Direct answer
Google’s model-card catalogue lists a Gemini 3 Pro model card. The Gemini API model gemini-3-pro-preview was shut down on March 9, 2026; that shutdown applies to the Gemini API and does not, by itself, establish the model’s status on Vertex AI. Google’s deprecation page names gemini-3.1-pro-preview as the recommended replacement for Gemini API users. (Google DeepMind model-card catalogue; Gemini API deprecations)
A Hacker News discussion identifies a PDF as “[Gemini 3 Pro] External Model Card - November 18, 2025 - v2” and transcribes selected contents. The discussion is secondary reporting; it is not a substitute for the Google-hosted model card. The specifications below cover selected details supported by that card, not a review of all its disclosures. (Hacker News discussion)
What the model card and API documentation establish
The Google-hosted model card describes Gemini 3 Pro as a sparse mixture-of-experts transformer. It lists text, image, audio, and video inputs, text output, a context window of up to 1 million tokens, and output of up to 64,000 tokens. These are stated model specifications; they do not establish that every API surface or deployment has identical limits. (Google-hosted Gemini 3 Pro model card)
Google’s Gemini API model page identifies the preview model as gemini-3-pro-preview and lists 1,048,576 input tokens and 65,536 output tokens. It also marks the model as shut down. Those details are useful for understanding a former integration, not as instructions to start a new one. (Gemini 3 Pro Preview model page; Gemini API deprecations)
The Hacker News thread reports that the model card gives a January 2025 knowledge cutoff. Treat that date as a secondary-reported detail until checked against the card; a stated cutoff is not a guarantee that the model knows or correctly answers every fact from before that date. (Hacker News discussion)
What the benchmark evidence means for selection
Google’s November 18, 2025 enterprise announcement highlights multimodal understanding, coding and front-end generation, long-context work, and tool use and planning. It also reports a 1,501 Elo LMArena score for Gemini 3 Pro. These are vendor claims in a launch announcement; customer results quoted in the post are attributed testimonials, not independent evaluations. (Google Cloud announcement)
The announcement says Gemini 3 has a 1-million-token context window in its coding section. That statement alone does not establish that Gemini 3 Pro has the same limit across API surfaces; check the current product-specific documentation before relying on it. (Google Cloud announcement)
A Hacker News commenter reproduces a benchmark table attributed to the card. For SWE-Bench Verified, it lists Gemini 3 Pro at 76.2%, Gemini 2.5 Pro at 59.6%, Claude Sonnet 4.5 at 77.2%, and GPT-5.1 at 76.3%. This is a secondary transcription, and the captured discussion does not provide enough methodological detail to establish direct comparability. It does not show Gemini 3 Pro leading every task. Treat the scores as a reason to test a candidate, not as a universal ranking. (Hacker News discussion)
For repository-level coding, evaluate a currently available model on representative issues from your own repositories, using your tools, test commands, and acceptance criteria. Compare task quality and tool errors alongside latency and token use; a benchmark score alone cannot predict your team’s results.
Production API status and migration
Google’s Gemini API deprecation page says gemini-3-pro-preview was shut down on March 9, 2026, and names gemini-3.1-pro-preview as the recommended replacement. The confirmed shutdown is specific to the Gemini API. (Gemini API deprecations)
The replacement is a different model, not a guarantee of identical outputs. Before migrating, check the replacement’s lifecycle status, model ID, inputs and outputs, limits, regions, access conditions, and commercial terms for the API product you intend to use.
Google’s November 18, 2025 announcement said Gemini 3 Pro was in preview on Vertex AI and Gemini Enterprise at that time. That historical statement does not establish current Vertex AI access or lifecycle status for Gemini 3 Pro; verify that separately. The Gemini API shutdown should not be extended to Google Cloud without product-specific evidence. (Google Cloud announcement)
Pricing: historical figures, not a current quote
The Hacker News thread reports historical Gemini 3 Pro API rates of US$2.00 per 1 million input tokens and US$12.00 per 1 million output tokens for contexts of 200,000 tokens or less, and US$4.00 input / US$18.00 output per 1 million tokens above 200,000. These are forum-reported historical figures, not current verified prices. Because gemini-3-pro-preview is shut down in the Gemini API, they are not a current price quote for that endpoint. They are also distinct from consumer subscriptions or enterprise plans. (Hacker News discussion)
For a purchase decision, check current official pricing for the specific replacement model and API or cloud product you plan to use. The captured Gemini 3 Pro API documentation establishes the preview model’s shutdown and technical details, but not a current price for it. (Gemini 3 Pro Preview model page)
Selection takeaway
The Gemini 3 Pro model card remains useful technical reading, but gemini-3-pro-preview is retired in the Gemini API. If you maintain an integration using that API model, use Google’s named replacement as a migration starting point and regression-test your application. For Vertex AI, confirm Gemini 3 Pro’s current product-specific status separately. For a new workload, base the choice on current documentation, terms, pricing, and evaluation on your own tasks—not on launch-era availability or a forum-transcribed benchmark table. (Google DeepMind model-card catalogue; Gemini API deprecations)
Sources
[Gemini 3 Pro] External Model Card — Google-hosted PDF
Gemini deprecations — Google AI for Developers
Gemini 3 Pro Preview — Gemini API