

FAQ

What is FrontierCode?
Frontier software-engineering benchmark from Cognition; the Diamond split is the hardest tier.
What is the current top score on FrontierCode?
The top reported score is 53.3% by GPT-6 Astra, across 4 models reporting (4 from frontier labs).