The GPT-6 release date still is not a date, but its price just did something dates rarely do: it moved a month closer overnight without becoming any more likely to happen at all. On Saturday, August 29, the Kalshi contract for a GPT-6-or-greater release before October 1, 2026 ran from a 20-cent bid to a 46-cent bid against a 47-cent ask, touching 68 cents on the spike before the book settled down. That is a 26-point repricing of a single month in a single day.
And here is the part that makes this board worth a second read: the rung for before November 1 barely moved, and the rung for before 2027 did not move up at all. The crowd pulled the release forward on the calendar while leaving the odds that it ever happens this year almost exactly where they were. That is a very specific claim about the world, and the rest of this page takes it apart the way a sims player takes apart a projection spike: what moved, what the move is built on, and the fine print that decides whether the people paying 46 cents can actually get paid. The fine print matters more than ever, because this board is still quietly two bets in one, and the second bet just got more expensive to ignore.
The Quick Answer
Kalshi now prices a public GPT-6-or-greater release before October 1, 2026 at 46 to 47 cents, up from a 20-cent bid a day earlier, on a book now 1 cent wide with more open interest than any other rung on the board. But before November 1 sits at 52 to 57 cents and before 2027 at 61 to 62 cents, both essentially unchanged, so the ladder pulled September forward without raising the year-end total. The market is pricing an imminent release, not a more likely one. OpenAI has announced no GPT-6; its confirmed next family is tentatively called Astra, and it may not carry the GPT-6 name at all. The one-day repricing, the rung-by-rung math, the naming clause, and what a price-blind model panel makes of the new number are all below.
Free: The Weekly PM Market Brief
The 8-model panel's graded record, the week's biggest market-vs-model gaps, and what's spiking next. One email, Sundays.
Unsubscribe anytime, one click. We never share your email.
The Market At A Glance
| Venue | Kalshi, a CFTC-regulated event-contract exchange (18+; availability varies by state as of August 2026) |
| The Contract | "When will OpenAI release GPT-6?" (KXGPT-OPEN), sold as cumulative yes/no rungs: before September 1, October 1, November 1, and before 2027. Each rung settles on its own; they are not mutually exclusive |
| The Clock | Before July 1 and before August 1 settled NO. Before September 1 is about to join them: it shows a 0-cent bid against a 4-cent ask with roughly 48 hours left |
| Depth | About 444,000 contracts traded lifetime across the event's six rungs (including the two already settled), per our August 30 sweep. The October rung alone holds 53,484 contracts of open interest on 120,739 traded |
| Timestamps | All prices pulled from the live Kalshi order book as of August 30, 2026 |
Those are the mechanics. What changed since our last read is one rung, and the size of that one change is the story.
What Moved: 26 Points In A Day
Some history keeps the jump honest. When this page first read the board on August 12, before October 1 traded in a wide 44-to-58-cent range with a last trade of 45, a midpoint near 51 on a book 14 cents wide. Over the following two weeks the rung sagged, and on Friday, August 28 it closed at a 20-cent bid. Then Saturday's session repriced it to 46 bid, 47 ask, with the day's trades running as high as 68 cents and the last print of the spike landing at 64, while the bid itself never crossed 49 and closed at 46. So part of Saturday's 26 points is a round trip back to just under the rung's mid-August level. What is genuinely new is the conviction around the number: about 4,900 contracts changed hands in 24 hours on that rung alone, and the book that emerged is 1 cent wide on more than 53,000 contracts of open interest, more than any other rung on the board. A Saturday session runs thinner than a weekday and deserves a notch of discount, but this was not one lonely bidder walking a thin book uphill; it was a crowd agreeing on a number fast, and agreeing hard enough to compress a 14-cent spread to a penny.
The trigger sits in that day's news cycle. On August 29, the AI-model tracker TestingCatalog reported that OpenAI had expanded internal testing of its next major model under the codename "mozaik-alpha-fdm," with leaked output samples circulating and separate reports suggesting a launch within weeks. None of that is an announcement. OpenAI has published no release date, no pricing, no model card, and no claim that anything ships in September. The market treated the testing footprint as unconfirmed but directionally credible, and repriced now rather than after the press release.
