Grok 4.7 trained for long-running tasks, but still struggles with reliability

xAI released Grok 4.7 with a training approach specifically designed to improve performance on extended coding tasks. The model uses reinforcement learning weighted toward harder problems—some spanning many hours—to help it maintain focus and catch mistakes across long sequences of work.

Benchmarks show sharp improvement—but context matters#

The results on specialized endurance tests are notable. Grok 4.7 jumped from 20.3% to 38.0% on Terminal-Bench 4.0, climbed from 40.4% to 46.3% on CursorBench 4.0 (which measures longer coding workflows inside editors), and improved from 1,546 to 1,657 on AA Briefcase v1.1, an evaluation of multi-hour professional tasks.

Yet these gains come with an important caveat: Anthropic's Claude Fable 5.1 still outperforms Grok 4.7, scoring 57.9% on Terminal-Bench versus Grok's 38.0%. More concerning, benchmarks of real-world codebases show that even the best-performing models fail over 60% of the time. The fundamental challenge of unattended execution remains: as agents work through hours of tasks, they accumulate error history and must verify each step before proceeding, or risk cascading failures.

Integration of tools into the training process#

xAI took the approach of embedding the Grok Bot harness directly into training, rather than leaving the model to learn tool schemas and formats through prompting at runtime. This integration mirrors OpenAI's recent move to formalize agent infrastructure through its Agents API, treating the scaffolding around models as a service layer.

The advantage is clear in theory: a model already fluent in its native harness reduces overhead in tool use and multi-step reasoning. The tradeoff, however, is lock-in. Developers may find it difficult to swap in a different model without sacrificing agent performance, since alternatives won't have the same harness-specific training. As agents take on larger portions of the development cycle—and as systems like Google's recent work on Go optimization shows—the optimization boundary is shifting. It's no longer just about the model; the entire surrounding infrastructure is being tailored for agent work.

Reliability remains the biggest barrier#

At $2 per million input tokens and $6 per million output tokens, Grok 4.7 is competitively priced for multi-hour agent runs. Affordability alone, however, doesn't solve the core problem: a 38% success rate on Terminal-Bench means the model still fails more often than it succeeds on standardized tests. In production scenarios, that gap widens further.