Grok 4.7 Targets Longer Agent Tasks With New Training and Safety Claims

Grok 4.7 Targets Longer Agent Tasks With New Training and Safety Claims

September 22, 2026

SpaceX has introduced Grok 4.7, a new model aimed at complex coding and agent workflows. The company says it trained the system on harder tasks for longer, improved its reinforcement-learning process, and added safeguards for sensitive cybersecurity and biology requests. The announcement arrives as AI vendors compete on more than benchmark scores: models increasingly need to plan, use tools, and stay reliable over extended jobs.

SiliconANGLE reports that SpaceX tested Grok 4.7 on CursorBench 4.0, a coding benchmark created by Cursor, which SpaceX recently acquired. The company reported an average task cost of $4.69 and said the model compared favorably with selected competing systems. On other evaluations, the reported picture was mixed: Grok 4.7 exceeded Anthropic’s model on two named tests but trailed OpenAI’s latest model on a chip-design benchmark.

Those results should be read as vendor-reported comparisons. Benchmark results can change with model settings, tools, hardware, task selection, and whether the comparison prioritizes answer quality or cost. The use of a benchmark connected to an acquired company also makes transparent methodology especially useful. Independent replication would help buyers understand how representative the scores are.

Why longer tasks matter

An agent that can answer a question in one turn is different from one that must inspect files, write code, run tests, correct errors, and summarize what changed. Each additional step creates chances to lose context, make a mistaken assumption, or take an unintended action. Longer-horizon performance therefore depends on planning, memory, tool reliability, and verification—not just fluent text generation.

SpaceX says Grok 4.7 is designed to work with its multi-agent Grok Bot harness, which can divide a larger task among agents and check their work. This approach can speed up parallelizable jobs, but it can also multiply errors if agents share faulty assumptions or if no one checks the final result. Teams should evaluate completed outcomes and audit trails, not simply the number of agents involved.

Safety claims need task-level tests

The company says the new model performed strongly on evaluations related to harmful biology and cybersecurity requests. Such safeguards matter because a model capable of writing code or reasoning through technical tasks can be useful to defenders and potentially misused. A high benchmark score does not settle how the model behaves in real interactions, especially when users combine instructions, external data, and tools.

Organizations considering Grok 4.7 should test it with their own permitted workloads, verify access controls, and define which actions require human approval. For software agents, a safe workflow restricts credentials and tools to the minimum required and records the actions the model takes. These operational controls remain important even when a model provider reports improved safeguards.

What buyers should watch

SpaceX lists API pricing from $2 per million input tokens and $6 per million output tokens, with a faster option at a higher price. Total cost depends on how many retries, tool calls, and verification steps a task requires. A cheap token rate can still produce an expensive workflow if the system loops or returns unusable work.

The most useful next evidence would include reproducible benchmark settings, performance on independent long-running tasks, measured failure rates, and clear documentation of safety testing. Grok 4.7’s release marks another step in the competition to make AI agents practical. The real test is whether developers can trust it to finish complex jobs accurately, safely, and at a predictable cost.

Sources

• SiliconANGLE: SpaceX launches Grok 4.7
• xAI: Grok
• NIST AI Risk Management Framework