A 26-point day invites a simple question: did GPT-6 just become much more likely, or just much sooner? The ladder answers that question cleanly, and the answer is the whole thesis of this page.
The Ladder Moved One Month, Not The Year
| Contract: GPT-6 Released Before... | Yes Bid-Ask | Last Trade | Implied Probability |
|---|---|---|---|
| September 1, 2026 | 0-4 cents | 4 cents | effectively dead; settles within days |
| October 1, 2026 | 46-47 cents | 64 cents (spike print; the live book is the price) | ~46-47% |
| November 1, 2026 | 52-57 cents | 54 cents | ~54.5% at the midpoint |
| January 1, 2027 ("Before 2027") | 61-62 cents | 62 cents | ~61-62% |
Live Kalshi order book, August 30, 2026. These are cumulative thresholds, so probabilities climb with the date; never add the rungs together.
More live boards from the same panel: the Anthropic Mythos release date board, where before-2027 trades at 17-20 cents, and the iPhone 18 release date board, where before-October trades at 2-3 cents, both as of August 30, 2026.
Because the rungs are cumulative, the information lives in the differences. Read the month-by-month slices off the midpoints: September now carries roughly 44 of the 46.5 points on the October rung, since the September 1 rung is priced at almost nothing. October itself adds only about 8 points, the gap between October's 46.5-cent midpoint and November's 54.5. November and December together add about 7 more, the gap up to before-2027's 61.5. One month is carrying the entire forecast, and it is the month that starts in two days.
Now hold that against the previous shape of this board, on both clocks. Over the single day, the story is stark: October jumped 26 points while November's book closed Thursday and Friday at 51-to-53-cent bids and sits at 52/57 now, and before-2027 closed those same days at bids of 60 to 66 and sits at 61/62. Over the two weeks since August 12, the picture is messier and worth stating honestly: the September window grew from about 38 points to about 44, though most of that growth is just the before-September-1 rung expiring toward zero; October's own slice held at 8 on both dates; the November-December tail actually fattened from about 2 points to 7; and the year-end total went from a 61-cent midpoint then, on a book that stretched from a 40-cent bid to an 82-cent ask, to 61.5 now on a book 1 cent wide. On the two-week clock, probability spread out in both directions. The clean pull-forward is Saturday's move specifically: one rung up 26 points on the day while every book behind it stayed put, with the year-end total flat on both clocks. That is a pull-forward, not a re-rating: the market's claim is "if it comes, it comes now," not "it is more likely to come." A repricing that fattens one month while leaving the annual total flat should make you ask what, exactly, got announced. The answer, as we are about to see, is nothing.
The Base Rate: What OpenAI's Cadence Actually Supports
Numbered flagships have never come fast. GPT-2 arrived in February 2019, GPT-3 in June 2020, GPT-4 in March 2023, and GPT-5 in August 2025. That is 16 months, then 33, then 29 between major numbers, an average of about two and a half years across the last two cycles. A GPT-6 before October 1, 2026 would land about 14 months after GPT-5, half the recent gap. A forecast built on cadence alone would price the September window in single digits, not at 44 points.
The counterargument is that the treadmill underneath the numbers has sped up. OpenAI shipped GPT-5.5 in April 2026 and GPT-5.6 in July 2026, a preview on June 26 and general availability on July 9 in three tiers named Luna, Terra, and Sol. Two ships three months apart is real velocity. But notice what neither ship did: neither one carried a new number. The point releases prove OpenAI can ship quickly and simultaneously prove it has been choosing not to spend the "6" on those ships. And the cadence prior is not just mine: the price-blind model panel you will meet below, shown the settlement rules and the evidence but never a price, put before October 1 between 5 and 15 percent across its six seats, in the same neighborhood a cadence-only forecast lands. So a market paying 46 cents on September is not pricing cadence at all. It is pricing a specific pile of news, and that pile deserves to be read item by item.
