SpaceXAI has released Grok 4.7 on Amazon Bedrock, a platform that offers foundation models from multiple providers. Grok 4.7 builds on the previous version with a larger base model and a longer reinforcement learning run on a harder mix of tasks, emphasizing sustained work and longer context management. The model exposes a reasoning effort setting with four levels—low, medium, high, and xhigh—that controls how much deliberation occurs before each response. In a benchmark, a tool‑calling agent built on Grok 4.7 was evaluated on three search‑based tasks: diagnosing a two‑service outage from alerts, scheduling maintenance under interacting constraints, and selecting an optimal set of migrations within a budget. Each task had a single correct answer verified by brute force. The agent ran six times per task at each reasoning level, totaling 72 runs. It answered 67 of 72 runs correctly. At the xhigh level, all 18 runs were correct; at medium and high, 17 of 18 were correct; at low, 15 of 18 were correct, using about one‑eighth the output tokens of xhigh. The study shows that higher reasoning effort improves accuracy but increases token usage, while low effort offers a strong cost‑to‑quality ratio. Grok 4.7 supports Bedrock’s OpenAI‑compatible Chat Completions and Responses APIs, the Bedrock Converse API, and provides a 500K token context window with text and image input and text output. The model is accessed via cross‑Region inference profiles, requiring IAM permissions for bedrock:InvokeModel on both the profile and the default project. Production use is recommended with short‑term bearer tokens derived from AWS credentials. The benchmark methodology emphasizes trustworthy measurement, including explicit grading rules, single‑answer tasks, and separate tracking of format and correctness. The results suggest that users can tune reasoning effort per request to balance accuracy and cost for agent workloads.
Build long-horizon agents with Grok 4.7 on Amazon Bedrock
SpaceXAI’s Grok 4.7 model is now available on Amazon Bedrock, offering a larger base model and longer reinforcement learning training focused on sustained tasks. The model exposes a reasoning effort setting with four levels (low, medium, high, xhigh) that controls deliberation before each response. In a benchmark, a tool‑calling agent built on Grok 4.7 completed three search‑based tasks correctly 67 of 72 times, achieving 100% accuracy at xhigh and 83% at low while using significantly fewer output tokens.
Why it matters
The ability to fine‑tune reasoning effort allows developers to balance performance and cost in agent workloads, a critical consideration as AI systems scale in production.