New AI models released this autumn: a founder's briefing

More than a dozen new AI models were released between September 1 and October 27, 2026, from Anthropic, OpenAI, Google and a run of open-weight labs. For a founder, the lesson is less about any single model and more about the pace: the model you pick this quarter will have a successor before your next board meeting.

This briefing sorts the releases by what they mean for a company building on top of them. Dates come from each vendor's own changelog or launch post, as of October 28, 2026.

What shipped, in one table

ClaudeGPT Si, a free AI news publication, has a dated story with sources on most of these in its AI Models section.

Three things the pace changes

1. Lock-in is now a model problem, not just a vendor problem. OpenAI's changelog shows five GPT-6 family models in under eight weeks. Anthropic shipped four Claude models in five weeks. If your prompts, evals and agent code assume one model's quirks, every release is a migration. Keep a thin layer between your product and the model call, and keep a test set you can rerun in an afternoon.

2. Tiers matter more than flagships. Both big labs now ship a ladder. Anthropic's models overview steers most work to Opus 5.5, keeps Fable 5.1 for the hardest reasoning, and points high-volume tasks at Haiku 5.5. OpenAI describes GPT-6.1 Terra, its October 26 release, as a fast, efficient model for visual reasoning and computer use, below Sol and Astra. Routing easy requests to a smaller tier is usually the quickest margin win a team can find.

3. Open weights are a real option again. Flint, Nanbeige4.3-33B and Unlimited OCR 2 come with permissive licences that allow commercial use, and Moonshot says Kimi K3.1 Pro weights follow by November 4. For document processing, internal tools or regulated data, a self-hosted model can now cover work that needed an API a year ago. Factor in the cost of running it yourself before you commit.

What to do this quarter

• Name one owner for model changes, the way you would for security patches.
• Rerun your evals on each new tier your provider ships, not only the flagship.
• Pilot one open-weight model on a narrow internal task.
• Pick one news source that dates its facts, so the team argues about decisions, not about which model came out when.

Staying informed without a research team

The vendors' own pages are the record, but they are scattered across changelogs, model cards and launch blogs. ClaudeGPT Si reads those primary sources, labels its analysis, and states that companies cannot pay for coverage, per its about page. A founder can scan the models section once a week, or subscribe to the RSS feed and let the releases come to them.

If your company depends on one model for its core product, also watch that provider's deprecations page directly. A retirement date matters more to you than any launch, and it is the one item a general news site may cover a day late.