Breaking

Cognition Adds Grok 4.6 to Devin as Model Ranks Third on FrontierCode

August 12, 2026 at 12:40 EDT

Foundation Models
AI Agents
Software Dev & Coding

Cognition has made SpaceXAI's Grok 4.6 available in Devin, adding the newly released model to its autonomous software-engineering agent. The integration arrives with a third-place result on Cognition's FrontierCode 1.1 coding benchmark, behind Claude Fable 5 and Claude Opus 5 but ahead of GPT-5.6 Sol.

August 12, 2026 · Cognition + SpaceXAI

Grok 4.6 lands in Devin

The new model takes third place on Cognition's production-code benchmark, closing much of the gap between Grok 4.5 and the leading Claude models.

48.0%

FrontierCode 1.1 Main

+5.6 pts

Gain over Grok 4.5

500K

Token context window

$2 / $6

Input / output per 1M tokens

FrontierCode 1.1 Main

Weighted score, shown from zero

Higher is better

42.4%

Grok 4.5

47.5%

GPT-5.6 Sol

48.0%

Grok 4.6

53.4%

Claude Opus 5

53.5%

Claude Fable 5

Takeaway: Grok 4.6 moves ahead of GPT-5.6 Sol, but remains 5.4 points behind Claude Opus 5.

How the model was strengthened

Longer supplemental training

→

Regenerated SFT trajectories

→

Agentic reinforcement learning

→

Devin model choice

Early upside

Competitive coding value

Developers see a promising secondary model with stronger benchmark performance than Grok 4.5.

Open questions

Workload proof still needed

Users want bug-for-bug tests against Fable 5, while some consider the 500K context window restrictive.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.