What forecasters put on AGI, by year
Consensus table — snapshot 2026-09-26 11:40:56 UTC
Independent venues, one anchor date per row: AGI before this date? Venues read this run: polymarket, kalshi, manifold. Series that entered the median this run: kalshi-any-company-announce, manifold-we-get-agi (of the 3 eligible: manifold-we-get-agi, kalshi-any-company-announce, metaculus-strong-agi); the reference columns ask narrower questions and are shown, never averaged. Every number is a third-party quote read at the time above; none is live and none is this site's opinion. Note that the series do not share one definition of AGI — each resolves on its own rules (listed in the next table), and the median mixes an announcement-settled market with judged forecasts.
| Anchor | Manifold Will we get AGI before D? | Kalshi Any company announces AGI before D | Metaculus Strong AGI devised, tested and publicly announced before D (community CDF) | Kalshi + Polymarket OpenAI announces AGI before D | Metaculus Weak AGI publicly announced before D (community CDF) | Median | Spread |
|---|---|---|---|---|---|---|---|
| before 2027 | 2% | 13% | — | 17% | — | 8% | 11% |
| before 2028 | 24% | 66% | — | 48% | — | 45% | 42% |
| before 2030 | 50% | 79% | — | 65% | — | 64% | 29% |
| before 2035 | 63% | — | — | — | — | — | — |
| before 2040 | 72% | — | — | — | — | — | — |
This run: before 2027 8% (spread 11%) · before 2028 45% (spread 42%) · before 2030 64% (spread 29%).
Where each series crosses 50%, and what each one actually resolves on
| Series | Implied 50% date | Resolution basis (from the venue's own text) |
|---|---|---|
| Manifold — Will we get AGI before D? | 2030-01-01 | Manifold: 'Resolves as YES if such a system is created and publicly announced before D'; since 2025-11-02 the family uses the creator's 'In what specific year will we hit AGI?' market as its baseline (creator/mod judgement). play money |
| Kalshi — Any company announces AGI before D | 2027-10-01 | Kalshi KXAGICO: settles on a company announcement per the series' settlement sources (SEC EDGAR filings and named news outlets); it does not judge capability. real money |
| Metaculus — Strong AGI devised, tested and publicly announced before D (community CDF) | — | Metaculus: resolves against the question's own tests (adversarial Turing test, robotic assembly, general game play, coding), judged by Metaculus admins. no money (reputation) |
| Kalshi + Polymarket — OpenAI announces AGI before D | 2028-03-23 | Polymarket/Kalshi: resolves on OpenAI or an official representative announcing it has created AGI (Polymarket: 'consensus of credible reporting' as secondary source). real money · reference only |
| Metaculus — Weak AGI publicly announced before D (community CDF) | — | Metaculus: a lower bar than the strong-AGI question (Turing test with text only, Winograd, SAT, Montezuma), judged by Metaculus admins. no money (reputation) · reference only |
A market that settles on an announcement pays when a company says it has AGI; a judged forecast resolves on the tests written into the question. The gap between them is a real number on this page, but read it with the resolution column, not the venue name.
Method (recompute it yourself)
For each anchor date D, every series contributes at most one row: the open market or question whose 'before' date equals D. A series is a venue whitelist plus a fixed regular expression on the question text; if a venue lists duplicates for one date, the row kept is the real-money one first, then the highest volume, then the most unique bettors. The consensus value is the plain median of the consensus series that have a row for D (minimum 2 series; otherwise no median is printed); spread is max minus min of those same values. Reference series are shown but never enter the median. The implied 50% date per series is the first crossing of 0.5 by linear interpolation over that series' own dated points (Metaculus: its community median date). Prices are read as the venue reports them at the fetch time and are never adjusted. The fetcher keeps every dated timeline row it found and, beyond those, only the highest-volume rows up to a fixed cap, so the full list below is 'all dated rows plus the top of the rest', not every match.
gen_market_board.py in the public repository regenerates
agi-consensus.json from /market-board.json with
python3 tools/gen_market_board.py --check, and fails if a byte differs. Markets are selected by a fixed keyword filter
on what the venues return and a fixed venue-whitelist × regex per series; nothing on this page is hand-picked.
Not read this run (stated, not padded):
- metaculus: METACULUS_TOKEN 未设置(未登录 403;runner 从 Secrets 注入)
Timestamp proofs
Versions of agi-consensus.json stamped since 2026-09-26 carry an OpenTimestamps proof: 2 version(s) listed, 1 with a Bitcoin block attestation, 1 pending (calendar receipt only, not in a block yet). This version (sha256 c327ae396014…) has proof agi-consensus.json.c327ae396014.ots, status pending (stamped 2026-09-26). Index: /ots/manifest.json. A proof shows that the exact bytes existed no later than the block it names; it says nothing about whether the numbers are right, and versions before 2026-09-26 rest on the public git log only. Verify with the client or web verifier at opentimestamps.org: ots verify <proof> -f agi-consensus.json.
Markets read this run (55 shown of 106 matched; all dated timeline rows kept, the rest top-by-volume)
Why a scorecard publishes odds it does not trade
A market prices what a crowd expects. This site grades what has been demonstrated, against criteria written down before the fact — the AGI-2027 Thesis Tracker reads >62.5/100 as of 2026-09-06. When the median above and the tracker diverge, the divergence is the story. That is also why there is no bet button here, and never will be. We are not an operator, not a broker and not an affiliate of one; we quote public prices the way a newspaper prints a closing index.
Third-party market data quoted with its as-of timestamp. This site takes no positions, accepts no wagers, earns nothing from any venue listed, and none of this is trading or betting advice.
Frequently asked questions
No. It is a median over independent public forecasting venues, computed by a published formula that anyone can rerun on the same snapshot. The scorecard's own position is the Thesis Tracker, which grades evidence, not expectations.
Snapshot 2026-09-26 11:40 UTC, median across kalshi-any-company-announce, manifold-we-get-agi series: before 2027: 8% (spread 11%); before 2028: 45% (spread 42%); before 2030: 64% (spread 29%). Each series resolves on its own rules, so treat the spread as part of the answer.
No. Each run records a snapshot with an exact UTC timestamp and links to the market so you can check the current price yourself. A quoted price that pretends to be live is the single most common way odds coverage misleads people.
Only series that ask about AGI in general enter the median: Manifold's 'Will we get AGI before D?', Kalshi's 'any company announces AGI before D' and Metaculus's strong-AGI question when it was read. 'OpenAI announces AGI' is one company's announcement and 'weak AGI' is a lower bar, so both are shown as reference columns instead. The series in the median still differ in what resolves them; the page lists each one's basis.
No. This site is not an operator, a broker, or an affiliate of one. There is no wager mechanism, no referral link, and no revenue from any venue listed.
By a fixed keyword filter applied to whatever the venues' own search and series endpoints return that run, then a fixed venue whitelist and regular expression per series. The board can look thin in a quiet week, and it will say so rather than pad itself.
The live scorecard updates as models ship and verdicts change.
View the live scorecard →