Google's Gemini 3 Is Here: A Special Early Look

Summary of Google's Gemini 3 Is Here: A Special Early Look

by The New York Times

26m•November 18, 2025
Listen to Episode

Overview of "Google's Gemini 3 Is Here: A Special Early Look"

This Hard Fork episode (The New York Times) gives an early look at Google’s newly announced Gemini 3 model. Hosts Kevin and Casey summarize Google’s briefing, the model’s headline capabilities, benchmark gains, rollout plan, and an interview with Demis Hassabis (CEO, Google DeepMind) and Josh Woodward (VP, Gemini). The conversation covers technical improvements (reasoning, multimodal generation, coding), new generative interfaces and agents, safety/testing, competitive implications, and product integration plans.

Key takeaways

  • Gemini 3 (Pro) shows substantial gains over Gemini 2.5 Pro across many benchmarks; Google positions it as state of the art.
  • New visible features: generative interfaces (custom interactive UIs), much-improved coding (including “vibe coding”), better long-form reasoning and multi-step thinking, and early agent features (e.g., inbox triage & reply suggestions).
  • Notable benchmark: “Humanity’s Last Exam” — Gemini 2.5 Pro ≈ 21.6%, Gemini 3 Pro ≈ 37.5%.
  • Rollout: available in the Gemini app and Google Search “AI” side tab this week; wider integration into Docs/Gmail/Workspace not yet dated.
  • Google is offering U.S. college students one year of free access to a paid Gemini tier.
  • Google emphasizes efficiency (cost-to-performance) to serve billions of users; they claim heavy safety testing and external review.
  • Leadership claims: Google believes it’s back in strong form and focused on rate-of-progress rather than declaring an outright “win” in the AI race.
  • Demis Hassabis retains a multi-year AGI timeline (roughly 5–10 years), expecting one or two further groundbreaking advances (memory, world models, reasoning).

What Gemini 3 can do (concrete examples)

  • Build interactive, custom interfaces on-the-fly (example: interactive tutorial about Van Gogh; a mortgage calculator for high-value homes).
  • Enhanced coding support — including front-end and creative “vibe coding” use cases.
  • Multimodal generation and best-in-class image editing (suggested demo: selfie edits on phone).
  • Agent-style workflows (e.g., reading your inbox, summarizing and proposing replies, organizing threads) — early demos shown as animated GIFs.
  • Improved long-chain reasoning: better at maintaining context and multi-step problem solving.

Benchmarks & performance

  • Broad improvement: Google presented a dozen+ benchmarks where Gemini 3 outperforms 2.5 Pro.
  • Highlighted metric: Humanity’s Last Exam (graduate/PhD-level interdisciplinary test) — 21.6% → 37.5%.
  • Other proxies mentioned: LM Arena ELO (cracking 1500), coding and multimodal benchmarks.
  • Google stresses benchmarks are proxies; product user satisfaction matters most.

Availability, rollout & pricing/offer details

  • Immediate availability: Gemini app and Google Search AI tab (this week).
  • Developer access: various product channels (details forthcoming).
  • Wider integration into core Google products (Docs, Gmail, Workspace) will follow but no firm timeline announced.
  • Offer: one year of paid Gemini access free for U.S. college students.
  • Google signals model-family variety (performance vs. cost tradeoffs) beyond the Pro variant.

Safety, testing & risks

  • Google reports extensive internal and external safety testing; calls Gemini 3 their most thoroughly tested model yet.
  • Specific safety attention on tool/function calling — valuable for development but increases cyber-misuse risk, so Google says it applied extra caution.
  • Google acknowledges more capabilities increase risk and they are monitoring misuse vectors.
  • Hosts noted industry competition and regulatory/ethical debates; Google positions itself to move carefully but broadly.

Industry context & strategy

  • Competitors reportedly nervous; some see Gemini 3 as a potentially disruptive release that could shift competitive positioning.
  • Google’s advantage: massive product ecosystem (Search, Maps, YouTube, Android, Workspace) to inject Gemini features at scale — both a distribution and data advantage.
  • Google emphasizes cost-efficiency and model distillation techniques to serve large user bases affordably.
  • Discussion on AI valuation: Hassabis says parts of the AI space may be in a bubble (e.g., overvalued seed rounds) but there’s also real opportunity in robotics, gaming, drug discovery, and product integration.

Notable quotes & lines

  • Hosts: “Learn anything” — a repeated phrase in Google’s messaging presenting Gemini as a learning/creation tool.
  • Casey (joking): “Step one, build an illegal monopoly.” — a quip about Google’s ecosystem advantage.
  • Demis Hassabis: Gemini 3 is “right on track” with past expectations; still expects “one or two more breakthroughs” for full AGI-level consistency.
  • Josh Woodward: Gemini 3 is “more succinct… more to the point” and “pleasant to brainstorm with.”

Practical things to try (suggested demos)

  • Try an image-edit selfie demo in the Gemini app (easy, visible test of multimodal/image capabilities).
  • Ask Gemini 3 to create a custom interface for a learning topic (e.g., quick interactive tutorial on an artist or concept).
  • Test coding help (front-end tasks, “vibe coding” prompts) if you’re a developer.
  • When agent features roll out, try inbox triage/summarization and suggested replies (watch for privacy controls and permissions).
  • Students: redeem the one-year free paid access if eligible.

Podcast/host context & disclosures

  • The hosts disclosed The New York Times Company is suing OpenAI and Microsoft over model training practices; one host mentioned a partner works at Anthropic. These are editorial disclosures included in the episode.

Bottom line

Gemini 3 represents a substantial iterative leap for Google’s multimodal foundation models: stronger reasoning, much better coding capabilities, and a new focus on generative, interactive interfaces and agent workflows. Google is positioning Gemini 3 for scale via efficiency gains and staggered integration into its massive product ecosystem. Safety testing and careful rollout are emphasized, and Google still frames AGI as multiple breakthroughs away even while celebrating the progress. The NYT team plans deeper hands-on testing and reviews once broader access is available.