OpenAI may already have another major model ready. The historical question is whether five weeks is enough time for it to reach public release.
BID NO
PastBehavior thinks OpenAI is less likely than the market implied to publicly release a qualifying GPT-6 before October 1.
Contract note: We could not verify the full KXGPT-OPEN settlement rule. If Kalshi counts a differently named successor as GPT-6, the 20% estimate would be too low.
Kalshi asks whether OpenAI will release GPT-6 before October 1, 2026. Two things have to happen inside five weeks. A qualifying model has to reach the public, and based on the rules available to us it probably has to count as GPT-6.
Public means available to ordinary customers. A private demo does not count. Partner-only access does not count. A waitlist does not count. An announcement with availability promised later does not count.
The naming requirement is our reading of the rules rather than a verified one, and it is the biggest uncertainty in the study.
OpenAI ships constantly. Since GPT-5 arrived in August 2025 the company has released 5.1, 5.2, 5.3-Codex, 5.4, 5.5 and the 5.6 family. Seven models in twelve months, with gaps this year of 28, 49 and 77 days. On cadence alone, something arriving before October is entirely plausible.
Two things complicate that.
The first is naming. GPT-5.5, codenamed Spud, was described by Greg Brockman as a new base and a new pre-train. It shipped as a point release. Orion was widely expected to be GPT-5 and shipped as GPT-4.5. A new pre-trained model does not automatically get a new integer.
The second matters more. Frequent shipping is not the same as frequent shipping of a frontier model carrying unresolved safety and government constraints. GPT-4 finished training in August 2022 and was held until March 2023. GPT-5.6 went to roughly twenty approved partners before it went generally available. Releases can stay gated.
PastBehavior is not forecasting that OpenAI stops shipping. It is forecasting that five weeks may be too short for this particular model to reach qualifying public availability.
On August 1, OpenAI named Astra as its next major model inside a research post about ten solved open problems in mathematics and theoretical computer science. Sam Altman had already demonstrated it privately to lawmakers in Washington. The model exists, and the capability claims came with published proof artifacts.
On August 7, OpenAI said it could not rule out that Astra reaches the Critical cybersecurity threshold under its own Preparedness Framework. No model had been placed there before. GPT-5.6 Sol was assessed at High.
On August 18, OpenAI said a significant number of Astra workloads remained paused pending migration to hardened environments, and that the Preparedness Framework itself needs extending.
As of the study date there was no public model card, no API model ID, no pricing, no waitlist and no general release.
Against all that, DevDay falls on September 29, two days inside the window. And OpenAI has launched with almost no warning before. GPT-5.3-Codex was announced and released on the same day.
This is where the disagreement actually sits.
Public release before October 1
Naming / qualification
Kalshi also prices Astra separately without the naming condition. Comparing the two markets implies roughly a 45% chance that a released Astra qualifies as GPT-6.
The naming question is almost a wash. The forecast comes down to timing.
History shows OpenAI can move fast. It also shows that frontier releases carrying unresolved safety and government constraints stay gated, get staged, or slip.
Five weeks requires several things to clear quickly and in sequence. Safety testing finishing. Security migration completing. A federal pre-release review starting and ending. Product surface, pricing and documentation appearing. None of it had visibly happened by the study date, and OpenAI's own most recent statement described constraints tightening rather than lifting.
The adversarial pass moved the number up rather than down. A missing preview is weak evidence given the same-day launches. The August 18 language may not apply to Astra in full. DevDay is a real path to YES. Those points took the estimate from 16% to 20%. They did not close the gap to 35%.
This is a paper signal recorded for the experiment. It is not advice to trade real money.
PastBehavior is testing whether historical patterns can produce forecasts that can be written down, compared with the market, and scored later. This one resolves on October 1.