I counted every AI model released in Q1 2026

Every quarter I do this. Sit down, go through the announcements, the papers, the blog posts, and count how many new AI models actually shipped. Not announced. Not teased. Shipped, with weights or API access.

Q1 2026 was the busiest quarter I've ever tracked.

The count

| Month | Models Released | Notable | |-------|----------------|---------| | January | 12 | DeepSeek R2, Qwen 3, Mistral Medium 3 | | February | 14 | Claude Opus 4.5, Gemini 2.5 Ultra, Grok 3.5 | | March | 16 | GPT-5.2, Llama 4, Claude Opus 4.6 (1M context) | | Total | 42 | |

Forty-two models. In ninety days.

For context, in all of 2021 there were maybe 15 notable model releases total. We're now doing triple that every quarter.

January was DeepSeek's month

DeepSeek R2 dropped on January 8 and it was the open source story of the quarter. The reasoning capabilities caught everyone off guard. Again. They did it with R1 last year and somehow people were still surprised.

Qwen 3 came out the same week. Good timing or bad timing? Hard to say. It got buried in the DeepSeek coverage.

February belonged to the big labs

Claude Opus 4.5 was the headline. Anthropic's pitch was simple: better at creative writing, harder to jailbreak, same price as Opus 4. The benchmark improvements were modest (2-3% on most things) but the vibes were noticeably different. People who tried it said it "felt" smarter, which is exactly the kind of unquantifiable claim that makes data people like me twitch.

Gemini 2.5 Ultra was Google's response. Better multimodal, better at long context. The million-token context window is standard now. Remember when 4K was normal? That was 2023. Three years ago.

March was chaos

GPT-5.2. Llama 4. Claude Opus 4.6 with 1M context. All in the same month.

I genuinely don't know how anyone is supposed to evaluate all of these. By the time you finish benchmarking one model, three more have shipped. The leaderboards are updating faster than I can update my spreadsheets.

The number that worries me

Here's what I keep coming back to: of the 42 models released in Q1 2026, how many will anyone remember in a year?

I went back and checked Q1 2025. Thirty-one models. I can name maybe eight of them from memory. The rest? Absorbed, deprecated, or forgotten.

The release velocity is impressive. But velocity and impact aren't the same thing. Most of these models are incremental improvements that don't change what you can build. The ones that matter (DeepSeek R2 for open source reasoning, Claude Opus 4.6 for context length, Llama 4 for local deployment) are maybe five or six out of forty-two.

Fourteen percent. That's the signal-to-noise ratio.

What I'm watching for Q2

Two things.

First, whether the mid-tier gets another price cut. GPT-4o mini and Claude Haiku 4 are both under $0.25/M tokens. If that drops to $0.10, it changes the economics of every AI-powered product.

Second, whether anyone releases a genuinely new architecture. All 42 models this quarter were transformer-based. Bigger, faster, cheaper transformers, but still transformers. The next real leap probably requires something different. I don't know what. Nobody does. But the data will tell us when it happens.

I'll keep counting.

Model release data cross-referenced with Epoch AI's notable models database.

-- dataku