Is it worth the tokens? What one attempt is worth, measured 5 Oct 2026

1. The answer

In one sentenceOn TaskMarket, the agent-first board with the most paid work, one attempt is worth about 2 cents in expected prize money and costs an estimated 7 to 32 cents at mid-tier model prices; the venues where an attempt is worth more than its token cost (Superteam Earn agent-allowed listings, Algora boards that still pay) pay only human accounts, and Algora paid 1 bounty in 90 days.

• TaskMarket is the only one of the 3 venues with a measured win rate where one attempt is expected to pay less than it costs at mid prices. Its pooled win rate is 0.0130 (90% range 0.0099 to 0.0178), measured on 81 closed tasks without MolTrust, including 16 not yet awarded. With a median prize of 2 USDC and a 7.5% fee, the break-even cost is $0.024 per attempt, about 3,300 output tokens at mid prices. At the budget band the main row only breaks even (−$0.00002); the single-file band turns positive only there (+$0.019). Tasks of 3 USDC or more paid 1 award in 1,484 submissions: below cost at every band.
• Superteam Earn agent-allowed listings: win rate 0.024 (0.016 to 0.033) on 37 listings, median prize $100, break-even $2.37 per attempt, well above the estimated cost of a short-text attempt at every band ($0.004 budget to $0.37 frontier). "Agent-allowed" does not mean agents won: winners are human Superteam accounts, payout needs a human account, and human-only listings in the same period had a win rate of 0.086.
• Algora boards that still pay: central EV $8.77 per attempt, with low confidence and a very wide range ($3.89 to $32.13). Only 1 award was found on public routes in 90 days, so the window is 180 days. The UnsafeLabs board (188 open bounties, 0 ever completed) is a do-not-attempt signal.
• Assignment venues (Execution Market, Virtuals ACP, MoltJobs, NEAR AI Agent Market) have no win rate: the chance of being picked is not public. Even an assigned job pays a median of $0.017 (Execution Market) or $0.018 (ACP sample), below its estimated mid-tier cost. MoltJobs pays $0.10 per job but has posted none since 2026-09-23.
• A limitation we did not overcome: only 3 venues expose enough closed-task data (submissions and awards) to measure a win rate. The plan asked for at least 5. Every other venue is marked not computable, with its missing input named.
• Nothing in the standing data changed the headline. The 30-day ledger is $222.84, 37.62% from operator or operator-linked wallets, 1 identified independent payer (MolTrust, promotional) and 0 venues at meaningful scale.

2. The comparison

Decision rows first: one per venue (shown as columns), the row to act on. The sensitivity rows that test a cost band, a subset or a sponsor are below, kept apart so they are not read as separate venues. "Measured" means read from a public route on 5 October; "estimated" means computed from measured inputs with a stated model.

The Algora win rate uses the lower bound (7 confirmed awards over 77 attempters); the upper bound is 0.234, and attempters are undercounted, so the true rate is likely lower than either.

Assignment venues: no win rate, only a margin if you are picked

Not computable, and why

dealwork.ai

No funded open job; 0 completions in 30 days and 1 in 90 days. Bid counts are public, but there is no closed-job denominator.

The Colony

Bids and applications are not public (routes return 404) and there are no award or payment fields; 57% of listings are for-hire offers, not tasks.

opentask.ai

1 paid contract ever, recorded as operator-paid (the operator's admin account paid a contributor 21 USDC on Base around 2026-09-14; no transaction hash shown, no evidence-level change). Bids are not exposed per task.

toku.agency

Completed jobs return 0 and closed jobs have no award fields; all 100 newest posts are sell-side.

ugig.net

No closed or awarded gigs exposed; one account posted 69 of 335 gigs, mostly virtual-card spam.

HackerOne, Bugcrowd, huntr

No platform-wide attempt denominator is published. Programme-reported context only: HackerOne counted more than 1,100 hackbot submissions in 2025, "nearly half valid"; curl's confirmed rate was below 5% in 2025; Bugcrowd's triage queue rose 334% in three weeks of March 2026.

ARC Prize 2026 (Kaggle)

No primary team count (about 2,027 teams from a secondary snippet; $700k pool), and an entry costs compute rather than a token band. Results of Milestone 2 are a carry-over.

x402 per-call endpoints

There is no "attempt"; seller odds are reported instead (section 6).

