◉CLAUDE LABJP
●WATERMARK — Since 11/16, text from Opus 4.6, Sonnet 4.5 and other models carries a watermark; non-default temperature and similar settings may drop it●CC 2.1.325 — Claude Code is still on 2.1.325 (11/13), with no newer release yet●SONNET 4.5 — Retires on the Claude API on 11/30, 13 days left. Sweep model IDs in code and saved settings in one pass●AUTOMATION — A user reports recurring AI jobs set up without run counts or a spending cap, burning through a weekly allowance●CONNECTOR — If connectors vanish only in new sessions, compare the /mcp list against an older session first●NEW — Cowork Feels Worse? Run the Same Request in Three Fresh Sessions●WATERMARK — Since 11/16, text from Opus 4.6, Sonnet 4.5 and other models carries a watermark; non-default temperature and similar settings may drop it●CC 2.1.325 — Claude Code is still on 2.1.325 (11/13), with no newer release yet●SONNET 4.5 — Retires on the Claude API on 11/30, 13 days left. Sweep model IDs in code and saved settings in one pass●AUTOMATION — A user reports recurring AI jobs set up without run counts or a spending cap, burning through a weekly allowance●CONNECTOR — If connectors vanish only in new sessions, compare the /mcp list against an older session first●NEW — Cowork Feels Worse? Run the Same Request in Three Fresh Sessions
Articles/API & SDK
⬡ API & SDK/2026-11-17Beginner

If You Rely on Claude's Text Watermark, Check temperature in the Request Body, Not Your Code

Since November 16, text from Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 carries Anthropic's text watermark, but non-default temperature, top_p, or top_k may drop it. Here is how I log the request body once and decide.

Claude API147Text watermarktemperature3Python SDK3TypeScript SDK

On the afternoon of November 17 I opened the Platform release notes and found one short new entry dated the day before. Output from Claude Opus 4.6, Opus 4.5, Sonnet 4.5, and Haiku 4.5 now carries Anthropic's text watermark on supported platforms.

The next sentence is where I stopped. If a request to one of these models sets temperature, top_p, or top_k to a non-default value, the output might not carry the watermark.

Back in October I had already stripped sampling parameters out of my own code to get ready for API version 2026-10-21, so my first thought was that I was fine. Then I read the Text watermark page, which adds one more line: your framework or client library might set these parameters for you. Not having written them doesn't mean they aren't being sent.

Most people don't need to do anything

The Text watermark page says the watermark doesn't change the meaning, quality, or readability of the output, adds no tokens or hidden characters, and carries no information about you or your organization. You don't need to change your requests or response handling.

The only case that needs work is when you rely on the output carrying the watermark — say, a team rule that generated text must stay traceable, or a wish to leave a provenance hint in something you ship. For me, as an indie developer, it's the drafts I make with Claude when localizing app store listings: I want to be able to trace later which text went through Claude.

Here is the order I decide in.

SituationWhat to do
You don't rely on the watermarkNothing (the docs say no change is needed)
You rely on it, use a covered model, and send one of the threeStop sending it
You rely on it and aren't sure what's sentLog the request body once, as below, then decide
You send API version 2026-10-21The three return a 400 even at default values, so they can't be going out

The Sampling settings section lists Opus 4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5, and Opus 3 (access on request). Sonnet 4.6 got the same caveat in the November 3 notes. Sonnet 4.5 also leaves the Claude API on November 30, which is 13 days away, so it's worth checking whether the model you move to is covered by that section too.

grep can't show you what was sent

When I removed sampling parameters in October, I searched the source and deleted every temperature= line. I wrote that up in Before upgrading the SDK turns temperature into a 400, I swept my call sites.

But a search only shows what you wrote. A framework, a shared in-house wrapper, a config loader — any of these layers can inject a default, and the body carries a number while the source stays clean. I think that gap is exactly why the docs tell you to check the request body.

So my rule became: check what you sent, not what you wrote.

