Every “release date” for the next generation of AI models belongs to one of three species, and most of the trouble in planning comes from mistaking one for another. Some dates the labs themselves published. Some are reported—sourced by journalists and trackers, plausible, unowned. And some are merely what the release cadence suggests, arithmetic wearing a calendar’s clothes. This companion to today’s survey of the current field keeps the three in separate pens, labeled, because a rumor in a confirmed date’s clothing is how budgets die.
Confirmed: dates the labs published themselves
Notice the shape before the dates: almost everything a lab will commit to in public is a price change or a funeral. Launches are the one thing nobody dates.
- End of October 2026: Mistral promised the weights for its previewed Large 4 by month’s end, per the October release changelog—the one genuine launch window on this list.
- November 21: OpenAI’s promotional $4/$20 rate on GPT-5.6 Sol is guaranteed “at least through” this date on its own pricing page. After it, the list price is the risk.
- November 30: Anthropic retires Claude Sonnet 4.5, per the current deprecation record. The same day, Google’s introductory pricing on Gemini 4 Argon expires.
- December 1: Gemini 4 Argon steps to its standard $4/$20, per Google’s pricing page.
- December 11: OpenAI’s custom GPTs reach their standard retirement on the timetable reported this month, and the deprecation trackers put the original GPT-5’s API shutdown in the same window.
- January 1, 2027: Google’s Flash rates double, $0.75/$3.75 to $1.50/$7.50, again per its pricing page.
- January 14, 2027: GPT-4o’s API shutdown, per the deprecation schedule—the end of the model that taught the world to talk to a computer.
If you build on any of these, your winter calendar is already written, and it is written by accountants and undertakers, not by launch teams.
Reported: sourced, and not promised
The release trackers who watch the labs’ pipelines for a living currently carry: a GPT-6.1 Terra estimated for October/November, filling the one gap in OpenAI’s weather-named family; a Claude Fable 5.5 estimated for November/December; a Grok 4.8 that was expected around October 1 and is now weeks late—the lateness is itself the data point, since xAI’s rhythm had been metronomic; a Qwen 4 and a GLM-6.0, both rumored and both late; and a DeepSeek V4.1 Pro tracked with no date at all. Separately, WhatsNew’s reporting has Google’s Gems becoming Skills for personal accounts in November. Treat every item here the same way: a reported date is a reason to keep your architecture flexible. It is never a reason to wait.
“no, not yet”—the release tracker’s standing answer to every “is it out?”, and the correct prior for this entire tier
Expected by cadence: my arithmetic, labeled as such
Now the inference, openly flagged. xAI shipped flagships on August 12 and September 21—a six-week rhythm that points at early November for the next Grok, late rumor or not. Anthropic refreshed an entire four-model family between September 1 and October 7; a shop that can turn a family in five weeks plausibly turns the next one around the new year. OpenAI answered GPT-6 with a point release twenty-six days later, so expect the 6.1 family to fill in—Terra is the known gap—before anyone should hold their breath for a 6.5. Google ships a Flash roughly monthly and a flagship step roughly quarterly, which suggests more Gemini 4.x fills before year-end. This is weather, not timetable: cadence tells you the season, never the day.
What a buyer does with a calendar like this
Four rules, bought with experience. Build on what is shipping: the $2 tier documented in the companion survey is more model than most products will ever need. Budget the dated steps, because the only dates certain to arrive are the price rises and the funerals—December 1 and January 1 will keep their appointments even if every launch slips. Pin your model versions in production, because deprecations now arrive quarterly and a retired endpoint is a breaking change with a date on it. And never architect around a rumored model: waiting is a pricing strategy—inference fell by as much as 80 percent last quarter, as this morning’s roundup traced—but it is a terrible launch strategy. The next generation will arrive the way the last one did: unannounced on a weekday, cheaper than feared, and immediately compared against five rivals by people who put the same question to all of them.
Sources: OpenAI API pricing and Google Gemini API pricing, read October 26, 2026; October 2026 model changelog, StackSpend; Anthropic pricing and deprecations, updated October 2026; OpenAI deprecation dates, October 2026; GPT and Gems transition timetables, October 22, 2026; AI model release tracker, read October 26, 2026.