Gemini 4 Carbon Release Odds 2026: 70% by 15 November

Google announced its next frontier model three days ago and you still cannot use it, which is the entire reason these Gemini 4 Carbon release odds are worth reading rather than guessing at. Polymarket has $271,820.36 riding on four Carbon boards and the three that price the public launch date do not agree with each other: the cumulative ladder caps a 7 to 15 November release at 60.80% while the two boards beside it price a narrower window at 70.00% and 72.55%. We read all four boards from the Gamma API at 16:08 UTC on 31 October 2026, pulled hourly CLOB history back to each board's first print, checked Carbon's status against Google's own Gemini API changelog, and timed the announcement move against the 9to5Google post that broke it.

The reason the disagreement matters is that one of these boards is the one most people will look at, and it is the one that looks wrong. Everything below is built from the resolution text and the order books rather than from the launch coverage, because the launch coverage and the markets are answering different questions.

1. What a Carbon release market actually resolves on.

2. The announcement moved the board 27 points in four hours.

3. The full distribution, day by day.

4. The Argon clock: 13 days from announcement to GA.

5. Three boards, three answers, one impossibility.

6. The 9.2-point gap is executable, and it opened today.

7. The daily board sums to 118% and most of it is empty books.

8. The HLE-Diamond board has an inverted rung.

9. Six checks before you trade a Carbon release market.

Every one of these markets turns on one sentence, and it is not the sentence the launch posts are about. The resolution text requires that a qualifying model be "launched and publicly accessible, including via open beta or open rolling waitlist signups", and then rules out the alternative explicitly: "A closed beta or any form of private access will not suffice."

That is the whole story of why the 31 October leg is priced for a loss. Google announced Gemini 4 Carbon on 28 October 2026 and described it as coming soon to developers, businesses and consumers, with first access going to selected cyber defenders in its Fairwind Program. Reporting on the announcement gives introductory API pricing of $2 per million input tokens and $10 per million output, rising later to $4 and $20, with cached input discounted 95%. Announced pricing is not access, and the market knows the difference.

The check we can run ourselves is Google's own API changelog, which is where a GA date actually appears. Its 13 October 2026 entry reads "Gemini 4 Argon (gemini-4-argon) generally available (GA)". There is no Carbon entry on that page at all as of today, which is the cleanest available confirmation that nothing has yet happened that would resolve any of these boards Yes.

Two more clauses matter. The qualifying model must be named Carbon or be confirmed to be the model Business Insider reported employees testing on 9 October 2026, so a rebadged Gemini 4.1 only counts if someone confirms the identity. And a model merely derived or distilled from Carbon does not qualify, which closes the loophole where a smaller public sibling ships first.

The announcement went up at 10:08 am PT on 28 October, which is 17:08 UTC, and the hourly CLOB series brackets it almost to the hour. Through the early afternoon UTC the board was flat: the 15 November leg printed 38.50% at every hour from 13:00 to 16:00, the 30 November leg 64.00% and the 31 December leg 69.50%.

By 21:00 UTC all three had repriced hard. The 15 November leg traded 70.00%, the 30 November leg 90.50% and the 31 December leg 96.55%, gains of 31.5, 26.5 and 27.05 points respectively. The 15 November leg moved first, printing 55.00% in the 17:00 hour while the two longer legs were still flat, which is what you would expect if the first traders to react were buying the nearest plausible date rather than the tail.

The direction is what makes the size interesting. An announcement that explicitly withheld public access pushed the probability of public access up by a quarter of the board, because the information in it was confirmation that Carbon exists as a shipping product with a price list. That collapses most of the variance in a market whose real risk had been that Carbon stayed an internal checkpoint.

The two days before the announcement rule out the simpler story that these were young, untraded boards drifting: the 30 November leg opened at 47.00% on 26 October, hit 86.50% by 09:00 UTC on 27 October and fell back to 56.00% by that evening.

The cumulative ladder is the board worth reading first, because it holds the most money ($106,568.41 of volume and $54,802.46 of liquidity) and because the gap between two adjacent rungs is the implied probability of landing between them. Here are the ten rungs at 16:08 UTC on 31 October 2026, with the band each pair implies.

Released by 31 October: 1.25% cumulative, implying 1.25% for 31 October or earlier. By 2 November: 2.50% cumulative, 1.25% band. By 4 November: 2.95% cumulative, 0.45% band. By 6 November: 7.20% cumulative, 4.25% band. By 9 November: 23.50% cumulative, 16.30% band. By 11 November: 41.00% cumulative, 17.50% band. By 13 November: 63.50% cumulative, 22.50% band. By 15 November: 68.00% cumulative, 4.50% band. By 30 November: 87.50% cumulative, 19.50% band. By 31 December: 95.00% cumulative, 7.50% band. No release at all in 2026: 5.00%.

