Short answer: The reported benchmarks are mixed and do not establish an overall winner. Google documents that the Gemini Developer API model Gemini 3 Pro Preview (gemini-3-pro-preview) shut down on March 9, 2026. That establishes its status on the documented Developer API route; the sources cited here do not establish whether a separate Vertex AI offering remains available. OpenAI’s API documentation lists GPT-5.2 as a previous flagship. So this is a historical model comparison—not evidence that two currently callable API endpoints are available to compare. Status checked November 9, 2026. (Google Gemini API model page; OpenAI GPT-5.2 API model page)
First, clarify what “Gemini 3 vs ChatGPT 5.2” means
“ChatGPT 5.2” can refer to a model in the ChatGPT product or to an API model. Those are different access and billing routes. OpenAI’s launch announcement said GPT-5.2 Instant, Thinking, and Pro would be available in ChatGPT and the API. The current API model page lists gpt-5.2 and the dated snapshot gpt-5.2-2025-12-11; the reviewed material does not establish a one-to-one mapping between that priced API entry and every Thinking or Pro benchmark row below. (OpenAI launch announcement; OpenAI API model page)
“Gemini 3” also refers to distinct variants. The benchmark roundup reports separate results for Gemini 3 Pro and Gemini 3 Deep Think; do not combine their scores. Google’s model page identifies the Gemini 3 Pro Preview model code as gemini-3-pro-preview and says the Gemini Developer API model shut down on March 9, 2026. Google’s launch announcement described preview access through the Gemini API in Google AI Studio and through Vertex AI for enterprise customers. The cited material does not establish the current status of a separate Vertex AI offering, so do not assume either that it remains available or that it was retired. (Google model page; Google launch announcement)
| API comparison point | OpenAI GPT-5.2 | Google Gemini 3 Pro Preview |
|---|---|---|
| Documented model identifier | gpt-5.2; snapshot gpt-5.2-2025-12-11 |
gemini-3-pro-preview |
| Status in the cited documentation | Listed by OpenAI as a previous flagship | Google says the Gemini Developer API model shut down March 9, 2026; current Vertex AI status is not established here |
| Documented token limits | 400,000-token context window; up to 128,000 output tokens | Historical model card: 1,048,576 input-token limit; 65,536 output-token limit |
| Price shown in the cited material | $1.75 per 1M input tokens; $0.175 per 1M cached input tokens; $14 per 1M output tokens | Historical launch rate: $2 per 1M input and $12 per 1M output for prompts of 200,000 tokens or less—not a current quote |
The Gemini limits and rates are historical model-card and launch information, not evidence that the Gemini Developer API model can still be called. The cited GPT-5.2 model page lists an API entry and rates, but the benchmark labels “Thinking” and “Pro” should not automatically be treated as interchangeable with that priced entry. The roundup also reports Gemini 3 Deep Think scores, but the cited material does not establish a corresponding API identifier, access route, or price for that variant. (OpenAI API model page; Google Gemini API model page; Google launch announcement)
Before selecting a model for an image, audio, or video workload, verify that the exact currently available API model supports the required inputs and outputs. The evidence cited here does not establish a complete modality inventory for the compared models.
What the reported benchmarks show
The figures below are from a December 11, 2025 comparison. Its author describes them as vendor-reported and pending independent verification. They compare specific model variants, not one generic “Gemini 3” with one generic “ChatGPT 5.2.” Read the benchmark roundup.
| Benchmark | GPT-5.2 Thinking | GPT-5.2 Pro | Gemini 3 Pro | Gemini 3 Deep Think | What the reported figures suggest |
|---|---|---|---|---|---|
| ARC-AGI-2 | 52.9% | 54.2% | 31.1% | 45.1% | The listed GPT-5.2 variants score higher than either listed Gemini variant. |
| GPQA Diamond | 92.4% | 93.2% | 91.9% | 93.8% | Gemini 3 Deep Think is slightly above GPT-5.2 Pro in this table. |
| Humanity’s Last Exam | 34.5% | 36.6% | 37.5% | 41.0% | Gemini 3 Deep Think is highest among these entries. |
| SWE-bench Verified | 80.0% | Not reported | 76.2% | Not reported | The table gives GPT-5.2 Thinking a small lead over Gemini 3 Pro; “not reported” does not mean zero. |
The ordering changes by benchmark: the table favors GPT-5.2 on ARC-AGI-2, but Gemini 3 Deep Think on Humanity’s Last Exam and GPQA Diamond. These vendor-reported results are not a controlled test of a production workload, and the variants are not interchangeable. The roundup’s AIME 2025 table labels the comparison “no tools,” while its prose says Gemini 3 Pro’s score used code execution. Because those conditions conflict, that result is not used here as a like-for-like comparison.
