Skip to content

Everyone’s posting charts about Gemini 3 being “the number one model”.

You know I love benchmarks, but as a builder, I'm more interested in what's under the hood. 👉 Gemini 3 is the first model in a while that feels like a clear #1 across the hardest benchmarks, and it’s already wired into real user surfaces.

3 min read

Originally posted on LinkedIn · 8 reactions · 1 comments · View original →

Everyone’s posting charts about Gemini 3 being “the number one model”. - image 1
Everyone’s posting charts about Gemini 3 being “the number one model”. - image 2

Everyone’s posting charts about Gemini 3 being “the number one model”.

You know I love benchmarks, but as a builder, I'm more interested in what's under the hood.

  • Gemini 3 is the first model in a while that feels like a clear #1 across the hardest benchmarks, and it’s already wired into real user surfaces.

    1. Benchmarks as a capability map

If you zoom out, the benchmark story looks like this:

  • Reasoning & knowledge: top scores on Humanity’s Last Exam and GPQA Diamond. Google DeepMind

  • Math: near‑perfect AIME 2025 and a huge jump on MathArena Apex, going from “basically unsolved” to “meaningfully competent”. Vellum

  • Multimodal & visual: state‑of‑the‑art on MMMU‑Pro, Video‑MMMU, ScreenSpot‑Pro (screen understanding), plus strong OCR and chart reasoning.

  • Coding & agents: competitive on SWE‑Bench, strong on LiveCodeBench and Terminal‑Bench, and leading long‑horizon behaviour on Vending‑Bench 2.

The takeaway for me: we’re not hitting a wall. The ceiling for “hard, messy, multi‑step work” just moved up again.

    1. One “brain”, many surfaces

From day one, the same Gemini 3 brain shows up as:

  • Search AI Mode – dynamic, AI‑generated result UIs on top of web search.

  • Gemini app + Gemini Agent – an assistant that can organise your inbox, act across Calendar/Gmail/Drive, and execute multi‑step tasks.

  • Developer surfaces – Gemini API / AI Studio / Vertex / Firebase, plus Antigravity, an agent‑first VS Code‑style IDE where multiple agents can touch editor, terminal and browser.

Oh, and it does all this with a 1M‑token multimodal context window across text, images, audio and video. 🤯

    1. What this means for people building with AI

A few things I’m paying attention to:

  • No wall mindset: don’t freeze your roadmap around what last quarter’s models could do. Benchmarks like MathArena, ARC‑AGI‑2, ScreenSpot‑Pro and Vending‑Bench 2 say the frontier is still moving fast.

  • Visual & multimodal as native: start designing flows where the natural input is a screen, timeline, repo or video, not a carefully curated text snippet.

  • Agentic UX becomes real work: with Gemini Agent, Antigravity and long‑horizon benchmarks lining up, we’re closer to agents that manage ongoing processes, not just respond to prompts. That puts more weight on questions like when should the agent act vs stay quiet? What can it see and remember?How do we evaluate behaviour over time, not just one‑off answers?

If you’re already playing with Gemini 3 via API, Gemini Agent or Antigravity — I’d love to hear your experience.