A subscriber wrote to me last week after a bad month. His task queue had stopped refilling, a project he had onboarded to paused without explanation, and the assignment he did get had nothing to do with the expertise on his resume. His summary of the platform was three words long: merciless and messy.

Then he asked the question I now get several times a week, and the one you may have a tab open for right now. Should he go find a new platform?

The instinct behind it is right and the conclusion is wrong. Concentration is genuinely dangerous, and I have written before that no single project should carry more than about sixty percent of your income. But the fix is not more platforms. For most people the fix is two platforms and much harsher selection inside them.

I can now show you why with something better than my own opinion. Contractors have been running this experiment in public, and four threads I read this week converge on one uncomfortable finding: for a lot of people, leaving does not change the work at all.

Start with what you are actually selling

Every argument about which platform is better collapses into one fact: you are selling hours. Not deliverables, not a product, not a business that compounds while you sleep. Hours.

That framing kills a lot of bad reasoning. A platform is not a company you join, it is a buyer for inventory that is finite and does not carry over. An hour spent reading a guideline doc for platform number four is an hour you did not bill anywhere.

It also means the most important fact about any project is not the headline rate. It is whether the project pays for time or pays for tasks.

Time based projects pay for hours logged, usually with a cap on how many hours a task may take. Your risk is the cap.

Task based projects pay per completed and approved unit. Your risk is everything: how long the task takes, whether the tool lags, whether the reviewer approves it, whether the guideline changed between your submission and the review.

Task based work moves approval risk from the buyer to you. Read it as a signal about how much a project trusts its own instructions. When the specification is clean, buyers are comfortable paying for time. When it is a moving target, paying per accepted unit lets the buyer stop paying for the confusion they created.

The thing nobody told you: the client follows you

Here is the finding that should change how you think about switching, and I had not seen it stated plainly anywhere.

In a Mercor workers thread, a contractor described leaving for Handshake and being offered the same project, from the same client, for about a dollar more per hour. Others in that thread and a related one describe the identical pattern: one client running one project simultaneously across Mercor, Handshake, and Outlier, pay differing by a couple of dollars, tasks recognizably the same work. One contractor noted that when tasks dry up on one platform they dry up on all of them, because it is a single client queue underneath.

That dismantles the usual escape plan. When you switch platforms in frustration, there is a real chance you are renting the same work from a different landlord. The chaos you are fleeing often originates with the lab that commissioned the project, not the marketplace that staffed it.

Which means the honest question is not “is there a better platform.” It is “is there a better project,” and that question has a completely different answer.

What contractors report about the platforms themselves

Three things showed up repeatedly across these threads, and none of them are what people expect.

The access and pay relationship is inverse, and it is steep. One contractor tested five platforms and summarized it in a line worth memorizing: high pay means harder and more selective, easy access means much lower pay. His own numbers ran from about a dollar per task for basic audio evaluation up to roughly $85 an hour for specialized legal work, on the same resume. Every platform you can join this afternoon sits at the bottom of that curve.

Mercor’s chaos is relative, and multi platform contractors rank it well. This surprised me. Several experienced contractors independently rated Mercor first on organization, naming the per domain management and project Slack channels specifically, while criticizing its pay and volatility. One said they would take a pay cut to have Mercor’s project management everywhere. Your frustration is real. It is also, apparently, the good version.

The mechanics differ more than the rates do. Contractors on one client’s project across two platforms report very different effective earnings purely from timer rules: one platform requires the timer stopped and restarted at every submission, logs no time during mandatory training videos, and takes quality screenshots after the clock stops. The other lets it run continuously. Same tasks, same client, about two dollars an hour apart on paper, and the contractor concluded the higher nominal rate lost. Another described three day review latency during which no further tasking was possible, capping a good rate at roughly two hours of real work.

Nominal rate is the number you compare. Effective rate is the number you get paid.

A note on sourcing: everything in this section is contractor testimony from public forums, not something I verified myself. I am labeling it as such because that is the deal here. Where it matches my own experience across ten plus Mercor projects I say so.

Mercor and Handshake AI, by the numbers

Mercor pays more at the top. Domain experts at frontier labs reach $175 to $200 and above per hour, mid tier roles $90 to $150. Handshake clusters tighter, mostly $75 to $125. A genuinely senior resume makes Mercor’s ceiling worth the friction.

Handshake is steadier. Structured part time blocks, typically 10 to 20 hours a week. Mercor is spiky, with a one to six week wait after acceptance and no choice of assignment.

They screen differently, and this decides which fits you. Mercor runs a 45 to 90 minute AI video interview that probes your resume hard and targets senior professionals. Handshake runs shorter human led interviews and openly recruits PhD students, postdocs, and recent graduates. Early career, Handshake. Senior specialist, Mercor.

Both cap your hours. Handshake also carries an undisclosed weekly earnings ceiling, and in May 2026 one large project paid out between 20 and 50 percent of logged hours. A nominal $60 to $100 an hour can become an effective $8 to $25 once a cap bites.

One hard eligibility fact. Handshake AI requires US work authorization. F-1 students may qualify through CPT or OPT, STEM OPT is excluded, and only a minority of postings are open worldwide.

If that rules you out, read this part carefully. Most of you are not in the US, so for you the sentence “hold Mercor and Handshake” is not advice, it is a locked door. Three things still apply. The two queue rule is about structure, not about those two names: your second queue is whichever platform actually hires from your region, and the logic for choosing it is identical. Everything below the paywall is platform neutral, because the filter, the timer questions, and the drop rule describe mechanics that every one of these marketplaces runs. And Handshake is not entirely shut: its broad fellowship track is open worldwide, which is the one door on that platform I would send a non US reader to. I am testing the platforms that do not require US work authorization now, and that roundup is the next free issue.

So: hold two queues, whichever two your passport permits. Replace neither.

And if you have decided Handshake is your second queue, do not paste your Mercor resume into it. The two platforms reward opposite things, and the profile that got you through Mercor’s funnel is close to the worst version of you that Handshake can see. Below, I walk through the conversion.

Below the paywall:

How to convert your Mercor profile into a Handshake one, field by field, and the single setting that blocks most non students from registering at all
The contractor who registered with ten platforms, and what exactly one of them was paying her
The real waiting times contractors report, platform by platform, and why that is the number that kills the scatter strategy
My seven question filter for accepting or declining any project, including the contract clause that should end the conversation
How I allocate my own hours right now, and the mechanical rule that decides when I drop a project
The ten hour test budget, if you insist on trying a third platform anyway
What the sieve actually selects for, and why breadth is the expensive mistake

World◦Model: the world is your oyster

Should I explore more AI training platforms?

Comments on this post are for subscribers