The Problem with Pay

"We are not willing to compromise our compensation principles, our principles of fairness, to respond individually to these offers."

That is Dario Amodei, chief executive of the artificial-intelligence company Anthropic, on the Big Technology Podcast in July 2025, explaining why his company would not counter the offers rivals were putting in front of its researchers. WIRED reported Meta packages of up to US$300 million over four years; Meta disputed how some offers were characterised, but nobody disputed the direction. Amodei's reasoning was simple: paying one person more because someone else tried to hire them is unfair to everyone who didn't get the call. The offers have not stopped since. Neither has the policy.

If you are running a company of forty people in Singapore, that can read as a position available only to firms with a great deal of money and a great deal of leeway. You have probably lost someone good over a few hundred dollars a month, and you may have paid a counter-offer yourself, quietly, hoping nobody would find out.

Stay with me. The stance is not what it first appears to be, and the part that can apply to smaller companies costs nothing.

The best-retaining lab is not the best-paying one

Anthropic pays extremely well. As of August 2026, Levels.fyi puts median total compensation around US$402,000 company-wide and about US$686,000 for software engineers, and visa filings have shown individual technical salaries above a million dollars. Those are estimates built from self-reported data and public filings, not company disclosure but nobody there is trading pay for purpose.

It is still not the top of the market. OpenAI's reported median is higher, largely through equity, though the samples behind these comparisons are small and self-selected. The clean version: both sit at the top of the market, and only one refuses to negotiate individual exceptions.

Now put that against retention. SignalFire's State of Talent report: the analysis Axios recirculated in August 2026, a year into the poaching war, found that 80 per cent of Anthropic staff hired at least two years earlier were still there at the end of their second year.

• Anthropic — 80%
• Google DeepMind — 78%
• OpenAI — 67%
• Meta — 64%

Source: SignalFire, State of Talent 2025. Based on public employment profiles, not company records; Meta's figure is company-wide.

These are self-reported profiles, so treat the figures as indicative and as correlation, not proof that the pay stance causes the retention. But the retention ranking is hard to miss. The lab that refuses to bid for individuals keeps the most people.

Two things follow, and the first matters more than the second.

Pay has to be right before any of this applies. Anthropic's floor is high. Nothing in this article is an argument for paying below the market, and a company that underpays cannot substitute meaning for money, it will simply lose people more slowly than it deserves to, and lose the wrong ones first. Make sure you have the right compensation structure that is on market, then read on.

Above that floor, more money stops being the thing that decides. Most founders discover this too late, usually while writing a counter-offer.

What a counter-offer really costs

The case against matching an outside offer is not that it fails to work. It often works, for a while. The case against it is the one Amodei made: it is unfair to everyone who did not go looking.

A company that negotiates individually teaches its people something quite specific that the reliable route to being paid properly is to get an offer elsewhere and bring it back. That lesson spreads faster than any policy, and it lands hardest on the people who never would have gone. You have rewarded the one person who was leaving and taxed everyone who stayed.

A consistent structure is the cheapest thing on this list. It costs a founder nothing but the discipline to hold the line in a difficult conversation. This is also what fair means in Singapore's employment framework: reward decided on ability, performance, contribution and experience, the same rule for everyone, not the same number. If an exception is genuinely warranted, the person was under-banded. Kf the role has outgrown its level, fix the band, not the individual, and check everyone else the fix touches. An undocumented exception is just a negotiation with extra steps.

The mission people stay for is the mission of the job

Here is where most companies reach for the wrong instrument. Told that people want meaning, a founder writes a mission statement, puts it on the wall, and repeats it at the quarterly meeting. Almost nobody has ever stayed in a job because of a sentence about the company.

What holds a senior person is much narrower: knowing what their particular role exists to change. Decades of work-design research points the same way, autonomy, task significance, and being able to see whether you are winning and none of it is delivered by a wall poster.

Most job descriptions never say. They list duties, manage the payroll cycle, oversee recruitment, handle compliance matters. Duties tell somebody how to spend a Tuesday. They do not say what is different in the business if the work is done well, which means they give a capable person no way to tell whether they are winning.

The repair is a single sentence, written before the offer goes out. This role exists so that we stop losing people in their first ninety days. This role exists so that the founder is no longer the final approver on every hire. This role exists so that we can pass a Ministry of Manpower inspection without three weeks of preparation.

That sentence does three jobs at once. It tells the person what they are for, it tells you what you are actually buying, and it makes the first performance conversation obvious rather than awkward.

Then you have to hand over the ground to do it on

A stated purpose without authority is worse than no statement at all, because now the person can see the goal and cannot reach it. Which brings up the most instructive example of a company solving this, at a moment when it looked like it could only be solved with money.