The Update: Astra's August, Item By Item
Three dated facts built this repricing, and they pull in different directions.
August 1: OpenAI confirmed its next major model family, tentatively called Astra, in the least product-like way imaginable: a research post reporting that an internal version had produced new results on ten long-standing open problems in mathematics and theoretical computer science, each with a machine-checkable proof. No release date, no pricing, no model card. Crucially, reporting on the announcement was explicit that OpenAI has not decided whether Astra ships as GPT-6 or as another GPT-5 point release. Remember that sentence; it is the second bet.
August 7: OpenAI disclosed that it could not rule out Astra having reached the "Critical" cybersecurity capability threshold under its Preparedness Framework, the first time the company has been unable to rule that tier out for one of its models. A Critical designation triggers additional safety controls and testing before any wide release, and OpenAI told Axios on August 18 that it had paused two weeks of deployment-focused reinforcement-learning training, with its largest planned frontier RL run remaining on hold. This is the item the 46-cent crowd has to argue past: the company's own safety process is publicly standing between the model and a ship date.
August 29: TestingCatalog reported the expanded internal testing described above, codename and leaked samples included, with launch-within-weeks chatter around it. This is the item the crowd bought.
Put the three together and the ladder's odd shape starts to make sense, because the three items map onto three different risks. The testing reports shrink ship risk, and only ship risk: the model being closer says nothing about whether it is allowed out, which the August 7 disclosure governs, or what it is called on the way out, which the August 1 ambiguity governs. Saturday repriced the one risk that got smaller and left the other two where they were. A market that believed all three items would do exactly what this one did: rush the near rung and refuse to raise the total. The crowd is not being irrational here. It is being precise about which uncertainty shrank, and it is the same two-column discipline a DFS player runs every slate: the projection is the base rate, the news is the update, and the number you actually use is neither one alone. The MLB DFS projections, ownership and stacks in the DataHub exist to run exactly this arithmetic on ballplayers instead of language models, and you can try the Sims free to watch base rates and news get blended thousands of times per slate.
The Naming Fine Print Splits The Bet
Here is the second bet I flagged in the opening, and Saturday's buyers now own more of it than anyone. The market's rules resolve YES only if OpenAI releases "a model called GPT-6 or greater" before the date, with the release public and outside a closed beta (a high-cost subscription tier still counts). Astra is not confirmed to be GPT-6. OpenAI has said the naming is undecided, and the two most recent flagship-adjacent ships, 5.5 and 5.6, both took the point-release exit. By the rules text, an on-time Astra that launches under the Astra name alone, or as a GPT-5.7, would not obviously settle these contracts YES.
That means the 53 to 54 cents of NO on the October rung is not buying one outcome; it is buying three at once. The release can slip past October 1, the safety process can hold it, or it can arrive on time wearing a badge that is not GPT-6 or greater, and the contract pays NO in all three branches. YES, by contrast, needs the calendar, the safety gate, and the christening to cooperate, three doors that all have to open in about four weeks. Recognizing a narrow path behind a confident price is the most transferable reading skill on this page. Read the resolution rules on the market page yourself before forming an opinion about any of these numbers.
What A Price-Blind Model Panel Makes Of 46 Cents
Last edition, this page promised that our price-blind AI model panel would score this board on its next refresh. Here is that read, with its date attached. On August 21, we put the four open rungs to a six-model panel, each model shown the settlement rules and the public evidence but never a market price, with every verdict stored and graded against real settlements as markets resolve. The medians: about 4 to 5 percent for before September 1, about 10 percent for before October 1, about 11 percent for before November 1, and about 18 percent for before 2027. The seats disagreed most about the far rung, ranging from 5 to 38 percent on before-2027, and least about September, where every seat sat in single digits or barely above.
