Why do AI labs pay engineers $1,000 per coding task?

Why do AI labs pay engineers $1,000 per coding task? The answer is compute. During reinforcement learning (RL), the lab runs each task against the model many times. That compute costs more than the engineer who writes the task. A broken task wastes all of that compute. A task with a weak check teaches the model to cheat. Thus labs pay for tasks that are hard, fair and safe from shortcuts. Only engineers who know a codebase well can write these tasks. They must find a change that strong models still get wrong. The Askable listing that started this post pays “US$1,000 for each task that passes our quality review”. The role is open to all countries.

Why do AI labs pay $1,000 per task? The short answer

What do you build for $1,000?

You do not label data. You select a real change from a repository that you know. You set up the repo in Askable’s platform in its condition before that change. Then you write three items: the brief for the model, the correct solution, and the checks that grade an answer.

Next, you attack your own checks. You try lazy, incorrect and shortcut solutions. Each one must fail. Then you run the task against the best models “until strong models often fail but a skilled engineer can clearly solve it”. You submit the task. A reviewer examines it and you make changes. The listing says that most tasks need more than one review. Askable pays only for an accepted task.

The requirements are specific. Askable wants 2 or more years of professional engineering. You must have code merged into a public open source repo, ideally the repo for your tasks. You must know Docker, Linux, Git and automated tests. You also need “experience writing tests, rubrics, or evaluation criteria”. Deep knowledge of databases, compilers, distributed systems, security or infrastructure is a plus. The minimum is 10 hours and one task each week.

Epoch AI wrote an FAQ on RL environments. The authors spoke to 18 people at environment startups and labs. Their description agrees with the Askable listing. An environment is usually a Docker container. A task has a prompt and a grader. The grader is usually a set of unit tests or an LLM rubric. The scores of the model’s attempts change the model’s weights. One startup employee said that RL requests are 10 to 20 times more frequent than benchmark requests. Thus most of these tasks are training data, not tests.

Why is one task worth that much?

Compute explains the price. Mechanize did the calculation in Cheap RL tasks will waste compute. A SWE-bench Verified transcript has approximately 20,000 tokens. Frontier RL tasks already had approximately 100,000. Transcript length increases approximately 5 times each year. Thus Mechanize predicts 500,000 tokens for each attempt. They use the Grok 4 price of $15 for one million output tokens. With 64 attempts in each run, one run of one task costs approximately $480. Five runs cost approximately $2,400.

Mechanize says that data and compute work together. If a lab uses $2,400 of GPU time on a task, a $100 task is a bad decision. They compare it to “cheap tires on a Ferrari”. They said that less than approximately $500 for each task is not efficient. They expected this number to increase five times in one year. They wrote this in August 2025.

Epoch did its interviews in January 2026. Most task prices were between $200 and $2,000. One founder said “$20k per task would be rare but possible”. Exclusive tasks cost 4 to 5 times more. The vendor keeps a part of the price. Thus a $1,000 payment to the author is in this range.

Why can’t labs just scrape public code?

There are two reasons. Both are in the job listings.

The first reason is contamination. OpenAI stopped using SWE-bench Verified. It found that GPT-5.2, Claude Opus 4.5 and Gemini 3 Flash saw the solutions during training. With only a task ID and a hint, these models wrote the fix again. The tasks came from public repos, and all labs collect public repos. OpenAI also examined 138 tasks that GPT-5.2 always failed. It found that 59.4% of them were broken. For example, some tests required a function name that the problem did not mention.

Thus Askable says “Don’t publish the work anywhere: no public pull request or public fork”. Askable also accepts a private codebase that you own. It does not accept code from an employer or a client. A task has value only if its answer is not already on GitHub.

The second reason is reward hacking. Epoch’s interviewees said that this is the most important quality criterion. Models search for the solution. They also check out future commits. One interviewee said “it takes many many iterations to check against reward hacking.” A check that passes when the model did no work is worse than no check. The model learns the shortcut. I wrote about the same failure in agent tests in delta verification. There, an absolute assertion passes when the agent did nothing.

The difficulty must also be correct. Some interviewees wanted a minimum of one success in 64 or 128 attempts. If no attempt solves a task, the task gives no signal. If all attempts solve it, the model learns nothing.

Who’s hiring, and what do they pay?

Askable is not the only company. The table shows the companies that I found. They pay senior engineers to build coding tasks or to review AI output. The pay data comes from listings and aggregator sites. It shows advertised pay, not real income. Most of these companies want you to make an account and pass a screen before you see work.

The range of pay is important. Hourly evaluation work is between approximately $60 and $150. In this work, you read model output and give it a score. Task authoring for RL gets payment for each accepted task, up to $1,000. The lab uses one good task in many runs. The author also accepts the risk of rejection.

Where does the money come from?

The money comes from the labs. The amounts are large. Mercor had more than $2B in gross annualized revenue in June 2026, as reported by The Information. Approximately two-thirds of that goes to contractors. Foundation model companies gave 91% of the revenue in the first half of the year. Mercor says that it pays its experts approximately $4M each day.

In September 2025, The Information reported a plan at Anthropic. Anthropic discussed a budget of more than $1B for RL environments for one year. Epoch says that this amount is small compared to compute costs. Datacurve operates Shipd. It got a $15M Series A to sell coding datasets and repo-level RL environments.

The money goes from the lab’s training budget to a vendor, and then from the vendor to you. The vendor does the difficult part. One of Epoch’s founders said: “Finding the experts isn’t that hard, but managing them and doing quality control is hard.” The vendor must also make sure that contractors do not make their tasks with an LLM. This is the slop grenade problem, but the damage goes into a training run.

Should a senior engineer take this work?

It all depends on your speed to an accepted task. At 10 hours each week, $1,000 for each task is good pay if one task takes one week. The pay is much worse if a task gets three reviews and then a rejection. Askable pays only for accepted tasks. Hourly platforms pay for your time. But aggregators report unpaid onboarding and training hours on some of them. Also, the work comes in batches, and a batch can stop.

The rights are also important. You must own the change and the solution. Thus you cannot use code from your employer. After Askable accepts a task, you cannot publish it. Do you maintain an open source project? Do you own a private codebase with an interesting history? Then you have the material that labs need. I think that this material is what the $1,000 pays for, more than the seniority on your CV.

So why do AI labs pay $1,000 per coding task?

The task is the less expensive input. A frontier RL task can use approximately $2,400 of compute during its life. Labs already pay $200 to $2,000 for each task. A task that shows its answer, or that permits shortcuts, wastes all of that compute. It also trains the model to cheat. Public code is contaminated. Thus labs need new changes from people who know a codebase well. These people select a change that strong models fail on. They write checks that a model cannot trick. Askable’s $1,000 pays for that judgment. You get it only when a reviewer agrees that your task is correct.