Every AI model released in 2026, in order.

As of September 28, 2026, the newest models from the biggest AI labs are Claude Sonnet 5.5 (Anthropic, September 28), GPT-6 Astra and GPT-6 Sol and Luna (OpenAI), Gemini 3.8 Flash (Google, September 2) and Grok 4.7 (xAI, September 21). OpenAI released Astra on September 3 and Sol and Luna on September 22.

The largest open-weight releases came from DeepSeek, Alibaba’s Qwen team, Moonshot AI and Z.ai, and each shipped a new one between July and September.

Below is every major model release of 2026 in date order, each linked to the lab’s own announcement. We cover language models and leave out image, video, voice and music models.

01What are the latest AI models right now?

The newest flagship from each major lab, checked on September 24 to 28, 2026, against each lab’s announcements and model documentation:

Which one to pick depends on your task and budget, which our guide on how to choose an AI model walks through. Prices per million tokens are in our API pricing comparison.

02The 2026 model timeline, January to March

Each date is the one on the lab’s own announcement, checked on September 24 to 28, 2026. A morning launch in Beijing or Hangzhou can fall on the previous day in the Americas.

03The 2026 model timeline, April to June

The second quarter brought an invitation-only frontier model, Claude Mythos Preview, and a US export control that briefly took Claude Fable 5 offline.

Access to Fable 5 and Mythos 5 lasted three days. On June 12, the US government applied export controls to both, and Anthropic suspended them for all users. The controls were lifted on June 30, and Fable 5 came back for everyone on July 1, according to Anthropic.

Google’s May 19 post said Gemini 3.5 Pro would roll out “next month.” As of September 28, 2026, Google’s model list still names Gemini 3.1 Pro, in preview, as its Pro model.

04The 2026 model timeline, July to September

The third quarter was the busiest: nineteen launches between July 9 and September 28, ending with three new frontier models and a new Sonnet in September.

Anthropic said Claude Sonnet 5.5 and Haiku 5.5 would follow “in the coming weeks.” Sonnet 5.5 shipped on September 28; Haiku 5.5 had not been released when we checked that day.

05Which open-weight models came out in 2026?

The biggest open-weight releases of 2026 came from Chinese labs, and they grew into the trillions of parameters. Open weights mean you can download and run the model yourself, which our guide to open vs closed AI models explains, license catches included.

• Largest: Kimi K3 (2.8 trillion parameters, custom Kimi K3 license) and Qwen3.8-Max (2.4 trillion, 95 billion active).

• Permissive licenses: Gemma 4 and Mistral Small 4 use Apache 2.0; GLM-5 and DeepSeek-V4.1-Flash use MIT.

• Laptop-sized: Gemma 4 runs from E2B up to 31B, and Qwen3.6 added a dense 27B model.

• Not in 2026: we found no new Llama model on Meta’s Hugging Face page, and OpenAI’s newest general-purpose open models are still the gpt-oss models from August 2025.

We checked each open-weight release against its Hugging Face repository on September 24, 2026. Our guide to running AI models locally covers the hardware each size needs.

06How do I know which AI model I am using?

Look at the tool, not the answer. A model’s reply about its own name comes from its training data and instructions, so it can name an older version or the wrong family entirely.

1. 01

   In a chat app, check the model picker

   The name in the picker is the model that answers. If the app has an automatic mode, it may route your request to a different model, so pick one yourself when it matters.

2. 02

   In a coding agent, run /status

   In Claude Code, /status shows the version, model and account, and /model switches models. In OpenAI’s Codex CLI, /status confirms the active model and /model changes it.

3. 03

   In an API, read the response and pin a version

   Every major API returns the model it used in its response. Aliases move over time: OpenAI’s gpt-5.6 alias pointed to GPT-5.6 Sol at launch, and Google’s gemini-flash-latest pointed to Gemini 3.5 Flash in May. Pin a full model ID, such as claude-haiku-4-5-20251001, when results must not change.

4. 04

   Check the date on the announcement, not the download page

   A Hugging Face repository can exist weeks before a launch. Mistral Small 4’s repository was created on January 23 and the model was announced on March 16. Kimi K3’s repository dates from June 13, a month before its July 16 launch.

07How fast are new AI models coming out?

About one major language model every six days in 2026, counting only the 46 releases in this list. Anthropic alone shipped a new Opus model about every two months, from Opus 4.6 in February to Opus 5.5 in September.

The pace makes benchmark claims age quickly. A model that leads in July can sit mid-table by September, which our coding agent leaderboard shows with dates. Before switching, run your own small test: our guide to AI benchmarks explains how to read the numbers labs publish at launch.

08FAQ

Read next: reasoning models explained for the effort settings most of these models now share, or how to choose an AI model for picking one.