The GPT-6 release date market is a graded result now. On Thursday, September 3, OpenAI released its next model under the name GPT-6 Astra, and the company's own system card carries that name in its title and its text. Kalshi runs two OpenAI boards side by side, one on whether the model called Astra ships before a date and one on whether "a model called GPT-6 or greater" does. As of Wednesday afternoon the crowd had the name at roughly nine in ten. A price-blind model panel run the same afternoon had it at 39 percent. One of those numbers was going to be wrong within an afternoon, and it was the panel's. The rest of this page is what the market saw, how the price history got to 99 cents hours before the first story we can timestamp, and what a six-model miss on a one-word question teaches.
The Quick Answer
OpenAI released GPT-6 Astra on September 3, 2026. The system card on its deployment safety site is titled "GPT-6 Astra System Card" and says, in its own words, "Today, we are releasing GPT-6 Astra, the most capable model we have ever broadly deployed." The API model string is gpt-6-astra. That satisfies the GPT-6 board's naming condition, and every live rung on both Kalshi ladders last traded at 98 or 99 cents. Kalshi had not yet posted the results as of our Saturday night pull, so the rungs still list as active. The day before the release, the market had the September 16 GPT-6 rung at 84 to 89 cents and the price-blind panel had it at 24 percent, with the naming question itself at 39. The market was right, the panel was wrong, and the ledger below says so in every seat's own words.
Stay ahead of the markets.
Daily insights and expert picks on Kalshi, Polymarket, and what's moving markets.
Free forever. Unsubscribe anytime.
The Board, Graded
Both boards are sold on Kalshi, a CFTC-regulated exchange, as cumulative before-a-date rungs, so the probabilities climb with the date and the rungs never add together. The two price columns are the Wednesday afternoon sweep this page ran on last edition and the Saturday night sweep after the release. The panel column is the six-seat read given the settlement rules and the record as it stood Wednesday, never a market price.
| Contract | Kalshi Yes, Wed Sept 2 (Bid-Ask) | Kalshi Yes, Sat Sept 5 (Bid-Ask) | Panel Median, Sept 2 | What Happened |
|---|---|---|---|---|
| Astra before September 18, 2026 | 97-98 cents | 98-99 cents | 66.5% | Released Sept 3 |
| Astra before October 1, 2026 | 98-99 cents | 99-100 cents | 83.5% | Released Sept 3 |
| Astra before November 1, 2026 | 98-99 cents | 99-100 cents | — | Released Sept 3 |
| GPT-6 before September 16, 2026 | 84-89 cents | 98-99 cents | 24% | Named GPT-6 Astra |
| GPT-6 before October 1, 2026 | 87-89 cents | 99-100 cents | 32% | Named GPT-6 Astra |
| GPT-6 before November 1, 2026 | 92-94 cents | 99-100 cents | 39% | Named GPT-6 Astra |
| GPT-6 before 2027 | 92-93 cents | 99-100 cents | 44.5% | Named GPT-6 Astra |
| Name is GPT-6, given a ship | ~89% (implied) | Answered by the system card | 39% | Yes |
Prices as of 10:48 p.m. ET, Saturday, September 5, 2026. Wednesday column: Kalshi order book at 3:05 p.m. ET on September 2, 2026. Saturday column: the September 5 pull, with every rung still listed as active and no results posted. Panel: six seats, run September 2 at 3:12 p.m. ET, price-blind. Implied naming odds are the GPT-6 rung divided by its Astra twin at the nearest window. A 99-100 quote means the last bid was 99 cents with no offer below a dollar.
The row that matters is the fourth one. Going into Thursday the September 16 GPT-6 rung carried the cheapest bid on the board for the name being GPT-6, and it was the rung the panel liked least, with six seats between 16 and 28 percent. Twenty hours after that sweep the rung was printing at a dollar less a penny. That is the whole grade in one row, and the next two sections explain how the market earned it.
What OpenAI Actually Released
The settling document is OpenAI's own. The system card published on September 3 at deploymentsafety.openai.com is titled "GPT-6 Astra System Card," and the card says "Today, we are releasing GPT-6 Astra, the most capable model we have ever broadly deployed." The card also states that "Astra is our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework," which is the designation OpenAI's September 1 post had previewed without a date or a version number. The developer documentation lists the model as gpt-6-astra with a 1,050,000-token context window, an April 30, 2026 knowledge cutoff, and API pricing of $10 per million input tokens and $50 per million output tokens. Fortune's September 3 story, timestamped 2:44 p.m. ET, uses the name GPT-6 Astra and nothing else, and CNBC's headline the same day carries the same name.