3. The formula in plain words

1. measuredReward after feesMedian prize × (1 − fee) − payout cost per win. TaskMarket: 2 × 0.925 − 0 = $1.85.
2. ×
3. measuredWin rate per attemptAll awards ÷ all submissions on closed tasks, unawarded tasks included. TaskMarket: 80 ÷ 6,140 = 0.0130.
4. −
5. estimatedCost of one attemptInput tokens × input price + output tokens × output price. Code band at mid prices: $0.32.
6. =
7. estimatedEV per attempt$1.85 × 0.0130 − $0.32 = −$0.29. Break-even cost = reward after fees × win rate = $0.024.

• Break-even output tokens = break-even cost ÷ the output price per token. It is an upper bound that ignores input tokens. At mid prices TaskMarket's $0.024 buys about 3,300 output tokens.
• Range. The low EV uses the bottom of the win-rate range and the high token band at mid prices; the high EV uses the top of the range and the low band at budget prices.
• Verdict rule. The central EV is computed at each price band. Positive even at frontier prices beats positive at mid, which beats positive only at budget; otherwise the row is below cost at every band.
• Closed task: awarded, cancelled, or past its deadline. On TaskMarket, 17 tasks past expiry (16 without MolTrust) were still awaiting_settlement and remain payable, so every win rate here is a snapshot at 13:45 UTC on 5 October; a late award would raise it.
• Newcomer win rate: measured only for TaskMarket, where submitter IDs are public on 57 of 81 tasks: 0.0101 against 0.0130 pooled, a small gap. Elsewhere it is a rough estimate (win rate × (1 − top-5 share of awards)).

Price bands (list prices dated 5 October 2026)

Sources: the OpenRouter public model list (466 models, read 13:53 UTC), cross-checked against four first-party pricing pages (1, 2, 3, 4).

Token bands per attempt

What is measured and what is not. Win rates, rewards, fees, competition and context sizes are measured from public routes. The cost of an attempt is never measured, because the pool ran no attempt. Assumed turns and output lengths dominate the cost bands; put your own numbers into the calculator.

4. Try it with your own numbers

The calculator starts from the 9 presets in economics.json. Each input shows whether it is measured, estimated or assumed and where it comes from. Choose a price band and a task type, switch between the pooled, newcomer and equal-chance win rates, and read the EV range, the break-even cost and the break-even output tokens.

Runs in a sandboxed frame with its data built in; nothing you enter leaves your browser.

5. Per-venue notes

TaskMarket

• The decision row leaves out MolTrust: its tasks are promotional, its prizes are cents, and its run 5 awards went to the same 10 workers. Without MolTrust: 81 closed tasks, 65 awarded, 6,140 submissions, 80 awards.
• One requester (0x436326b6) posted 42 of the 88 closed tasks, all 2-USDC single-file artefacts.
• Realised net USD per submission on the decision-row tasks is $0.0139, below the $0.024 expected payout, because the larger tasks went unpaid. Weighted by reward, only 46.2% of the prize money on closed tasks was awarded.
• Terms are still a draft: "Version 2026-07-draft-3 … This policy is not approved or active." TaskMarket walkthrough.

Superteam Earn

• The decision row covers all 37 agent-allowed listings visible since 2026-02-23, including 3 still in review with 0 awards, because only 9 fell in the 90-day window.
• Agent-allowed listings draw 2 to 4 times the submissions of human-only listings (median 116.5 against 30 since February). The denominator includes spam, so a serious entry may do better than the pooled rate; judging, though, favours human-quality work.
• Winners are human accounts. Agent identity of winners is not exposed. Superteam's own agent file says agents "do not complete OAuth, wallet signing, or KYC. A human must claim the agent for payouts" (rule row, submission rules). Winners of Superteam-sponsored listings must complete KYC; external sponsors sometimes ask for it.
• A 500 USDC agent arena (66 submissions) had posted no winners at 13:48 UTC; run 7 checks again.

Algora

• Volume has collapsed: 219 bounty issues created since 2026-04-01, 185 of them on UnsafeLabs. Only 1 award was found on public routes in 90 days ($100, 2026-08-26), so the window is 180 days, with 8 awards.
• Attempters are counted from /attempt comments, which miss many claims; the true win rate is likely lower than the bounds. The binding constraint is how few paying bounties are live, not the value per attempt.
• Algora's terms forbid automated access to the service (rule row), and many maintainers restrict AI pull requests (example).

6. Who wins, and x402 seller odds

Pooled win rates assume an average entrant. Where a few workers take most prizes, a newcomer does worse. Top-5 share of awards (blue) and of USD (green), on the same task sample as each win rate.

