AI-generated editorial illustration; not a documentary photograph.
SpaceXAI released Grok 4.7 on Sunday with training focused on coding-agent tasks that can run for many hours. The company says an extended reinforcement-learning run helped the model manage longer context and verify its own work, but benchmark results show that the model remains inconsistent on extended software tasks.
Higher scores on long tasks
Grok 4.7 scored 38.0% on Terminal-Bench 4.0, compared with 20.3% for Grok 4.6. Its CursorBench 4.0 result rose from 40.4% to 46.3%, while its AA Briefcase v1.1 score increased from 1,546 to 1,657. These evaluations cover terminal work, longer coding workflows and multi-hour professional tasks, respectively.
The improvement does not put Grok 4.7 at the top of every comparison. Anthropic’s Claude Fable 5.1 scored 57.9% on Terminal-Bench 4.0 according to the independent leaderboard cited by The New Stack. The source also reports that a recent private-codebase benchmark saw even the best-performing model fail more than 60% of the time, without identifying Grok 4.7 as that model.
Model and harness move closer
SpaceXAI trained Grok 4.7 to understand the Grok Bot harness natively. A harness supplies tools, formats terminal responses, returns execution results to the model and determines what happens next. Embedding those interaction patterns in training could reduce the prompting overhead of multi-step tool use, although the company has not disclosed how much of the benchmark gain comes from this approach.
The design may also make model substitution harder. If an agent depends on particular tool schemas, context formats or execution environments, developers could lose performance when switching models. The New Stack compares this model-centered strategy with Google’s work on changing the development environment to make Go easier for agents to use.
Reliability and cost remain limits
The source says SpaceXAI has not explained whether Grok 4.7’s context-management gains come from architectural changes, summarization, retrieval or better retention across long sequences. It also has not detailed how self-verification was evaluated during reinforcement learning. Those omissions make it difficult to isolate why the model improved or how reliably it can detect a bad assumption before building on it.
Grok 4.7 starts at $2 per million input tokens and $6 per million output tokens. Lower token costs may make multi-hour runs more economical, but the Terminal-Bench gap with Claude Fable 5.1 shows that price does not resolve the model’s reliability limitation.
Source: Grok 4.7 was built to work for hours. It still fails most of the time.
SIGNAL TITANS
Leave a Reply