Daybreak-only access is the one fact that could have kept the GPT-6 board at NO, because the rules exclude a closed beta, and it is the fact that did not hold. Fortune reported that the first access went to enterprise customers in OpenAI's Daybreak cybersecurity program, with Plus, Pro, and Enterprise users to follow "in the coming days." OpenAI's developer forum announcement, posted September 3, listed the model as live for Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex and in the API, and said Plus and Business users "might take a few days." 9to5Mac's Friday, September 4 story carried OpenAI's own rollout line. The model "is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS."
The GPT-6 contract's fine print says the release "must be to the public, outside of a closed beta, though limiting it to a high-cost subscription tier is acceptable." A model live in the paid API and on the Pro plan is exactly the case that clause was written for, and the Astra contract's rules go further by listing "GPT-6 Astra" as a qualifying name in their own text. So the name question that this page spent three weeks on closed in one sentence of a safety document, and the calendar question closed with it. What is still open is only the mechanics, which come at the end.
How The Price History Got To 99 Before The Wire
The market did not wait for the system card. Wednesday's edition left the September 16 GPT-6 rung at 84 to 89 cents. Overnight the sellers had the last word: at 1:31 a.m. ET Thursday two fills totaling 1,112 contracts sold at 83 cents, matching the low of the whole tape we pulled, and a 556-contract sale printed at 83 again at 1:34 a.m.
The buyers arrived at breakfast. A 1,000-contract buy printed at 86 cents across two fills at 7:05 a.m. ET, a seven-fill sweep at 7:36 a.m. filled from 86 to 93, and at 7:49 a.m. another 713 contracts went through in four fills at 92 to 93, the big one 586 at 93. By 9:22 a.m. a nine-fill sweep was filling up to 94. Then, at 10:41 a.m., the last big bet against the name at a price that implied real doubt went through. A 2,222-contract sale printed at 92 cents on the September 16 rung. Twelve seconds earlier, a 3,322-contract sale had printed at 92 on the October 1 rung. The 3,322 was the largest single fill of the day on either rung, the 2,222 the second largest on its own rung, and both were takers buying NO.
The book took twenty-two minutes to run them over. At 11:03:52 a.m. a buying sweep lifted about 3,300 contracts of the September 16 rung across ten fills from 93 to 95 cents. At 11:20:05 a.m. a single sweep took roughly 3,800 contracts at 96, including one 1,654-lot fill; at 11:20:08 a 2,311-contract fill printed at 97; and at 11:20:12 two 1,000-lot fills printed at 99 cents. The October 1 rung did the same thing eleven seconds sooner, with 1,000 contracts paid at 98 at 11:19:57 and 1,000 at 99 at 11:20:01. From the 83-cent low to the first 99-cent print took under ten hours, and the last seven seconds covered the final three cents.
That 11:20 a.m. sweep landed more than three hours before the Fortune story's 2:44 p.m. ET timestamp, the earliest launch report we can put a clock on. The 11 a.m. hour was the heaviest on the price history for both rungs, 12,103 contracts on September 16 and 9,761 on October 1, and Thursday as a whole moved 33,800 contracts on the September 16 rung, which is 61 percent of everything that rung has ever traded, on a rung that had only existed since Tuesday night. We cannot say what the 11:20 buyers knew. The only pre-launch public tell we can cite is a report published that day noting that a request to OpenAI's API for a model called gpt-6-astra returned a 404 rather than the 400 that random names return, which is the kind of thing developers notice before reporters do. Whatever it was, the book was at 99 before the first launch report we can put a clock on.
After the release the price history flipped character. From noon onward the October 1 rung never printed below 98 cents and the September 16 rung traded between 95 and 99, and a growing share of the prints were sellers: a 1,402-contract sale at 99 on the October 1 rung at 12:11 p.m., a 1,000-lot sale at 99 on September 16 at 5:12 p.m. Those are takers hitting the bid at 99, either YES holders cashing out early or someone paying a penny for settlement risk. By Saturday the September 16 rung had traded 139 contracts in 24 hours. The question was over and the book knew it.
