Gemini 3's Launch: Benchmarks, Antigravity, and Agents

RÉSUMÉ

En 30 secondes

Matthew Berman reviews Gemini 3's launch-era benchmark claims, Google Antigravity, and an assistant demonstration for organizing email. The short video describes initial features and availability; rankings, access, pricing, and limits can change, and benchmark scores depend on test conditions.

1Launch benchmarksBerman cites strong results including a perfect AIME 2025 score; such a number requires the exact model and evaluation conditions and is not a universal capability claim.
2AntigravityGoogle's development platform lets users coordinate coding agents; its free preview and quotas were launch-time details.
3Assistant workflowsA demo shows an agent proposing ways to organize an inbox for the user to review.

À retenir

Gemini 3's launch included benchmark results and a multimodal context window.
Google introduced Antigravity as an agent-oriented development environment.
A Gemini Agent demo showed proposed email organization for user review.
Launch-era scores and availability should not be assumed to describe current products.

ThèmesGemini Benchmarks Agents de code Personal agents

TEXTE INTÉGRAL

Lire le texte remanié, chapitre par chapitre

Afficher la transcription

Transcription seule

Gemini 3 and benchmarks

Berman summarizes model performance reported at launch.

DIAPOSITIVE 1 / 2Launch-era model leaderboardThe displayed ranking reflects a particular leaderboard snapshot, not a permanent ordering.

In November 2025, Matthew Berman introduces Google's Gemini 3 and points to launch-era leaderboard and math benchmark results. He cites a perfect AIME 2025 score; benchmark results depend on the model variant, settings, and tools, and do not establish that a model is best at every task. Google's later comparison lists Gemini 3 Pro at 95.0% on AIME 2025 without tools.

Antigravity and agents

He describes Google's development platform and an email organization demo.

DIAPOSITIVE 2 / 2Gemini Agent demonstrationThe launch page describes an assistant that plans multi-step tasks with user control.

Berman also highlights Google Antigravity, a development environment designed around coding agents. He says Gemini 3 Pro was available in its free preview at the time. Pricing, quotas, and availability can change, so this historical description is not a current offer.

A Gemini Agent example shows an assistant proposing ways to organize an email inbox, with the user reviewing suggested actions. Berman also cites a one-million-token input context and 64,000-token output limit for the launch model, plus text, image, video, code, and audio inputs. Actual capabilities and limits depend on the specific product and model version.