SpaceXAI has released Grok 4.7, targeting coding and professional knowledge work at half the operating cost of rival systems.
The deployment matches the execution speed and API pricing of its direct predecessor, Grok 4.6, billing developers at $2 per million input tokens and $6 per million output tokens. For latency-sensitive production environments, SpaceXAI provides an expedited variant that delivers double the output generation speed at twice the standard price.
Detailing the architecture in its technical release, SpaceXAI said: “Grok 4.7 is our most capable model for coding and knowledge work. It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.”
Reinforcement training drives multi-hour reasoning performance
The system relies on an expanded base architecture trained through extended reinforcement learning runs. Machine learning teams focused this compute on complex multi-hour problem sets, expanding context handling and native self-verification mechanisms. Engineers also integrated direct compatibility with the Grok Bot harness, refining conversational dialogue and unstructured knowledge processing across enterprise environments.
Benchmark evaluations reflect these computational trade-offs across specialised domains. On CursorBench 4.0, an evaluation measuring sustained programming assignments, Grok 4.7 recorded a 46.3 percent success mark at an average cost of $11.95 per completed task. That result outpaced Grok 4.6 at 40.4 percent and GPT-5.6 Sol at 41.7 percent, while trailing Fable 5.1 at 51.8 percent.
On DeepSWE v1.1, the model achieved 71 percent under high-effort parameters, beating both Grok 4.6 at 65.2 percent and Fable 5.1 at 70 percent. Terminal-Bench 4.0 scores rose to 38 percent from 20.3 percent in the prior generation.
Specialised professional evaluations showed similar patterns. Grok 4.7 posted an Elo rating of 1695 on the GDPval benchmark, finishing behind Fable 5.1 at 1735 Elo but ahead of Grok 4.6 at 1605 Elo and GPT-6 Astra at 1542 Elo. The model scored 64.0 percent on the EEBench electrical engineering test, surpassing GPT-5.6 Sol at 39.4 percent.
On the Harvey Legal Agent Benchmark, the system reached 19.6 percent, exceeding Grok 4.6 at 15.8 percent, Fable 5.1 at 6.7 percent, and GPT-5.6 Sol at 2.5 percent. Clinical evaluations on HealthBench Professional placed the system at 56.7 percent, marking progress over Grok 4.6 at 48.5 percent while finishing below GPT-5.6 Sol at 60.5 percent and Fable 5.1 at 62.1 percent.
Hardened security stack curtails dual-use hazards
SpaceXAI rebuilt its defensive filtering architecture to address dual-use risks without impeding authorised technical exploration. The model leads LatchBio’s biological safety benchmark with a 62.4 percent rating by combining benign query utility with refusal mechanisms for biological hazards.
On HackerBench v0.3, an evaluation covering destructive cyber operations, the software restricted unauthorised completions to 3.3 percent of risky prompts while permitting legitimate security workflows. Selected cybersecurity partners have received restricted, invite-only access to evaluate red-team capabilities for defensive research.
General distribution has commenced across commercial developer pipelines. Teams can query the model through the Grok API, third-party model routers, and managed cloud environments. Integrated development environments including Cursor and Grok Build have enabled live access immediately, alongside terminal installers distributed via standard shell environments.
See also: Google’s Android Bench 2.0 tests AI models on complex tasks
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
Developer is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.