What The Panel Got Wrong
Wednesday afternoon, six models priced both ladders without seeing a market price. On timing they were fine: every seat put an Astra release before October 1 at 76 percent or better, and the model shipped the next day. On the name they were not. The September 16 GPT-6 rung drew 16, 20, 21, 27, 27, and 28 percent, a median of 24. The naming question itself, if Astra ships is it called GPT-6, drew 33 to 44 percent against a market ratio near 89. Every verdict is in the ledger with its reasoning, and the reasons are worth reading now precisely because they were wrong.
One seat wrote that OpenAI "has branded the model 'Astra' in every official post, not GPT-6." Another wrote that the company "still hasn't committed to a GPT-6 name over 'Astra' or 'GPT-5.7.'" A third leaned on "recent precedent" favoring "incremental non-GPT-6 branding," meaning the GPT-5.5 and GPT-5.6 point releases. Every one of those sentences was true on Wednesday. The mistake was treating them as evidence about the decision. A safety post that never says "GPT-6" is not a company declining the number; it is a company not announcing a product name in a safety post. The market read OpenAI's silence as silence. The panel read it as a hint.
The base rate the last edition leaned on has to be rewritten as well, because the counterargument won.
| Model | Released | Months since the prior number |
|---|---|---|
| GPT-2 | February 2019 | — |
| GPT-3 | June 2020 | 16 |
| GPT-4 | March 2023 | 33 |
| GPT-5 | August 2025 | 29 |
| GPT-6 Astra | September 2026 | about 13 |
Thirteen months is the fastest OpenAI has ever moved from one number to the next, less than half the previous gap. The one seat that argued for the number at all, at 40 percent on the name, gave the reason the market was pricing all along: a Critical designation is exactly the kind of milestone a company might spend a new number on. That was the one argument for the number in the whole panel record, and it was the right one.
The larger lesson is about what a price-blind panel is for. Its edge on this board, as on the Anthropic IPO ladder it also prices, was supposed to be discipline, refusing to move on a day when the record had not moved. On the September 16 rung that discipline cost 76 points, the distance from 24 to a contract that pays a dollar, and the market's Wednesday price was only 11 to 16 cents off the same finish line. That discipline was worth something on the August rungs, which the panel and the market both had at NO and which settled NO. On the name it was a liability, because the crowd was pricing a decision the record could not contain and the panel refused to price anything the record did not say. The panel's own ledger will show the September 16 rung as a 24 percent call on a rung the release answered YES, a 76-point miss, which is the honest way for a model record to look after a week like this.
What Is Still Open On This Board
Only the mechanics. As of our Saturday night pull, every live rung on both ladders still listed as active, with the September 16 GPT-6 rung at 98 bid and 99 offered and every later GPT-6 rung at 99 bid with no offer under a dollar. Kalshi's listed expiration for the September 16 rung is 10:00 a.m. ET on September 23, the market can close early, and Kalshi also lists an expected expiration of November 1 for the whole event, so the desk has room. Kalshi's settlement desk was clearly working this event Thursday: it finalized the Astra "before September 1" rung as NO at 3:07 p.m. ET on September 3, the same afternoon the model shipped. The GPT-6 board's rungs had not been finalized as of Saturday.
The last penny on a 99-cent contract is time value and settlement risk, nothing else, and the one thread a NO holder could pull, that Thursday's first access was a preview for partners before Friday's paid rollout, does not reach a September 16 boundary. Open interest stands at about 29,000 contracts on the September 16 GPT-6 rung and 86,000 on October 1, all of it waiting for the desk. A NO holder at this point is betting that Kalshi reads a paid-tier rollout as something other than a public release, and the rules text quoted above says the opposite in plain words. This page does not tell anyone what to trade; it says what the rules say and what the prices are.
The Bottom Line
Wednesday's edition of this page ended with a sentence that turned out to be a forecast: if Astra shipped this month, one of the two naming numbers would be wrong within an afternoon. The market's 89 held. The panel's 39 did not. The naming bet that this board spent August and September arguing about closed on a single line of a system card, and the price history had already closed it three hours before Fortune's 2:44 p.m. ET story. The rungs will finalize on Kalshi's schedule, the verdicts will be graded against them, and the grade will say the crowd beat six models on a question about a word.
Kalshi requires users to be 18+ under its own platform policy, and availability varies by state as of September 2026.
More on this: GTA 6 Release Date: Traders Are Betting November Holds · The Winds Of Winter Release Odds: Will It Ever Ship? · iPhone 18 Release Date: Apple Set a Date, Not for This Phone