The ladder is coherent, which is worth saying because the comparable Arena debut board on Polymarket is not: every rung is at least as high as the one before it, the eleven bands are all positive, and they sum to exactly 100.00%. That is the baseline against which the next three sections are failures.

The distribution itself is tight and front-loaded. The three bands covering 7 to 13 November carry 56.30 points between them, and the single fattest band is 12 to 13 November at 22.50%. The market is not spreading risk across the quarter; it is betting on one working week and treating December as a 7.50% tail.

There is one precedent and the market is clearly using it. Gemini 4 Argon followed the identical pattern four weeks ago: announced on 30 September 2026, access restricted to vetted cyber defenders in the same Fairwind Program, broader availability promised without a date. Google's API changelog then recorded Argon as generally available on 13 October 2026.

That is a 13-day gap from announcement to GA. Apply it to Carbon and you get 10 November 2026. Here is how that single data point lines up against what each board prices.

Argon announced 30 September 2026, GA 13 October 2026, lag 13 days. Carbon announced 28 October 2026, Argon-implied GA 10 November 2026. Daily board modal day: 10 November at 26.55%, which is day 13. Bucket board modal window: 9 to 15 November at 70.00%, which contains day 13. Ladder's fattest implied band: 12 to 13 November at 22.50%, which is days 15 to 16.

So the daily and weekly boards are pricing the Argon lag almost exactly, and the ladder is pricing it two or three days late. That is a difference of opinion rather than an error, and there is a reason to lean later: Carbon's staged rollout is justified on cybersecurity grounds, Google says it is in a voluntary United States government pre-release review, and a model built to find and exploit vulnerabilities is a plausible candidate for a longer gate than Argon's.

What you should not do is treat 13 days as a distribution. It is one observation, four weeks old, and the launch coverage quoting mid-November is leaning on the same single precedent rather than on anything Google has said.

Polymarket runs three separate Carbon release boards and they are not arbitraged against each other. Lined up on the same question, the public launch date, they say different things, and one of the differences is not a disagreement but an arithmetic impossibility.

On the ladder, release by 6 November is 7.20% and release by 15 November is 68.00%, so the probability of a release in the 7 to 15 November window is 60.80% and cannot be higher. The bucket board prices the strictly narrower 9 to 15 November window at 70.00%. The daily board's seven legs covering 9 to 15 November sum to 72.55%. A subset cannot be likelier than the set that contains it, so the ladder is 9.20 points below the floor the bucket board implies and 11.75 points below the daily board's.

Here is the comparison in full at 16:08 UTC. Release on or before 31 October: ladder 1.25%, daily board 6.90%. Release 9 to 15 November: ladder at most 60.80%, bucket board 70.00%, daily board 72.55%. Release by 30 November: ladder 87.50%, daily board 87.50% (its "no release by 30 November" leg at 12.50%). No release in 2026: ladder 5.00%, bucket board 0.50% by 3 January.

The tails agree exactly and the middle does not, which tells you where the attention is. The 30 November figure matches to the basis point across two boards; the week-two legs disagree by nine points because that is where today's flow is, and it is not hitting all three books.

The gap is also new. At 08:00 UTC today the ladder had 15 November at 76.00% and 6 November at 7.25%, a window ceiling of 68.75% against the bucket board's 72.50% - a 3.75-point excess. Over the session the ladder's week-two legs drifted down eight points (13 November went 71.50% to 63.50%) while the bucket board gave up only 2.50. The inconsistency was manufactured this afternoon by one board repricing and its neighbour not following.

A pricing inconsistency only tells you something if it survives the spread, so here is the construction at the quoted asks rather than the mid prices. Buy the ladder's 15 November Yes at its 0.69 ask. Buy the ladder's 6 November No at 0.937, which is one minus its 0.063 Yes bid. Buy the bucket board's 9 to 15 November No at 0.31, one minus its 0.69 Yes bid. The three cost 1.937 together.

The payoff is at least 2.00 in every state of the world. If Carbon goes public on or before 6 November the first leg pays 1 and the third pays 1. If it lands between 7 and 8 November all three pay, for 3.00. In the contested 9 to 15 November window the first two pay and the third does not, for 2.00. After 15 November the second and third pay, for 2.00. The worst case locks 6.3 cents on a 1.937 outlay, about 3.25% on capital, with a 1.063 windfall if the release lands in the two-day seam the boards disagree about.

