One request replaces the HIT
No template, no hosting and no approval queue. Send the items, the question and the votes per item, and read partial results while the job runs.
Amazon Mechanical Turk alternative
Amazon closed Mechanical Turk. Move your tasks to Datapoint.
Amazon Mechanical Turk stopped accepting tasks on September 30, 2026, and the MTurk workforce was removed from SageMaker Ground Truth and Amazon Augmented AI the same day. Datapoint collects the same human judgments through one API: comparisons, ratings, rankings and multiple choice on image, video, audio and text, with first responses in minutes.
Thinking Machines
Amazon Mechanical Turk alternative
What you did on MTurk, without the HIT.
Mechanical Turk gave you a marketplace of workers and left the task, the review and the aggregation to you. Datapoint takes the items and the question and returns the answer.
| When you need |  | Amazon Mechanical Turk |
|---|---|---|
| A service that is running | Yes: Open to new accounts, with signup credits | No: Closed on September 30, 2026 |
| No task page to build | Yes: One API request: media, question, votes per item | Partly: You built each HIT from a template or your own HTML |
| Workers verified as real, unique people | Yes: Every rater, before their first task | Partly: An open marketplace that you screened with qualifications |
| Quality checks without extra setup | Yes: Timing, attention, honeypot and bias checks on every answer | Partly: You reviewed the work and approved or rejected each assignment |
| Agreement scores on every label | Yes: Confidence, agreement and a reliability-weighted verdict | No: Raw assignments that you aggregated yourself |
| A price you can state before the job runs | Yes: One flat rate per response, quoted up front | Partly: A reward you set per assignment, plus Amazon’s fee |
| Image, video and audio tasks | Yes: Hosted for you, with control over how the media plays | Partly: Possible in a custom HIT that you built and hosted |
| Raters by country and demographic | Yes: Country, region, city, age and gender filters, plus screening questions | Partly: Location and premium qualifications |
| Open-ended work such as transcription or web research | No: Not the right fit | Yes: A common use of the marketplace |
Amazon Mechanical Turk closed on September 30, 2026. Details are from mturk.com and the AWS documentation, checked November 2026.
The difference
MTurk gave you a crowd. Datapoint gives you the answer.
On Mechanical Turk you were the requester. You designed the HIT, set a reward, picked qualifications, then approved or rejected each assignment and aggregated the results yourself. Amazon ended that on September 30, 2026. Transaction history stays available until January 28, 2027, so export what you need before then.
On Datapoint one API request replaces the HIT. You send the media, the question and the number of votes you want per item. Raters are everyday people reached inside apps they already use, and each one is verified as a real, unique person.
Each item comes back as a verdict with its agreement score, a reliability-weighted version and every vote behind it. The review and aggregation steps you ran on MTurk are already done.
GET /jobs/{id}/resultsOne comparison, ten votes{
"datapoint_index": 0,
"votes": { "A": 7, "B": 3 },
"total_responses": 10,
"consensus": "A",
"confidence": 0.7,
"agreement_rate": 0.7,
"weighted_consensus": "A",
"weighted_confidence": 0.718
}Every verdict comes with its agreement score, a reliability-weighted version and the raw votes behind it.
Why teams switch
Why teams moving off MTurk choose Datapoint
It is running, there is less to build, and the quality checks come with it.
Quality is checked for you
MTurk left review to the requester. Datapoint verifies every rater, runs timing, attention, honeypot and bias checks on every answer, and weights votes by rater reliability.
Built for model evaluation
Pairwise comparisons, ratings and rankings of image, video, audio and text outputs, returned as chosen and rejected pairs with vote margins for RLHF, DPO and leaderboards.
Proof
Built for training volume.
The vote records behind the DP-Bench image, video and speech leaderboards are public on Hugging Face, including 2M+ text-to-image votes.
10,000+human judgments per minute
1B+people reachable through apps
200+countries
FAQ
Datapoint vs Amazon Mechanical Turk, answered.
Yes. Amazon closed Mechanical Turk on September 30, 2026. It stopped accepting new customers on July 30, 2026, and announced the full closure in August. The MTurk site now says the service is no longer available.
It depends on the task. For human judgments of AI outputs, such as comparing, rating or ranking images, video, audio and text, Datapoint is an API that returns scored results in minutes. For academic surveys and behavioural experiments, a research panel such as Prolific is the usual choice. For expert annotation against a detailed rubric, use a managed annotation service.
Yes, if the HIT asked people to compare, rate, rank, choose from options or give a short text answer. Those map directly to Datapoint task types: send the same items through the API with the question and the number of votes per item. Open-ended work such as transcription or web research is not a fit.
The Mechanical Turk worker type is no longer available for labeling jobs in SageMaker Ground Truth or for human review workflows in Amazon Augmented AI as of September 30, 2026.
Amazon says transaction history remains available until January 28, 2027. Requesters could approve or reject submitted HITs until October 30, 2026, and prepaid balances were due to be refunded within 30 days of the closure.
Everyday people in 200+ countries, reached inside apps through direct partnerships with app developers and an in-app ad network. Each one is verified as a real, unique person, and every answer passes behaviour, attention, honeypot and bias checks before it counts.
On MTurk the requester handled quality: qualifications, review and approval or rejection of each assignment. Datapoint builds the checks into every job, weights votes by rater reliability and reports agreement on every verdict, with stable anonymized rater IDs so you can audit it yourself.
Datapoint delivers 10,000+ judgments a minute across its network. First responses typically arrive within minutes, and you can read partial results while a job runs.
Yes. Filter by country, region, city, postal area, US area income, gender or age range, and add a screening question for anything else, such as language or profession. Screened-out raters are never billed.
You pay per response collected. Targeting adds a per-response surcharge that is fixed when the job is created, sandbox runs and screened-out raters cost nothing, and new accounts get signup credits. Volume pricing is set on a call.
Send a HIT you used to run. Get it back as an API job.
Bring the task, the number of items and the votes per item. We come back with the setup, the cost and the turnaround within a working day.