Google's November 18, 2025 launch of Gemini 3, with Gemini 3 Pro available in public preview, marks a deliberate push toward agent-first development workflows. The announcement positioned Gemini 3 Pro as a model built for deep multimodal reasoning and practical tool use while outlining where developers and teams can access it: Google AI Studio, Vertex AI, the Gemini CLI and the new Antigravity IDE.
The public preview is documented in Google’s Vertex/Vertex AI materials under the model ID gemini-3-pro-preview and lists a knowledge cutoff of January 2025. The release bundles a very large context window, specialized developer tooling, and an agent-centric UX that together aim to change how engineers build, test and verify code with AI assistance.
Launch, identity and release stage
Gemini 3 and Gemini 3 Pro were officially announced by Google on November 18, 2025, with the Pro tier offered in public preview. Google’s documentation and blog post explicitly record the model ID as gemini-3-pro-preview and detail the preview terms, availability and initial gatekeeping for special variants like Deep Think.
The public preview status matters for developers evaluating tradeoffs between early access and potential rate limits or usage policies. Google framed the preview as broadly accessible: Antigravity is available on Windows, macOS and Linux, and Google described “generous rate limits” for Gemini 3 Pro usage in that preview window.
Official channels (Google’s product blog and Vertex model docs) provide the canonical record of specs, release date and the claims Google made about the model’s capabilities and safety evaluations. Press coverage from outlets like The Verge, VentureBeat and others added independent demos and breakdowns of the initial benchmark claims.
Massive context windows and developer capabilities
Technically, Gemini 3 Pro dramatically increases the context ceiling: Google reports support for roughly 1,048,576 input tokens (≈1M) and about 65,536 output tokens (≈64k). That scale enables workflows where an agent can reason across very large codebases, long document histories or complex multimodal inputs without continual context juggling.
Alongside the context window, Vertex and developer docs list built-in support for code execution, function calling, grounding (Search), structured outputs and streaming/agentic tool use. These features are designed to let models run as active agents that call functions, evaluate results, and integrate external search and runtime outputs into their responses.
For developers this means fewer artificial splits between prompt engineering, tool orchestration and execution. Gemini 3 Pro’s architecture and tooling aim to make it possible to feed the model large design documents, test outputs or full repository contexts and receive structured, executable guidance in return.
Antigravity: an agent-first IDE
Google introduced Antigravity as an “agent-first” integrated development environment built around Gemini 3 Pro. The central idea is to elevate agents to first-class participants in the edit-run-debug loop: agents can access the editor, terminal and embedded browser directly, and Antigravity exposes two main views for that work.
The Editor view includes an agent sidebar where one or more agents can suggest edits, run snippets, or explain decisions inline. The Manager view is focused on orchestration: teams can run, monitor and coordinate multiple agents working on different parts of a project or a single complex task.
Antigravity’s design treats agents like teammates with permissions and audit artifacts rather than opaque function-call logs. That reorientation is intended to make agent-led workflows feel less like sending black-box prompts and more like collaborating with an always-available, disciplined assistant.
Artifacts, verification and agent memory
A key Antigravity innovation is the notion of Artifacts: human-verifiable records that agents produce as they plan and act. Artifacts include task lists, step-by-step plans, screenshots and browser recordings that document what agents will do and what they did, and Google argues these are easier for users to verify than raw tool-call logs.
Antigravity also supports persistent agent memory. Agents can retain snippets and steps from past sessions for reuse, enabling progressive improvement: common fixes, project-specific heuristics and release checklists become part of the agent’s working memory and speed up subsequent sessions.
By combining Artifacts with memory, Antigravity aims to address two perennial concerns: transparency and repeatability. Users can inspect artifacts to confirm intent before actions are taken, and agents can rely on prior lessons to avoid repeating mistakes.
Benchmarks, Deep Think mode and safety
Google published benchmark highlights for Gemini 3 Pro in its launch post. Notable numbers include a WebDev Arena rating of 1487 Elo, Terminal‑Bench 2.0 = 54.2% for terminal/tool use, and SWE‑bench Verified = 76.2% on an agentic coding benchmark. Additional reported figures from press coverage listed top-line performances such as MMMU‑Pro ≈81%, Video MMMU ≈87.6%, MathArena Apex ≈23.4% and ARC‑AGI‑2 ≈31.1% with Humanity’s Last Exam near 37, 38% on some metrics.
Google also introduced a “Deep Think” variant of Gemini 3, intended for deeper, slower reasoning and initially gated for AI Ultra/safety testers. Google and press reports suggest Deep Think scores higher on several complex reasoning tasks, illustrating a trade-off Google is formalizing between throughput and deeper deliberation.
On safety, Google said Gemini 3 underwent the most comprehensive set of safety evaluations to date for a Google AI model, with improvements in prompt-injection resistance and reduced sycophancy. The company published a model card and referenced independent assessments as part of the release narrative.
Developer access, pricing, multi‑model support and ecosystem
Gemini 3 is accessible across Google AI Studio, Vertex AI, the Gemini CLI and Antigravity, and Google listed partnerships with third-party tooling vendors such as Cursor, GitHub, JetBrains, Manus and Replit. Google also said Antigravity supports other models (for example Anthropic Claude Sonnet 4.5 and some open-source models like GPT‑OSS), enabling multi‑model agent workflows.
Preview API pricing reported by trade press put baseline preview rates at roughly $2 per 1M input tokens and $12 per 1M output tokens for prompts up to 200K tokens, with higher tiers (reported) at about $4/$18 per 1M beyond that threshold. Google described generous preview rate limits and media reporting noted limits refresh every five hours, with only a very small fraction of power users likely to hit them during preview.
Sundar Pichai’s announcement framed Gemini 3 in the context of Google’s broader AI reach, citing AI Overviews at about 2 billion users/month, the Gemini app at ~650 million monthly users, and ~13 million developers who have built with Google’s generative models. For developers evaluating adoption, the combination of multi-model support, partner integrations and accessible entry points lowers the barrier for trying agent-first code workflows.
Vibe coding and generative interfaces
Google described Gemini 3 as enabling “vibe coding” and richer generative UIs where natural language, sketches or voice can create interactive apps, UI components and even 3D or voxel art. Demonstrations shown at launch highlighted rapid production of interactive code artifacts from high-level prompts and multimodal inputs.
Vibe coding reframes the act of building as iterative guidance: a developer describes the intent, agents generate scaffolded code and prototypes, then the developer refines the design with targeted instructions. Antigravity’s integrated editor, terminal and browser access shortens feedback loops so that the generated UI components can be run and validated immediately.
As those interfaces mature, the boundary between prototyping and production blurs, developers still need rigorous tests and human verification, but generative interfaces promise faster exploration, new UX patterns for non-coders, and practical boosts in productivity for engineers.
The combination of Gemini 3 Pro’s large context, developer-focused tooling and Antigravity’s agent-centric UX marks a meaningful step toward agent-first coding. Artifacts and persistent memory are explicit design choices that prioritize transparency and iterative improvement, while Deep Think and safety evaluations show Google aiming to balance capability with responsible release practices.
For teams and individual developers, the new stack, Gemini 3 Pro via AI Studio, Vertex, CLI or Antigravity, offers a flexible way to experiment with agentic workflows, multi-model orchestration and generative UI patterns. As the preview expands and independent benchmarks and developer experiences accumulate, Antigravity and Gemini 3 Pro will be important to watch for anyone serious about the future of coding with AI.
