No Varnish

Research

We read every open role at 89–90 AI companies. Here is what we found.

Not a survey and not an aggregator. Every figure below is counted at the source: on the employers’ own applicant tracking systems, which is the only place a posted salary survives intact, and on the job boards themselves.

roles in the latest scan
8,239
company job boards
90
oldest figure here
25 August 2026

  • Who tells you what a job pays

    57% of 8,204 AI roles state a salary

    Every published figure on this question comes from an aggregator, and aggregators strip the pay field. This counts it at the employers' own boards instead.

  • What those salaries actually are

    The typical posted range is $68k wide

    Anyone can approximate a median. The number nobody publishes is how much room a posted range leaves — two people hired against one advert can be that far apart and both be inside it.

  • Whose disclosure is worth reading

    Bands run from 18.2% to 55% of the midpoint

    Every posting here states a salary, and they are not equivalent. One employer's median band resolves the number to within a fifth of it; another leaves more than half open.

  • Which tools employers ask for

    MCP is named at 47 of 90 boards

    Every survey of AI tool adoption asks developers what they use. This counts what employers wrote in the requirements, which is a different question with a different answer.

  • Claude Code vs Cursor vs Copilot

    46 boards to 16 to 9

    The order is the reverse of what install-base figures would predict, and the reason is the useful part: ubiquity is exactly what stops a tool being written into a job advert.

  • Which job boards tell you the pay

    18% of 154 listings on other boards state a figure

    The only page here that measures the boards rather than the employers, this one included and sorted by the same rule. A salary is on a listing or it is not, which is the rare claim about a competitor that needs no judgement.

Why this is measured rather than asked

A salary survey asks people what they earn and reports what they say. This reads what the employer wrote on the advert for a role it was actually trying to fill — so nothing here is self-reported, modelled or extrapolated.

It is also the reason the numbers are lower than a survey’s. A posted band is what a company committed to in writing, not what the best-paid person in the room negotiated.

Each page states its own method and the date it was measured. Where a figure cannot be derived honestly — a level with too few postings, a range in another currency — it is counted in the totals and left out of the ranking rather than filled in. How listings are sourced covers the same discipline applied to the board itself.