As of November 9, 2026, the lifecycle status is different for the two exact API models discussed here: OpenAI labels gpt-5-2025-08-07 deprecated, while Google says gemini-3-pro-preview was shut down on March 9, 2026 in the Gemini API. Deprecation is not itself proof that GPT-5 can no longer be called. The available OpenAI model page does not state a shutdown date, so check the model’s current status before deciding whether an existing integration can continue. (OpenAI’s GPT-5 model page; Google’s Gemini API model page)
Quick verdict
- Starting a new OpenAI API integration? GPT-5 is marked deprecated, so favor a currently supported model rather than building on it. That is lifecycle-risk advice—not a claim that GPT-5 has already shut down.
- Using GPT-5 now? OpenAI’s model page identifies
gpt-5-2025-08-07and lists endpoints including Responses API (v1/responses) and Chat Completions (v1/chat/completions). Because the model is deprecated, verify its shutdown date and current availability for your account before relying on it. (OpenAI’s GPT-5 model page; OpenAI’s API deprecations guidance) - Considering
gemini-3-pro-previewthrough the Gemini API? Google’s model page says it was shut down on March 9, 2026, so it is not a current option through that API. This notice does not, by itself, establish the model’s lifecycle status on Vertex AI or another distribution channel. (Google’s Gemini API model page)
This is a comparison of the OpenAI API’s GPT-5 model entry and the Gemini Developer API’s Gemini 3 Pro Preview entry—not a general comparison of every GPT-5 or Gemini 3 family model. GPT-5, GPT-5.1, GPT-5.2, and Gemini 3 Pro Preview are not interchangeable labels.
What the original API model specifications said
The following are specifications listed on those model pages; they do not establish current callability for every account or channel.
| API model | Documented inputs and limits | Other listed capabilities | Lifecycle status on the cited page |
|---|---|---|---|
GPT-5 (gpt-5-2025-08-07) |
Text and image input; text output; 400,000-token context window; up to 128,000 output tokens | Function calling and structured outputs; the page lists Responses API and Chat Completions among the supported endpoints | OpenAI labels it deprecated. The model page does not state a shutdown date. (OpenAI) |
Gemini 3 Pro Preview (gemini-3-pro-preview) |
Text, image, video, audio, and PDF input; text output; 1,048,576-token input limit; up to 65,536 output tokens | Function calling, code execution, search grounding, URL context, and thinking | Google says it was shut down in the Gemini API on March 9, 2026. (Google) |
A larger context limit is a capacity specification, not evidence that a model will be more accurate, faster, or cheaper for a particular workload. These historical specifications also do not establish a head-to-head quality, latency, or reliability winner.
Which differences would have mattered for a workload?
- Large mixed-media inputs: Gemini 3 Pro Preview’s Gemini API documentation listed video, audio, and PDF inputs, as well as a larger input-token limit. Those capabilities would matter for workflows involving long recordings or large multimodal documents—but the documented Gemini API model has been shut down. (Google)
- Text-and-image tasks requiring structured output or function calls: GPT-5’s OpenAI API documentation listed text and image inputs, function calling, and structured outputs. These are relevant features to check in a replacement model. GPT-5 itself is marked deprecated, not documented on that page as shut down. (OpenAI)
- Tool use or grounded answers: The Gemini API page listed code execution, function calling, and search grounding for Gemini 3 Pro Preview. Check that the exact replacement model and API surface support the tools your application requires; do not assume family-level names guarantee feature parity. (Google)
Production API fit: what to check before migrating
If you mean the exact original models
For the OpenAI API, GPT-5 is deprecated, and its model page lists gpt-5-2025-08-07 plus endpoints including Responses API and Chat Completions. That page does not establish whether the model is callable for your account today or give a shutdown date. OpenAI’s deprecations guidance says to check the model’s shutdown_date through the List models or Retrieve a model endpoint; use that information, along with any account-specific notice, before deciding whether an existing integration can continue. (OpenAI’s GPT-5 model page; OpenAI’s API deprecations guidance)
For the Gemini Developer API, Google explicitly says gemini-3-pro-preview was shut down on March 9, 2026. That statement is specific to the Gemini API documentation cited here. I have not established the same lifecycle status for Vertex AI or another distribution channel; check the applicable model documentation if that is where you deploy. (Google’s Gemini API model page)
A deprecated model is a lifecycle-risk choice for a new integration, even when the documentation does not establish that it is already unavailable. A shut-down model is a different case: Google’s notice says the Gemini API model has already reached that status.
If you mean the broader model families
Compare the specific model IDs and API surfaces you can actually use rather than carrying older GPT-5-versus-Gemini-3 claims forward. Check each candidate’s lifecycle status, input modalities, context and output limits, function/tool-calling behavior, rate limits, data handling, and migration policy in the documentation for the relevant provider and API.
A practical selection test
Suppose your application summarizes customer-submitted PDFs and images, extracts fields into a fixed JSON schema, and calls an internal function when a condition is met. Do not decide from a generic benchmark or a retired model’s context limit. Instead:
- Select currently callable candidate API models and record their exact IDs and lifecycle status.
- Run the same representative inputs and instructions through each, including difficult PDFs, image-heavy examples, incomplete records, and cases where the correct outcome is to abstain or request clarification.
- Score schema validity, extraction accuracy, correct function selection, failure handling, latency, and cost using your expected input and output volumes.
- Compare only candidates that meet your minimum quality and operational requirements.
This is a proposed evaluation procedure, not a test we conducted. It gives you a workload-specific decision without treating historical qualitative comparisons as proof of a current winner.
Pricing: verify the active model IDs
Do not use historical GPT-5 or Gemini 3 Pro Preview prices as the basis for a new implementation decision without checking the current rates for the exact model IDs and API tiers you plan to use. Compare like-for-like billing categories—input, cached input, and output tokens—and include applicable tool or service charges. Token rates alone do not establish total workload cost. (OpenAI API pricing; Google Gemini API pricing)
Bottom line
For a new integration, do not treat “Gemini 3 vs GPT-5” as a settled choice between two equally available APIs. OpenAI marks GPT-5 deprecated, but the model page does not say it has shut down; confirm its shutdown date and callability for your account. Google says Gemini 3 Pro Preview was shut down in the Gemini API on March 9, 2026; do not extend that conclusion to Vertex AI without channel-specific evidence. Choose current, exact model IDs and test them on your own workload before deciding which performs better.