Mechanical Turk closed. Here’s where human review goes next.

Amazon closed Mechanical Turk on September 30, 2026, and its workers left SageMaker Ground Truth and Augmented AI with it. If your models, pipelines or agents sent work there, here’s what changed and how to replace it.

What changed

September 30, 2026
Mechanical Turk closed. Unsubmitted work expired, and the Mechanical Turk workforce stopped being available in Amazon SageMaker Ground Truth and Amazon Augmented AI (A2I).
October 30, 2026
Last day for requesters to approve or reject submitted work. Anything left is approved automatically.
January 28, 2027
Transaction history is no longer available. Export what you need before then.

Pick the replacement for the job

Labeling data to train a model, in bulk

A training data vendor or an annotation tool

Built for large batch projects with managed annotators and review stages.

Surveys and behavioral research

A research panel such as Prolific

Built for studies, with participant screening and fair pay rules.

A person checking an AI decision before it acts

Plures

One API call returns a live, consensus answer from ID verified people, with disagreements sent to your team.

Moving a review loop to Plures

On Mechanical TurkOn Plures
A HITA task: a question, 2 to 8 options, and the text, image, recording or video to judge
Assignments per HITverifiers: 1, 3, 5, 7 or 9 people. If they split, Plures asks more, up to max_verifiers
Qualifications and MastersEveryone passes a government ID and selfie check and is scored on hidden known answer tasks
Approving or rejecting workNot needed. Answers are combined by consensus, and anything short of agreement can go to your own team
Polling for resultsA signed webhook when it's decided, or poll the task
Reward plus Amazon's fee$0.20 per answer from prepaid credit, more for recordings and videos by length. Test mode is free
An A2I human loopOne call from your code, an SDK, or the MCP tool inside an agent
from plures import Plures

plures = Plures()  # reads PLURES_API_KEY; a pk_test_ key is free and instant

task = plures.verify(
    question="Does this invoice match the purchase order?",
    context="PO 4471: 40 units at $12.50. Invoice: 40 units at $15.00.",
    options=["Matches", "Does not match"],
    verifiers=3,
)
print(task["result"]["answer"], task["result"]["agreement"])

Install with pip install plures or npm install plures. Agents can call it through the MCP server. Full details in the docs.

Questions

Is Plures a drop in replacement for the Mechanical Turk API?

No. Plures is a smaller API built for one job: a person checking an AI decision. Most review loops move over by replacing the HIT creation and result polling with one call to create a task and a webhook handler. Start with a free test key and move one flow at a time.

Can I still use SageMaker Ground Truth and Augmented AI?

Yes. Amazon closed only the Mechanical Turk workforce in them. You can keep using them with a private workforce of your own people or a vendor workforce. Plures is an option when you want a live answer from people you don't have to hire.

Is Plures more expensive than Mechanical Turk was?

Per answer, yes. Mechanical Turk work often cost a few cents an item. Plures is $0.20 per answer, so $0.60 for a decision with 3 people who agree, because every person is ID verified, scored, and confirmed with Face ID or a passkey, answers come back live, and every decision is recorded on a ledger. It's built for decisions where a wrong call costs more than that.

Can my Mechanical Turk workers join Plures?

Anyone 18 or older can apply to be a Plures verifier at plures.si/earn. They pass an ID and selfie check before their first task.