# $500, then $200: what one posting tells us about the price of a coding session

> In October 2026 Mercor paid developers $500 per recorded coding session, then cut it to $200 within a week. What the price says about the market for agent session data.

Source: https://groundedby.com/blog/mercor-500-then-200-coding-session-price (updated 2026-10-29)

On October 8, 2026, a posting went up on Mercor's job board: developers could record a session of their own coding work in an AI agent and be paid **$500 for each accepted session**. A session meant about 50 turns over two to three hours, using models Mercor provided. The buyer was described only as a leading AI research organization.

Within a week the rate was **$200**. A few days after that the posting stopped taking applicants.

That one posting is the clearest public price for a developer's agent session so far. Here is what we read in it.

## The price fell because supply showed up

Developers have far more of this than anyone expected. Anyone who has used Claude Code or Codex daily for a few months has dozens of sessions matching that description sitting on their laptop. A $500 rate for something many people already do every day fills a quota fast, and the price falls to where the buyer stops getting more than it can use.

## $200 is still a high floor

Compare it with other priced agent work in 2026:

- Snorkel AI contractors were paid **$280 for about an hour** comparing Claude Code outputs, according to Business Insider in June.
- Outlier pays **up to $1,800** for a terminal task with automated checks, which takes six to seven hours to write.

Per hour, a recorded session at $200 for two to three hours sits close to commissioned labeling work. Passive data rarely holds that price. The buyer paid it because the sessions were **real work on real projects**, which is hard to commission.

## What the buyer specified

The terms describe the product: your own project, about 50 turns, a few hours, recorded in the agent the buyer chose. That rules out quick questions, invented tasks and other people's code. It is the same shape our [evaluator](/evaluate) calls a deep session.

## What it does not tell us

- **How many sessions were bought.** A one-off quota and a standing market price look the same in one posting.
- **Whether old sessions sell.** Mercor asked for new recordings in its own setup, not history people already had.
- **Who else pays.** One buyer, one posting, one week.

## Where that leaves your history

The price shows that real, long agent sessions are worth paying for. It also shows how fast a price can move once supply appears. We use $200 as the top of our deep-session range and explain the rest in the [valuation method](/docs/valuation-method). When the next public price appears, we will update both.

## Sources

- [Mercor, "Share Your AI Coding Sessions" posting (archived Oct 8, 2026)](https://web.archive.org/web/20261008012525/https://work.mercor.com/jobs/list_AAABoRiaKGOq5ceLrrpLR6zE/builders-developers-share-your-ai-coding-sessions)
- [aidata.broker, Mercor posting history](https://aidata.broker/sources/work-mercor-com-av5gz9)
- [Business Insider, Snorkel AI contractors improving Claude Code (June 1, 2026)](https://www.businessinsider.com/anthropic-improve-claude-code-snorkel-data-training-contractors-2026-6)
- [Outlier, Terminal-Bench expert tasks](https://outlier.ai/experts/terminal-bench)