xAI's Grok 4.7 lands on Amazon Bedrock

A 500K-token window, four reasoning tiers. The top one roughly doubles output tokens per task.

xAI's Grok 4.7 is now on Amazon Bedrock. AWS customers get what xAI calls its most capable model yet for coding and knowledge work. The model ships with a 500K token context window and four reasoning effort levels — low, medium, high and xhigh. It is served through cross-Region inference profiles rather than a bare model ID. It supports the Responses, Chat Completions, InvokeModel and Converse APIs. Teams already using OpenAI-compatible clients or AWS's own SDK can both reach it.

xAI says the model was trained on a new, larger base. The reinforcement learning run was longer, too, and weighted toward tasks that take many hours to complete. The result, xAI says, is a model that verifies its own output before moving on and makes better use of long context. That self-checking behaviour matters most for agents. A model that catches its own mistakes mid-task tends not to compound an early error through dozens of later steps. xAI also reports a rebuilt safeguard stack. It says the stack allows only a small fraction of risky dual-use cyber security prompts through, while rarely blocking legitimate work. It has begun giving selected security partners invite-only red-team access.

Artificial Analysis, which runs independent evaluations rather than taking vendor numbers at face value, reports Grok 4.7 improving across its suite. Its Intelligence Index rose to 46 from 44 for Grok 4.6. Its Coding Agent Index rose to 56 from 47. The largest single gains came on long-horizon agentic knowledge work: an Elo of 1,657 on its AA-Briefcase benchmark, up from 1,546. Hallucination rates fell too, to 29% on AA-Omniscience from 34%. But the same evaluation puts output tokens per Intelligence Index task at roughly 81,000. That is more than double Grok 4.6's roughly 38,000, all measured at the xhigh reasoning setting.

On Bedrock, the model supports implicit prompt caching for repeated prompt prefixes. It also gets Guardrails for content filtering and PII redaction, structured JSON Schema outputs, and CloudWatch invocation logging that captures reasoning token counts — useful for auditing long, unattended agent runs. Two inference profiles are on offer. A Global profile, priced lower, routes requests to any supported commercial Region. A US geographic profile serves data-residency or latency-sensitive traffic instead. Pricing runs on three service tiers — Standard, Priority and Flex — set per request.

Written from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.