SpaceXAI's latest AI coding agent jumped from 20.3% to 38.0% accuracy on a benchmark testing multi-hour programming tasks, according to Grok 4.7's release details published Sunday. The improvement comes from extended training focused specifically on lengthy, difficult problems and the agent's ability to catch its own mistakes before they compound. The company designed the update to solve a critical weakness: when coding agents run unattended for hours, a single error early in the process can cascade through every subsequent decision unless the system spots it.
Grok 4.7's performance gains showed up most clearly on tests measuring endurance rather than raw speed. The model climbed from 40.4% to 46.3% on CursorBench 4.0, which evaluates extended coding workflows inside an editor, and rose from 1,546 to 1,657 on AA Briefcase v1.1, an assessment of professional tasks lasting multiple hours. Anthropic's Claude Fable 5.1 still leads on Terminal-Bench 4.0 with 57.9%, leaving Grok 4.7 nearly 20 percentage points behind on that particular measure. The new model costs $2 per million input tokens and $6 per million output tokens.
SpaceXAI attributes the gains to pairing a larger foundation model with reinforcement learning deliberately tilted toward harder, multi-hour challenges. The company says the training specifically strengthened two skills essential for long-running execution: verifying its own work and managing extended context. The training run was longer than previous versions and weighted toward problems that take "many hours" to finish, though SpaceXAI hasn't disclosed whether the context improvements stem from architectural changes, summarization techniques, retrieval methods, or better retention across long sequences.
An agent operating without supervision for hours accumulates a growing record of interactions while needing to confirm each step succeeded before proceeding to the next. According to the release, SpaceXAI trained Grok 4.7 to natively understand the Grok Bot harness, tightening the connection between the model and the surrounding infrastructure that handles tool exposure, terminal response formatting, and execution feedback. That approach mirrors OpenAI's recent move to open its Codex harness as the Agents API, turning long-running agent infrastructure into a managed service. Training models around specific tool formats and execution environments could make it harder for developers to switch models without losing agent performance, a concern that grows as agents take over more of the development cycle.
Grok 4.7's 38.0% success rate on Terminal-Bench still leaves reliability as an open question, even with lower per-token costs for multi-hour runs. The company hasn't explained how it measured self-verification during reinforcement learning or published enough detail to show how much of the performance gain comes from harness-specific training versus the base model improvements. Recent benchmarks of private codebases showed even top-performing models failing more than 60% of the time, underscoring the gap between controlled tests and real-world deployment. As models and the systems around them evolve together—whether through training models to fit infrastructure or reshaping development environments to suit AI—the boundaries of what counts as "the model" continue to shift. Enterprises evaluating agentic systems will need to weigh vendor lock-in risk against performance as tooling becomes increasingly model-specific, particularly for workflows where marginal accuracy gains translate directly to hours of developer time saved or wasted.