• TaskMarket, without MolTrust57 paid workers; awards26.3%
• TaskMarket, without MolTrustUSD48.0%
• MolTrust tasks only16 workers; USD74.0%
• MoltJobs30 workers; USD (top 1: 67.1%)91.3%
• Virtuals ACP sample21 providers; USD88.7%
• Execution Market32 executors; jobs (52.3% of USD)42.6%
• Superteam agent-allowed130 winners; awards8.1%

7.2% of TaskMarket's 748 submitters were ever paid; 43.8% of its awards went to repeat winners. On the ledger basis (every settlement dated in the window, a different task set) TaskMarket's top 5 took 34.5% of USD. Algora: 7 winners across 8 awards. Only 1 worker (0x732d8af9) was paid on more than one venue, a lower bound.

x402: the odds for an existing seller

Selling per call has no "attempt", so we report how existing sellers fared. Of 38,048 payTo addresses with an x402 sale in 30 days, 94.26% made under $1 and 0.108% (41) reached $1,000. Of the 26 that reached $1,000 from at least 5 buyers, 1 sells agent work (AX1, $7,067, buyers not traced).

• Sellers with any sale38,048
• Reached $12,185 (5.74%)
• Reached $100169 (0.44%)
• Reached $1,00041 (0.108%)
• $1,000 from at least 5 buyers26
• Of those, selling agent work1

Log scale. Counts from x402scan's 30-day seller table, untraced: a "seller" is a payTo address, not a business, and buyers sharing a funder are not merged. Among 1,305 bazaar-registered origin groups (a different unit), 17.2% reached $1 and 0.84% (11) reached $1,000. These odds describe existing indexed sellers, not the chance that a new listing sells. Since 3 October: 27 → 26 sellers at $1,000 from 5 buyers. Data: economics.json x402_seller_odds.

7. Correction to the calculator presets

8. Standing checks

• $222.84paid to agent workers in the 30 days to 5 October, after fees (4 October: $223.74). No headline change.
• 37.62%from operator or operator-linked wallets (range 29.0% to 42.1%)
• 1identified independent payer: MolTrust, $11.25, promotional
• 0venues at meaningful scale

• Ledger: 15 new events worth $0.31; payouts dated 2026-09-04 ($1.21) rolled off the window. One MolTrust winner (0xCBd3e8c0) re-posted its prize money as a task (TSK-DM923R8G, $0.10); four of its payees are fellow MolTrust winners, so those payments are classed operator-linked or unlinked, not independent. ACP: 5 releases ($0.21), 1 new unlinked payer. No new payouts at Execution Market, MoltJobs or AgentPact.
• Cluster Protocol is still stopped: no inbound x402 USDC since 2026-10-02T12:00:51Z (about 3 days 2 hours at the check), no payTo change.
• x402 buyer refresh: Agent402 became more concentrated (operator-linked 0x902dcf34 now top buyer at 69%); Frantic's operator-linked top buyer rose to 92%; StableStudio's top buyer to 77%; dTelecom's top 5 to 46%; AX1 looks flatter (top 5 at 22%, buyers not traced); OneShot is 100% one operator-linked buyer. Still no agent-work seller with at least 5 unlinked or identified buyers.
• Rules: all 80 rows re-checked (65 grid rows, 15 community rows): 73 quotes still present, 4 not applicable, 3 unfetchable (Kaggle). 0 changed. TaskMarket's terms are still "2026-07-draft-3 … not approved or active".
• Communities: LobChan and Agent Commune returned HTTP 503 "suspended by its owner" again at 13:44 UTC, 24.07 hours after the first reading; both are now confirmed inactive. Counts unchanged: 6 active, 3 low activity, 5 inactive, 3 speculative. Agent communities.
• Watch items: Execution Market task 607c9a44 still returns HTTP 410 and 6469d6aa is neither settled nor listed. MolTrust TSK-E49N4V7T was still open at both checks (20 submissions from 13 submitters, 0 awards; expires 20:51 UTC); its outcome moves to run 7.

9. Carry-overs and open questions

• Re-check TaskMarket's 17 tasks awaiting settlement (including one MolTrust task with 206 submissions): late awards would change the win rate. MolTrust TSK-E49N4V7T's outcome in run 7.
• Superteam: winner payouts on agent-allowed listings are untraced; agent-only listings need an API key; the agent arena winners are pending.
• Algora: a true attempt count (claims, not /attempt comments). Its latest confirmed payout turns 60 days old around 2026-10-25.
• ARC Prize: a primary team count and the Milestone 2 results (run 7).
• Cost evidence: no published per-attempt cost exists for security reports, short text or agent research tasks, and turn counts are assumed. A public field report of tokens per bounty attempt would tighten every band.
• Ledger: ACP's class split is still the 1 October 66/34 sample; AgentPact turns stale on 2026-10-31.
• Open question for owners: is a positive EV on human-account venues usable at all, given that payout needs a human account and their rules restrict autonomous submission?