SpaceXAI has released Grok 4.7, a new frontier model built for coding, agentic tasks, and professional knowledge work. The model uses a larger base than Grok 4.6, received longer reinforcement learning on multi-hour tasks, and was trained to work natively with the Grok Bot agent harness.
Grok 4.7 supports a 500,000-token context window, text and image input, function calling, code execution, and configurable reasoning effort. API pricing for prompts below 200,000 tokens starts at ＄2 per million input tokens, ＄0.50 per million cached tokens, and ＄6 per million output tokens. Rates double above that threshold.
SpaceXAI reports meaningful gains over Grok 4.6:
- CursorBench 4.0: 46.3% versus 40.4%
- DeepSWE v1.1: 71.0% at high reasoning effort versus 65.2%
- Terminal-Bench 4.0: 38.0% versus 20.3%
- AA Briefcase v1.1: 1,657 versus 1,546
These are vendor-reported results, but the direction is notable: the largest improvement appears in long-running terminal work, exactly where agent reliability often breaks down.
Why it matters
For builders, Grok 4.7 is less about another benchmark lead and more about longer-horizon execution at accessible API pricing. The model is available through the Grok API, Grok Build, Cursor, model routers, and a gradual GitHub Copilot rollout—making it immediately testable across several production agent harnesses.
The harness-aware training also connects directly to the always-on teammate architecture SpaceXAI introduced with Grok Bot: model capability and the surrounding execution system are increasingly being trained as one product rather than assembled independently.
In the Responses API, encrypted reasoning is returned automatically. Applications can pass that encrypted state back on later calls, preserving useful reasoning context without exposing the underlying reasoning text.