The table cannot tell you which model will better handle your repository, documents, or agent workflow. It also does not establish current API availability, latency, reliability, or production cost.
Production API fit and cost
OpenAI’s API documentation lists GPT-5.2 at $1.75 per 1 million input tokens, $0.175 per 1 million cached input tokens, and $14 per 1 million output tokens. It lists a 400,000-token context window and a 128,000-token maximum output. OpenAI labels GPT-5.2 a previous flagship. (OpenAI GPT-5.2 API model page)
Google’s November 2025 launch announcement listed Gemini 3 Pro Preview at $2 per 1 million input tokens and $12 per 1 million output tokens for prompts of 200,000 tokens or less. Those are historical launch rates for the Gemini API route: Google documents that gemini-3-pro-preview shut down on the Gemini Developer API on March 9, 2026. This does not establish the current status or pricing of a separate Vertex AI offering. (Google launch announcement; Google Gemini API model page)
Historical token-cost illustration—not a current API comparison
For illustration, suppose a request uses 100,000 input tokens and 20,000 total billable output tokens. The output assumption includes visible answer tokens and any reasoning tokens billed as output; it excludes cached input, tools, and other billable services.
- GPT-5.2 API rates: (100,000 ÷ 1,000,000 × $1.75) + (20,000 ÷ 1,000,000 × $14) = $0.455.
- Gemini 3 Pro Preview launch rates: (100,000 ÷ 1,000,000 × $2) + (20,000 ÷ 1,000,000 × $12) = $0.44.
At those historical rates, the Gemini example is 1.5 cents cheaper for that token mix. This is not a current two-provider API price comparison: the Gemini 3 Pro Preview model is documented as shut down on the Gemini Developer API, and the cited evidence does not establish the current status of a separate Vertex AI offering. The GPT-5.2 API price also does not establish the price of a separately labeled Pro benchmark variant. OpenAI documents that reasoning tokens are billed at the output-token rate, so budgeting only visible response text can understate output usage. (OpenAI model page; OpenAI usage documentation)
Keep API charges separate from ChatGPT or Google AI consumer subscriptions. A subscription price does not establish per-token API pricing or access to a particular API model.
How to decide for a real workload
- If you are evaluating GPT-5.2: OpenAI documents the
gpt-5.2API entry and its rates, while labeling it a previous flagship. Confirm the precise model and settings you intend to call; do not transfer Thinking or Pro benchmark results to a priced API variant without confirming the mapping. - If you are evaluating Gemini 3 Pro: treat its published benchmark results and preview-era price as historical. Google documents that the Gemini Developer API model ID
gemini-3-pro-previewshut down on March 9, 2026. The cited research does not confirm a separately callable Vertex AI offering today, so verify any Vertex-specific access directly before relying on it. Gemini 3 Deep Think’s reported scores likewise do not establish a currently callable, priced API candidate. - For a new production comparison: choose exact, currently documented API model IDs first, then run the same representative inputs through both. Verify required modality support for the exact models, score task success and error severity, and record latency and billed input/output tokens—including reasoning tokens and any caching or tools used. This is a proposed pilot, not a test performed for this article. A newer Gemini variant would be a new comparison; the Gemini 3 Pro Preview benchmark scores above should not be carried over to it.
Bottom line: the historical benchmark evidence is mixed. Google documents that the Gemini 3 Pro Preview model shut down on the Gemini Developer API, while the cited evidence does not establish current Vertex AI availability. GPT-5.2 remains listed by OpenAI as a previous-flagship API model. The sources therefore support a historical comparison, not a current, executable two-model purchasing recommendation. Verify the exact live model IDs, access route, required modality support, and rates before choosing a production pair.
Sources
Gemini 3 Pro preview | Gemini API | Google AI for Developers
Gemini 3 for developers: New reasoning, agentic capabilities
Observability and usage | OpenAI API
How GPT-5.2 stacks up against Gemini 3.0 and Claude Opus 4.5