Frontier Code
A third-party programming benchmark maintained by Cognition, the developer of Devin AI. It is designed to measure the capabilities of long-horizon coding agents and is considered closer to real-world production scenarios than HumanEval.
Core Facts
FrontierCode 1.1 is an evaluation benchmark for AI programming capabilities that simulates real software engineering tasks, including scenarios such as code completion, bug fixing, multi-file refactoring, and unit test generation.
Claude Fable 5.2 scored 49.1% on the FrontierCode 2.0 benchmark, ranking first.
Claude Opus 5 approaches Fable-level performance on the FrontierCode 1.1 benchmark while costing only half as much
Timeline (last 90 days)
On the Terminal Bench and Frontier Code benchmarks, Claude Opus 5.5 performed slightly better than Gemini 4 Neon.
Fable 5.2 scored 49.1% on the Frontier Code coding benchmark and 69.9% on the Frontier SWE agentic coding benchmark, ranking first on both.
Claude Fable 5.2 scored 49.1% on the FrontierCode 2.0 benchmark, ranking first.
On the Frontier Code reasoning benchmark, the medium reasoning tier performed better than the high and extra high tiers, approaching the max tier; this phenomenon occurred with both Opus and Fable.
On long-horizon programming tasks such as the Frontier Code benchmark, Opus 5.5 may slightly outperform Astra
Excessive reasoning can hurt performance, with x-high scoring higher than Max on the Frontier Code benchmark.
Sonnet 5.5's scores on Frontier Code-related benchmarks have risen from 42 to 46 and 52.
On Frontier Code, Opus 5.5's medium tier (under $1 per task) actually scored higher than its max tier (over $5 per task).
On Frontier Code, the performance gap among GPT-6 Sol, Luna, and Opus 5.5 is only about 6–7 percentage points, but their prices range from 10 cents, to $2, to $4.
In the Frontier Code test, GPT-6 Luna Max achieved 66.6%, with a cost of only 22 cents per task.
4 more timeline events
All Facts (14)
FrontierCode 1.1 is an evaluation benchmark for AI programming capabilities that simulates real software engineering tasks, including scenarios such as code completion, bug fixing, multi-file refactoring, and unit test generation.
75%VerifiedClaude Fable 5.2 scored 49.1% on the FrontierCode 2.0 benchmark, ranking first.
70%VerifiedClaude Opus 5 approaches Fable-level performance on the FrontierCode 1.1 benchmark while costing only half as much
70%UnverifiedOn the Terminal Bench and Frontier Code benchmarks, Claude Opus 5.5 performed slightly better than Gemini 4 Neon.
50%UnverifiedFable 5.2 scored 49.1% on the Frontier Code coding benchmark and 69.9% on the Frontier SWE agentic coding benchmark, ranking first on both.
50%UnverifiedOn the Frontier Code reasoning benchmark, the medium reasoning tier performed better than the high and extra high tiers, approaching the max tier; this phenomenon occurred with both Opus and Fable.
50%UnverifiedExcessive reasoning can hurt performance, with x-high scoring higher than Max on the Frontier Code benchmark.
50%UnverifiedSonnet 5.5's scores on Frontier Code-related benchmarks have risen from 42 to 46 and 52.
50%UnverifiedOn Frontier Code, Opus 5.5's medium tier (under $1 per task) actually scored higher than its max tier (over $5 per task).
50%UnverifiedOn Frontier Code, the performance gap among GPT-6 Sol, Luna, and Opus 5.5 is only about 6–7 percentage points, but their prices range from 10 cents, to $2, to $4.
50%UnverifiedIn the Frontier Code test, GPT-6 Luna Max achieved 66.6%, with a cost of only 22 cents per task.
50%UnverifiedCognition has released coding model Sway 2, achieving 50.0% on the Frontier Code benchmark, claiming to be within 1 point of stronger models while costing 64% less.
50%UnverifiedOn the Frontier Code programming benchmark, the old model scored 34.4%, while Gemini 3.7 Flash improved to 43.6%.
50%UnverifiedIn Cognition's Frontier Code evaluation, Fable 5 ranked first among all frontier models even with only medium reasoning intensity enabled.
50%