Grok 4.6 delivers strong agent performance and lower price than direct rivals

Grok 4.6 is released, and this time the story centers on professional agent work rather than model size. The model outperforms GPT-5.6 Sol Max and Fable 5 Max on two real-world agent benchmarks while also being cheaper per token.

In the Artificial Analysis metric, which aggregates general capabilities, Grok 4.6 scores 61, exactly the same as GPT-5.6 Sol Max. The price difference is $2 per million input tokens and $6 per million output tokens. Against Fable 5 Max, which positions itself as the market flagship, Grok 4.6 reaches 98.4% of its Intelligence Index score, but at a cost five times lower on input and eight times lower on output. This shifts the performance-to-cost balance toward the new model.

The most obvious advantage of Grok 4.6 lies in professional agent tasks. On GDPVal-AA v2 it records 1753, and on AA-Briefcase 1577, both ahead of Sol Max and Fable 5 Max. These metrics test open-ended knowledge-agent work such as research, analysis and multi-file generation, making the results more relevant to laboratory use than trivia tests. In image coding, Grok 4.6 achieves 69.9% on CursorBench v3.2, surpassing Sol Max’s 67.2%, but lags behind on DeepSWE and Terminal-Bench.

8.

SpaceXAI attributes the leap to several changes in the training process. The run was longer, and the training firm re-processed the data stack: SFT (Supervised Fine-Tuning) paths were rebuilt, model-based trace filtering was applied, and agentic RL (Reinforcement Learning) was performed across coding, web development, CAD and kernel optimization. The company also reports more extensive self-checks on long traces, with the model reviewing its work before proceeding. The stated goal is to tackle error accumulation that arises across multi-step agent chains. All figures come from the company’s report and have not been independently verified.