Python SDK: one line from an httpx event hook

The official Python SDK accepts an httpx client via http_client. I add a hook that runs before each request and prints only those three keys from Messages API calls. The prompt itself never hits the log.

import json
 
import anthropic
import httpx
 
SAMPLING_KEYS = ("temperature", "top_p", "top_k")
 
 
def log_sampling(request: httpx.Request) -> None:
    # Only look at Messages API calls (check count_tokens and batches separately)
    if not request.url.path.endswith("/v1/messages"):
        return
    try:
        body = json.loads(request.content or b"{}")
    except (ValueError, httpx.RequestNotRead):
        print("[sampling] a request body could not be read")
        return
    found = {k: body[k] for k in SAMPLING_KEYS if k in body}
    print(f"[sampling] model={body.get('model')} {found or 'none'}")
 
 
# The API key is read from the ANTHROPIC_API_KEY environment variable
client = anthropic.Anthropic(
    http_client=httpx.Client(event_hooks={"request": [log_sampling]}),
)
 
client.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=64,
    messages=[{"role": "user", "content": "Reply with one word."}],
)

If your framework lets you pass in the client it uses, attach the same hook there. If it doesn't, the next best thing is to read that framework's settings for the three defaults.

The hook only prints. Rewriting values inside it would hide where they came from; when something shows up, I fix the layer that injects it.

TypeScript SDK: wrap fetch and do the same

The TypeScript SDK lets you swap in your own fetch. Same idea: read only the three keys from the body.

import Anthropic from "@anthropic-ai/sdk";
 
const KEYS = ["temperature", "top_p", "top_k"] as const;
 
const client = new Anthropic({
  fetch: async (url, init) => {
    if (String(url).endsWith("/v1/messages") && typeof init?.body === "string") {
      const body = JSON.parse(init.body);
      const found = KEYS.filter((k) => k in body).map((k) => `${k}=${body[k]}`);
      console.log(`[sampling] model=${body.model} ${found.join(" ") || "none"}`);
    }
    return fetch(url, init);
  },
});

Neither is meant to stay in production. Run each watermark-dependent job through it once, read the log, and take it out.

The line I drew after trying it

I attached the hook and ran each drafting job I want watermarked once. In my setup none of the three keys appeared — but only because I had removed them in October. If you have a framework in the middle, your result may well differ.

I did notice something else, though. The jobs that rely on the watermark and the ones that don't were sharing the same client. Don't send the parameters in one, use them freely in the other — if that split isn't visible in the code, the day someone helpfully adds temperature, the assumption breaks without anyone noticing.

Now the watermark-dependent calls live in one function that simply doesn't accept those three arguments. Make the calls that rely on the watermark unable to send them. An assumption kept by shape seems to last longer than one kept by care.

For the difference between the text watermark and the provenance metadata attached to files, see What to Settle Before Watermarked Claude Text Ships Inside Your App. For counting Sonnet 4.5 callers, see 43 Days Until Sonnet 4.5 Retires: I Counted Callers From the Usage CSV, Not From grep.

What to do next

If even one job relies on the watermark, attach the hook to that job alone and run it once. If the log says "none," nothing in today's notice changes for you. If a number shows up, trace back through the layers until you find the one that sets it.

I started with the job that drafts my store listings. Having looked at what I actually sent, once, makes the next release note much calmer to read.

Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • ✦Copy-paste ready implementation code
  • ✦New advanced guides published daily
  • ✦$5/mo or $10 for lifetime access
View Membership →

If you found this article helpful, a small tip ($1.50) would mean a lot to us. Your support helps keep this site ad-free and covers server and hosting costs.

Related Articles

⬡ API & SDK2026-10-27
Three Places Python SDK 2.0.0 Breaks Your Code — A Checking Order Built on a Mocked Transport
Upgrading to Python SDK 2.0.0 changes behavior in three places: the temperature argument, the auth header, and double credentials. Here are the outputs I reproduced on a mocked transport, the fix for proxy setups, and the order I check in.
⬡ API & SDK2026-10-25
Before upgrading the SDK turns temperature into a 400, I swept my call sites
With API version 2026-10-21, sending temperature, top_p, or top_k returns a 400 even at the default value. Here is how the switch happens through an SDK upgrade, a small scanner for the lines to fix first, and the order I'd fix them in.
⬡ API & SDK2026-10-20
Moving Structured Outputs to output_config.format — What to Check Before You Drop the Beta Header
A checklist for moving structured outputs from output_format and the beta header to output_config.format, drawn from the official docs: the Python SDK TypeError, prompt-cache invalidation, schema limits, and the cases where output won't match your schema.