By 2010, much of Facebook's original product group had gone. Adam D'Angelo, the chief technology officer, left in 2008; Charlie Cheever followed in 2009, and the two built Quora. Co-founder Dustin Moskovitz left in November 2008 and started Asana. Dave Morin left in 2010 for Path. Matt Cohler, vice president of product management, joined the venture firm Benchmark in 2008. Arriving in the other direction was a professional executive bench, much of it hired out of Google.

Facebook's answer was to buy six small companies over roughly eighteen months, shut almost all of their products, and keep the founders.

• FriendFeed, Aug 2009 — Bret Taylor: Chief technology officer, from 2010
• Nextstop, Jul 2010 — Carl Sjogreen: The developer platform and Open Graph
• Chai Labs, Aug 2010 — Gokul Rajaram: Product director for advertising
• Hot Potato, Aug 2010 — Justin Shaffer: Facebook Groups, later video
• Drop.io, Oct 2010 — Sam Lessin: The profile, then Timeline and Identity
• Beluga, Mar 2011 — Lucy Zhang, Ben Davenport, Jon Perlow: Facebook Messenger

Mark Zuckerberg described the intent plainly at Y Combinator's Startup School in October 2010: "Facebook has not once bought a company for the company itself. We buy companies to get excellent people."

The purchases were the expensive half. The half that transfers was free.

Hot Potato closed on 20 August 2010. Seven weeks later, on 6 October, its founder was the product manager fronting the launch of the rebuilt Facebook Groups — one of the largest releases of that year. Asked whether the work had been under way before he arrived, Zuckerberg agreed it had, then said it "wouldn't have gotten finished" otherwise. Beluga was acquired on 1 March 2011; its three founders shipped Facebook Messenger on 9 August, five months later. Sam Lessin was not parked in an innovation unit, he was given the profile, the most-viewed surface the company owned.

None of that required a cheque. It required a chief executive willing to hand a seven-week-old employee a launch, and then to say publicly that it would not have shipped without them.

The people already inside were watching. Aaron Sittig, one of Facebook's first designers, left in June 2010 after five years, and was back by January 2011. He said the pull was partly the company's ambition and partly the people arriving around him "the steady influx of talented people like Sam Lessin and Justin Shaffer convinced me to say yes when I was approached with an offer to rejoin." Scope given to new arrivals is read by everyone who stayed.

What doesn't transfer

Three honest limits, because this is not a clean playbook.

The sourcing was extravagant. Those deals ran into the tens of millions for a handful of people each. Buying companies to acquire staff is not available to an SME and it is not even a retention trick: US Census research found acquired employees leave faster than ordinary hires do. What Facebook's case shows is the speed of real ownership, not a hiring strategy.

Neither company can be copied on reputation. Anthropic and Facebook are among the most sought-after employers on earth. An SME saying join us for the mission is making a much weaker offer, which is exactly why the mission of the specific role has to carry the weight the logo cannot.

It did not buy permanent loyalty. Sjogreen left in 2012, Rajaram in 2013, Lessin in 2014. Three or four years of a senior person's best work is a good result, but plan the handover rather than assuming they will stay forever. And remember that ground handed to a new arrival was usually being held by somebody already there that conversation is a required step, not an optional one.

So the question is not what to pay

It is what the role is for, who can see that it belongs to this person, and whether you can hold a fair structure when someone puts a rival offer on your desk.

Set the structure, then stop negotiating individually. Benchmark your bands properly, correct anyone sitting below them before they have to ask, and decline counter-offers as policy rather than as a judgement about the person.

Write what the role exists to change, before the offer. One sentence, in the offer conversation, and again in the first review.

Hand over real ground inside ninety days, and say whose it is. Ninety days is an operating benchmark, not a law — set it by the risk of the role. Credit in front of the company is what converts a task into ownership.

Both halves of this are what we do. We build compensation structures that hold up to scrutiny, and we do the part most companies skip — designing what a senior role actually owns, so that the person you worked hard to hire has somewhere to put their ambition. We work as fractional HR: senior, part-time, inside the business rather than presenting to it.

If you are about to write a counter-offer, or you have hired well and it still is not landing, that is the moment to talk rather than the month after. Tell us the story no one else hears.

Pay settles what someone is worth. It cannot tell them what they are for.

Sources: Dario Amodei on the Big Technology Podcast, July 2025, as reported by Fortune; WIRED on Meta's disputed offers; Levels.fyi compensation estimates, accessed 12 August 2026; SignalFire's State of Talent 2025, recirculated by Axios, 3 August 2026; Singapore's Tripartite Guidelines on Fair Employment Practices; Humphrey, Nahrgang & Morgeson's 2007 meta-analysis of work design; the US Census study of high-tech startup acquisitions; and contemporaneous Facebook reporting by TechCrunch and AllThingsD.