
Grok 4.6 Landed on Amazon Bedrock, So I Built a Roast Bot
Grok 4.6 just landed on Amazon Bedrock. I built an app that roasts your code or GitHub repo in 10 languages, and the roast itself comes back translated too.
Grok 4.6 just landed on Amazon Bedrock. Launch date August 18, 2026, and the model card was already live on AWS docs the next morning. It's cheap enough to run comfortably inside a normal AWS Community Builder credit allowance, so I didn't have much of an excuse not to build something with it. What I ended up with is Grok Roast, an app that lets Grok 4.6 tear apart your code or your GitHub repo, in whichever language you want the insults delivered in.
- Live app: https://grok-roast.vercel.app
Why a roast bot
Grok has two things going for it that no other model on Bedrock has right now. Reasoning effort is exposed as four discrete levels, low through xhigh, and the model itself carries xAI's blunt, sarcastic house style. Put those together and the obvious thing to build is a code reviewer with zero filter, where the reasoning dial doubles as a savagery dial. Low effort gets you a one-line jab. Xhigh gets you a full teardown of your design decisions.

Finding out Converse wasn't the right door
My first attempt was the obvious one, straight through the Converse API.
aws bedrock-runtime converse \ --region us-east-1 \ --model-id "us.xai.grok-4.6" \ --messages '[{"role":"user","content":[{"text":"Introduce yourself in one line."}]}]'An error occurred (ValidationException) when calling the Converse operation: The provided model identifier is invalid.list-foundation-models didn't show xAI as a provider at all in that account, in any region. My first instinct was that this needed the old Model Access opt-in flow, but that turned out to be the wrong theory. Recent launches like Claude Opus 5 and GLM 5 show up in the catalog without any manual access request, so Bedrock had clearly moved on from that gate for most models.The real answer was that Grok 4.6 doesn't live behind
bedrock-runtime at all. It's served from a separate endpoint called bedrock-mantle , an OpenAI-compatible Responses API, and reasoning effort is only configurable there.curl -sS -i -X POST "https://bedrock-mantle.us-west-2.api.aws/openai/v1/responses" \ -H "Content-Type: application/json" \ -d '{"model":"xai.grok-4.6","input":"hi"}'HTTP/2 401{"error":{"code":"invalid_api_key","message":"Missing 'authorization' or 'x-api-key' header","param":null,"type":"permission_denied_error"}}A 401 instead of a 404 told me the service existed and just wanted credentials. AWS's own sample code only documents the console-issued long-term Bedrock API key for this endpoint, but I tried plain IAM SigV4 signing first out of curiosity, and it worked on the first try.
from botocore.auth import SigV4Authfrom botocore.awsrequest import AWSRequest
req = AWSRequest(method="POST", url=url, data=body, headers={"Content-Type": "application/json"})SigV4Auth(creds, "bedrock", "us-west-2").add_auth(req)The trick is signing with service name
"bedrock" even though the hostname says bedrock-mantle . No console trip needed, just a regular IAM access key. The app itself ports the same logic to TypeScript with @smithy/signature-v4 .const signer = new SignatureV4({ credentials: { accessKeyId, secretAccessKey }, region: "us-west-2", service: "bedrock", sha256: Sha256,});A second permission I didn't know existed
Local testing worked fine, so I deployed to Vercel and hit a wall immediately.
User: arn:aws:iam::xxxxxxxxxxxx:user/grok-roast-vercel is not authorized to perform:bedrock-mantle:CreateInference on resource: arn:aws:bedrock-mantle:us-west-2:xxxxxxxxxxxx:project/defaultbecause no identity-based policy allows the bedrock-mantle:CreateInference actionMy IAM policy only granted
bedrock:InvokeModel . Turns out bedrock-mantle runs on its own action namespace entirely, bedrock-mantle:CreateInference , which nothing in the docs I'd read mentioned. I only found it by deploying and reading the actual error. Adding one line to the policy fixed it without a redeploy.{ "Effect": "Allow", "Action": [ "bedrock:InvokeModel", "bedrock:InvokeModelWithResponseStream", "bedrock-mantle:CreateInference" ], "Resource": "*"}What the roast actually looks like
I fed it this on purpose:
function add(a,b){var x=a;var x=b;return x+x}At xhigh effort, here's what came back.
It calls itselfaddand then just returnsb + b. The momentagets passed in, its purpose in life is already over.This isn't addition, it's a function that doubles whatever value arrived last. It fully exploits JS letting you redeclarevar, which quietly kills the first assignment. If that was intentional, genius. If not, an ordinary bug.Dressed up as an addition function, this is really "b gets doubled, a gets ignored" wearing a fake name tag.
It caught the
var redeclaration bug and made fun of it in the same breath. I also built a GitHub repo mode, and pointing it at one of my own projects (a quantum fortune-telling app) got a full architectural teardown, from the fake randomness theater to an unauthenticated public Lambda Function URL. Raising the reasoning effort visibly raises the resolution of what it notices, not just the length of the response.The multilingual part is the part I actually care about
The UI supports ten languages: Japanese, English, Chinese, Korean, Spanish, French, German, Russian, Portuguese, and Arabic. The easy part is translating button labels. What I wanted instead was for the roast text itself, the actual model output, to come back written natively in whichever language the user picked, not just the interface around it.
The system prompt stays in English, and I append one line telling it what language to answer in.
const system = `${SYSTEM_PROMPT}
Respond entirely in ${languageName}, using the tone and idiom a native speaker would actually use for savage-but-fair humor.`;Switching to French and pointing it at the same repo got me this:
On va pas se mentir : le README a le culot d'un prospectus IonQ, le dépôt a l'âme d'unnp.random.choice(p=[0.17, 0.35, …])habillé en pèlerinage shinto.
That's not a translated version of an English joke, it reads like something written directly in French idiom. Arabic flips the whole layout to RTL automatically when selected.
English-MEDIUM

Arabic-MEDIUM

Result

The tone in the English version was too harsh, which resulted in an error for this article.
One thing that got caught before publishing
While picking a thumbnail image, I almost used an AI-generated photo that made a real public figure look like they were shaking hands at an actual AWS event that never happened. It got flagged as something that could pass for a real photo and mislead people, so I swapped it for an obviously illustrated version instead. Small thing, but worth mentioning since it's the kind of mistake that's easy to wave through when you're just trying to finish a blog post.
Next thing I want to try is surfacing the xhigh reasoning summary in the UI, so you can watch it build the case against your code before it delivers the verdict.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article