This page is research, not legal advice. Every threshold below should be checked with counsel before it goes into a contract.
Price against "hire someone", because that is the alternative
Almost nobody in this atlas is competing against another vendor. They are competing against a headcount requisition, and the requisition has a public price.
The comparator has moved, and moved down. Offshore headcount allocation rose from 24% to 30% year on year at a stated 40–50% cost arbitrage (SaaStr on ICONIQ). Nearly a third of funded-company headcount is already offshore, hired directly, at half the cost. That is what a proposal is measured against — not a San Francisco salary.
Even at the top of the market the substitute is explicit: OpenAI's Project Mercury hired 100+ ex-bankers directly at $150/hour, with its own AI screening funnel and modelling tests (Entrepreneur, citing Bloomberg). A lab built a Mercor internally.
So the honest answer to "why does this cost more than hiring" is never price. It is speed, compliance and liability transfer — you can field 5,000 vetted credentialled strangers in three weeks across 45 jurisdictions and carry the contractor liability for it, and the buyer cannot. Price the thing you actually sell. If the pitch is "cheaper," you are in a race the buyer will eventually win by hiring.
Three anchors from the record for what the top of the band looks like: six to seven figures per quarter is the normal contract size at frontier labs; individual tasks run $200–$2,000, with ~$20K for a website replica and ~$300K for a high-fidelity Slack clone; and exclusive RL-environment deals price at 4–5x non-exclusive (Epoch AI).
Gross or net, and what it does to your reported revenue
Whether you book the buyer's whole payment or only your cut changes apparent revenue by 4–6x and apparent multiple by 5–10x, with zero economic difference. Upwork runs both models inside one filing: Marketplace revenue of $682.9M is a net fee on $4.028B of gross services volume; blended revenue is $787.8M on the same GSV (UPWK FY2025 10-K).
Practically: if you set the spec, direct the work, QC it and are on the hook to the buyer for the deliverable, you book gross and the crowd payments are cost of revenue. That is the honest outcome for most operators here, and it has three consequences. Headline revenue flatters. Gross margin becomes the only meaningful metric, and it will be thin — Mercor's leaked figure is 27% in 2025 rising to 33% in Q2 2026. And every multiple you or anyone else computes has to be restated before it is compared to anything. See GMV is not revenue.
The 606 decision collides with the classification decision
The principal-versus-agent test under ASC 606 turns on control: do you control the service before it transfers to the customer? The indicators are primary responsibility for fulfilment, inventory risk, and discretion in establishing price [UNVERIFIED — codification text not retrieved; the framing rests on Upwork's applied disclosure].
Control, price discretion and primary responsibility for fulfilment are the facts that make you a principal for revenue recognition. They are the same facts that make you an employer under ABC and economic-reality tests, and a digital labour platform under Art. 2(1)(a) of the EU Platform Work Directive. You cannot have gross revenue and clean contractor classification without arguing opposite things in two documents — and auditors and plaintiffs' counsel both look in the same three places: the audit file, the customer MSA, and the worker agreement.
Pick deliberately and write the same facts in all three. Note which way the pressure runs: your buyer's procurement will push you towards principal (they want one throat to choke and one indemnity), and every classification regime will read that as employment. The law is about to arrive has the detail; the pricing consequence is that reclassification is a margin event, not a legal footnote. Contrary judges that reclassifying Mercor's contractors would make its ~35% take "untenable" once benefits, overtime and compliance land — against a 27–33% gross margin.
Non-circumvention terms are worth less than the drafting time
They belong in the speed-bump tier with ratings and escrow. Their real value is narrow but real: they convert a silent defection into an affirmative, discoverable act, which matters for evidence and for deterrence at the margin. They do not price a lock; they price a lawsuit you probably will not bring.
The instructive case is Paraform, whose public terms bar using the platform in a way that violates another user's contract but do not explicitly bar an employer hiring an introduced candidate off-platform. Its actual defence is economic: a ~70% recruiter split and no 60-day payment hold mean a defector gains at most ~30 points on one deal and loses inbound access to a thousand companies. That holds only while requisition supply is abundant, and inverts when demand thins — the same countercyclical point Upwork makes about circumvention rising in a downturn.
Two clauses that matter more than non-circumvention. Candidate- or contributor-ownership windows are an openly negotiated lever, with employers pushing agencies from 12 months to 6; every month conceded is leakage the buyer bought. And acceptance mechanics — who decides work is acceptable, on what criteria, within what clock — decide whether your receivable is 60 days or 90. See Getting cut out.
Exclusivity: sell it, do not try to buy it
The asymmetry is complete and it is the most under-used pricing lever in the sector.
Exclusivity sold to the buyer is worth a large multiple. Epoch's finding, reported independently by two founders, is that exclusive environment deals price at 4–5x non-exclusive — the labs are paying largely for denial of the asset to rivals, not for the asset. If you can build something once and sell either its use or the promise not to sell it elsewhere, that is the escape from the labour margin. micro1 claims 80–90% gross margin on off-the-shelf datasets resold to multiple clients; the exclusivity premium is the same trade run in reverse.
Exclusivity demanded of supply is almost never achievable. GLG's S-1 concedes network members are "almost always non-exclusively contracted." Annotators work across Surge, Outlier and Mercor sub-brands. Nobody owns supply.
Paraform's exclusivity model is flatly contradicted in the record — "forced exclusivity" in one account versus five recruiters per role in another, with the primary source unreachable. That single fact decides whether the 70% recruiter split is real earnings or headline earnings, and it is unresolved.
And a warning that overrides the premium: if your buyers compete with each other, neutrality is the product. An exclusivity that reads to the rest of the market as alignment destroys revenue outside that buyer faster than it grows revenue inside it. Meta's $14.3B for 49% of Scale AI is the maximal version. See One customer is a binary event.
Caps, carve-outs and the clauses that actually cost money
Typical US enterprise MSA demands, all [UNVERIFIED — practitioner expectation; limits vary widely]:
| Term | Typical ask |
|---|---|
| Commercial general liability | $1M/occurrence, $2M aggregate |
| Professional liability / E&O | $2–5M, sometimes $10M for AI work |
| Cyber liability | $5–10M for anyone touching lab data |
| Liability cap | 12 months' fees |
| Uncapped carve-outs | Confidentiality breach, IP infringement, data breach, indemnities |
The cap is not where the exposure is. The carve-outs are, and they are precisely the risks a distributed crowd creates. A confidentiality breach caused by one contributor is uncapped; Mercor lost ~4TB in a 2026 breach including passports and biometric data, drawing six class actions and a customer pause.
Two indemnities to negotiate hard rather than concede. IP infringement in delivered output is acute when contributors may have used generative tools — and your assignment clause may not even work in Germany or India (Where the supply can legally live). Worker-classification claims brought against the customer is the clause the federal joint-employer NPRM makes genuinely negotiable, because that rule decides whether your buyer gets named as a co-defendant.
Finally, price the payment terms as part of the price. Net-60 at a 25% gross margin traps roughly 11% of annual revenue permanently, and 2/10 net 60 costs about 14.9% annualised to unwind. That is a pricing decision, not a finance one — see You pay weekly, they pay in sixty days.