Grok 4.7 lands on Amazon Bedrock with a 500K-token window
Serverless on AWS across three cost tiers, with four reasoning levels and a real Coding Agent Index jump — but each task now burns ~81K output tokens.

Copy markdown
Serverless, three APIs, two region profiles
Grok 4.7 is now fully managed and serverless on Bedrock, callable through the Responses, Chat Completions, and Converse APIs — no infra to stand up. Choose the us.xai.grok-4.7 profile for data residency or global.xai.grok-4.7 for cheaper routing.
The coding jump is real, not marketing
Artificial Analysis (cited by AWS) scores Grok 4.7's Coding Agent Index at 56, up from 4.6's 47, with an Intelligence Index of 46 vs 44 and hallucinations down to 29% from 34%. Four reasoning-effort levels (low to xhigh) plus built-in self-verification are aimed at keeping long agent runs from derailing.
Watch the bill: output doubled to ~81K tokens/task
The catch — average output per task roughly doubled to ~81K tokens versus 4.6's ~38K, so the same job costs materially more at Standard rates. Push non-urgent work to the Flex tier and cap reasoning effort where you can.
Weekend build: seven tiny LLMs, all in the browser
MicroLLM Lab hit the Hacker News front page: it runs seven small quantized models fully client-side via WebGPU/WASM — no keys, no server, no data leaving the tab. A fast way to prototype private, offline, zero-cost inference for a side project.