---
title: 'Why Every Frontier Lab Uses Our Webhook Standard'
authors: ['abhi']
date: 2026-11-18T12:00:00
tags: ['technical', 'ai', 'webhooks']
summary: 'OpenAI, Anthropic, Google, and xAI all sign webhooks with Standard Webhooks. We go over what each lab ships, why none of them invented a signature scheme, and where they still differ.'
---

OpenAI, Anthropic, Google, and xAI all sign their webhooks with [Standard Webhooks](https://www.standardwebhooks.com/), the open spec Svix started in 2023. AI jobs now run for minutes or hours, so polling for results stopped making sense. That pushed the frontier labs to adopt webhooks, and specifically, Standard Webhooks.

When my CEO, Tom, [announced the spec](/standard-webhooks/), he wrote that it could "do to webhooks what JWT did to API authentication." The committee included members from Zapier, Twilio, Lob, Mux, ngrok, Supabase, and Kong. Three years later, the four labs with the most watched APIs in software have all landed on it. All of the work for Standard Webhooks is developed in a [public repo](https://github.com/standard-webhooks/standard-webhooks).

## What each lab ships

| Lab       | What uses Standard Webhooks                                        | Docs                                                                                                                                                                                     |
| --------- | ------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| OpenAI    | Background responses, batch jobs, fine-tuning jobs, agent sessions | [Webhooks guide](https://developers.openai.com/api/docs/guides/webhooks)                                                                                                                 |
| Google    | Gemini API batch jobs, Deep Research, video generation             | [Gemini API webhooks](https://aistudio.google.com/docs/webhooks)                                                                                                                         |
| Anthropic | Managed Agents session and vault events, organization webhooks     | [Managed Agents webhooks](https://platform.claude.com/docs/en/managed-agents/webhooks), [organization webhooks](https://platform.claude.com/docs/en/manage-claude/organization-webhooks) |
| xAI       | Webhook triggers that start Grok automations                       | [Webhook triggers](https://docs.x.ai/grok/automations/webhooks)                                                                                                                          |

OpenAI [went first](https://community.openai.com/t/deep-research-in-the-api-webhooks-and-web-search-with-o3/1299919), in June 2025. Its guide lets you verify with the OpenAI SDK or just grab any Standard Webhooks library. Google followed in May 2026, and the [launch post](https://blog.google/innovation-and-ai/technology/developers-tools/event-driven-webhooks/) says the Gemini implementation strictly adheres to the spec.

Anthropic [announced webhooks](https://claude.com/resources/articles/new-in-claude-managed-agents) for Managed Agents that same week, and you can see the adoption right in the code. `unwrap()` in the [TypeScript SDK](https://github.com/anthropics/anthropic-sdk-typescript/blob/main/src/resources/beta/webhooks.ts) is the `standardwebhooks` package with a JSON parse on the end:

```ts
const wh = new Webhook(keyStr)
wh.verify(body, headers)
return JSON.parse(body) as BetaWebhookEvent
```

xAI is the odd one out, in a good way. Grok is the receiver: you sign the request that kicks off an automation, and Grok verifies it with the Standard Webhooks signature format.

## Long jobs killed polling

You've heard this one enough times already, but the reason is agents. A chat completion comes back in seconds, so for years AI APIs didn't need webhooks at all. But once the requests started including deep research, batch jobs over thousands of prompts, video generation, and long running interactive agent sessions, something had to give.

OpenAI shipped webhooks in the same release as deep research in the API. Google's launch post talks about operations that take minutes or hours, and developers calling `GET` in a loop to check on them. Anthropic's webhooks tell you when an agent session has gone idle and is waiting on you.

Tom argued in [The Future is Event-Driven](/future-is-event-driven/) that agents make polling more expensive for both sides. Unsurprisingly, the labs felt the polling load and likely decided that it was worth acting on. This issue isn't isolated to the labs, though; it generalizes to the greater AI ecosystem and applies to anyone building an API that agents will poll. It's only a matter of time before webhooks (or a more AI-native primitive) become the obvious standard for any API an agent touches.

## So why are they not rolling their own?

None of the labs has written up why they picked the spec, so this is mostly speculative.

Signing a webhook is not particularly hard; you can HMAC a body and get a v1 out pretty fast. What's hard is everything around it: replay attacks, idempotency, [zero downtime secret rotation](/zero-downtime-secret-rotation-webhooks/), and the rest of the [ways signature schemes break](/common-failure-modes-for-webhook-signatures/).

Standard Webhooks handles all of that, and it saves the labs' customers work too. A lab that rolls its own scheme makes every customer write and test one more verifier, and plenty of those customers call more than one lab. A lab that adopts the spec gets working libraries, a [signature debugger](https://www.standardwebhooks.com/verify), and built-in support in gateways like ngrok and Kong. More and more, that receiver is written by a coding agent anyway, and a single [receiving-webhooks skill](/introducing-webhooks-skills-for-agents/) covers every sender on this list.

Also, this is what they chose starting clean. These are new companies shipping new products, with engineers hired from the best companies in the world. Stripe and GitHub picked their signature schemes years before the spec existed, and changing them now would break every customer's verifier. It's not an unfair interpretation to suggest that Standard Webhooks is what every top engineer would start with if they weren't burdened with legacy.

## Read each lab's docs

The labs agree on the signature and not much else. OpenAI retries for up to 72 hours, Gemini for 24. Anthropic makes up to three attempts within a few minutes for Managed Agents events, then drops the event and expects you to reconcile through the API.

## Send what the labs send

The labs went event-driven first because they got polled first. If agents call your API, you're next. When you get there, sending Standard Webhooks means your customers verify you with the same libraries, and the same few lines that already work for OpenAI, Anthropic, and Gemini.

The spec is free to adopt. It's [open source](https://github.com/standard-webhooks/standard-webhooks), and you don't need Svix to use it.

From there it's the usual build vs. buy. While the signature is fairly cheap, making everything durable for every one of your customers is a hard problem, and that's what [Svix](https://www.svix.com/) does.