Then the catch, which is the same catch as every incoherence on these AI boards: size. The thinnest leg in that structure is the ladder's 6 November market at $2,929.81 of liquidity, against $5,325.60 on the 15 November leg and $3,359.49 on the bucket leg. A few thousand dollars of locked arbitrage is not a trade for most people, and the spreads will eat a careless fill.

Its value is diagnostic. A free 6.3 cents sitting all session on two events carrying $172,462.10 between them means nobody is enforcing coherence, so neither week-two price is a considered forecast. Read the ladder's 60.80% ceiling and the bucket board's 70.00% as the honest range instead.

The daily board looks like the most precise of the three and is the least trustworthy. Its 32 legs are mutually exclusive and jointly exhaustive, so the Yes prices should sum to 100%. They sum to 118.45%, an 18.45-point overround, and the excess is not spread evenly.

The clearest illustration is the one leg that is nearly settled. It is 31 October as we write, so the "on or prior to 31 October" leg has hours left and needs Google to make Carbon generally available tonight. The ladder quotes that at 1.25%; the daily board quotes it at 6.90%. Neither is a price anyone is bidding: the ladder leg has no bid and a 2.5-cent ask, the daily leg has no bid and a 13.8-cent ask, and in both cases the quoted probability is exactly half the ask because that is what a midpoint does to a one-sided book. The useful reading is not "the market says 6.90%". It is "someone will sell you this for 2.5 cents on one board and 13.8 cents on the other, and you should buy the cheap one or neither".

The same artefact inflates the middle of the board. The 13 November leg shows 11.10% from a 0.4-cent bid and a 21.8-cent ask, a 21.4-point spread. Only three legs have real two-sided books: 10 November at 26.55% on a 5.5-point spread, 11 November at 20.25% on 5.3, and 17 November at 6.50% on 1.0.

The bucket board is the better-behaved sibling at 104.60%, a 4.60-point overround, and its concentration is honest: $28,828.20 of its $65,893.69 volume sits in the single 9 to 15 November leg. If you want one number off these boards, take that one and discount it for the ladder's disagreement rather than averaging across all three.

The fourth Carbon board asks how well rather than when: whether the Carbon model added to the Humanity's Last Exam results at agi.safe.ai clears a given HLE-Diamond accuracy, read at noon ET the day after it first appears. It carries $49,995.26 of volume and runs to 1 July 2027.

Its five rungs quote 50%+ at 97.05%, 55%+ at 98.35%, 60%+ at 88.75%, 65%+ at 82.50% and 70%+ at 25.50%. The second rung is the problem: 55%+ is strictly harder to clear than 50%+, so it cannot be likelier, and pricing it 1.30 points higher implies a negative 1.30% chance of a score landing between 50 and 55. Every other band is positive - 2.95% below 50, 9.60% for 55 to 60, 6.25% for 60 to 65, 57.00% for 65 to 70 and 25.50% above 70 - and the six sum to 100.00%.

The inversion is thin rather than deep: the 50%+ leg holds $1,138.63 of liquidity against a 3.7-point spread and the 55%+ leg $894.35 against 2.5, so the crossing lives inside the spreads. The signal is the fourth rung instead - $24,446.99 of volume, nearly half the board, sits on 65%+ at 82.50%.

One caveat: Carbon has not appeared on agi.safe.ai, so there is nothing to anchor against, and Google's announcement numbers are cybersecurity benchmarks - 88.2% on SEC-bench Pro, 54.8% on ExploitGym, 71.0% on CWE-bench - that say nothing about HLE. The only third-party figure in the coverage is Artificial Analysis placing Carbon first on its v5.0 Intelligence Index at 59.

First, read the access clause rather than the headline. These markets need public accessibility, open beta or an open rolling waitlist; the Fairwind Program is a closed beta and does not count, however much coverage the model gets.

Second, check Google's API changelog, not the news. The Argon GA date came from that page, and a Carbon entry appearing there is the single most likely trigger for every release leg on all three boards at once.

Third, pick the board before you pick the direction. The three boards differ by nine to twelve points on the same week, and on a position that size the choice of venue matters more than the view.

Fourth, read the bid, not the quoted probability. On the daily board most legs have no bid at all, and the displayed percentage is half an ask that nobody has tested.

Fifth, treat the 13-day Argon lag as one observation. It is the only precedent and the market has already priced it; the edge, if there is one, is in whether a security-gated model takes longer, not in rediscovering the base case.

Sixth, do not expect a second venue to check against. We scanned all 12,000 open Kalshi markets on 31 October and found no Gemini release market, so unlike the Fed and election boards there is no cross-exchange price to arbitrage these against. Polymarket is the only quote, which is exactly why its three internal quotes disagreeing is worth knowing about.