Two honest caveats cut in opposite directions. The panel's read predates the August 29 testing reports, so it never saw the news that moved the market 26 points; a fresh run would likely land higher on the near rungs. On the other side, the panel's earlier skepticism has at least aged well so far: an eight-model version of this read, published August 16 on our tech release-date board, put before-2027 at 18.6 percent while the market read near 78 cents, and two weeks later that rung trades at 61 to 62. One honest asterisk on that walk: the 78 came off a wide, thinly traded book whose midpoint had sat near 61 days earlier, so some of those 16 points are a spread collapsing rather than opinion moving. And to be clear about what any of it is and is not: the rung has not settled, and a market drifting toward a forecast is agreement, not proof; the panel's actual scoring happens when these contracts resolve and every stored verdict gets graded. So the panel-versus-market picture today is roughly 10 percent against 46 cents on October, a chasm, and the entire difference is how much weight you put on one weekend of leak reporting versus a cadence, a safety hold, and a naming clause. The panel is not an oracle; it is a disciplined base-rate reader that cannot be swept up in a spike, which is exactly the second opinion a 26-point day calls for.
A Worked Example: From Cents To Probability
A 46-cent contract is the crowd pricing roughly a 46% chance of a $1 payout, but fees shift the effective number, and Saturday's move is a live lesson in how. Kalshi's published trading fee formula is 0.07 times contracts times price times one minus price, which peaks at 50 cents, so the jump from 20 to 46 nearly doubled the per-contract fee: 0.07 x 100 x 0.46 x 0.54 is about $1.74 on 100 contracts, versus about $1.12 for the same fill at 20 cents. Measured against what you actually stake, though, the fee got lighter, 5.6% of a $20 position at the old price against 3.8% of $46 now, because the stake grew faster than the fee did. Either way, a buyer at 46 cents needs closer to a 48% true probability to break even, and that two-point tax matters most exactly where this rung now sits, in the middle of the curve where fees bite hardest.
The October rung also teaches this board's other reading habit, one the before-2027 rung taught last edition from the opposite direction. October's last trade says 64 cents, but the live book is 46 bid, 47 ask. That 64 is a spike print left over from Saturday afternoon, a stale opinion the tape has already walked away from. Two weeks ago the lesson ran the other way here: before-2027 showed an 80-cent last trade on a book 42 cents wide, and the real number turned out to live far below the print. Same rule both times: read the live range, not the last trade, and when the two disagree, trust the book over the print.
A standing reminder before the close: every number on this page, the market's and the panel's alike, is a set of model estimates, not predictions of fact and not financial advice. Event contracts are for adults 18 and older where legal, with availability varying by state as of August 2026, and one OpenAI blog post can repaint this entire curve in an afternoon.
- GTA 6 Release Date: Will November Hold After The Leaks?
- The Winds Of Winter Release Odds: Will It Ever Ship?
- iPhone 18 Release Date: Apple Set a Date, Not for This Phone
- Michigan Governor Odds: Seven Raters, Not One Says Safe
- Crypto 2026 Odds: Can Any Of These Ten Finish The Year Green?
The Bottom Line
The GPT-6 release date market spent two weeks as a cadence argument and became an event trade in one Saturday session: a 26-point September pull-forward built on leak reports, sitting on top of an unchanged year-end total, a public safety hold, and a naming clause that can void an on-time release. The flat before-2027 rung is the tell; the crowd moved the date, not the destiny. Our price-blind panel's last read, taken August 21 before the leak reports, put October near 10 percent and the year-end rung at 18, and the last time the market and the panel disagreed this loudly on this board, the market drifted 16 points toward the panel in two weeks, some of it spread collapse, none of it settled yet. Prices will move again; the habit of asking "sooner, or likelier?" is the durable part. For more boards read this way, start at our sports prediction markets hub; the cumulative-ladder habit transfers directly to the MLB division race odds, and the narrow-path read, a confident price that needs several doors to open at once, repeats on the next-team boards for Bronny James and Tyreek Hill.
