GPT-5.2 vs Gemini 3: the short answer
In one published coding comparison, the result depended on the task: Gemini 3.0 did better on a tightly specified Python task, while GPT-5.2 was reported as more thorough on two larger codebase tasks. That is a limited coding comparison—not a universal ranking or an API performance benchmark.
The model names matter. Kilo’s historical tests used Gemini 3.0. The API comparison below concerns the separately named Gemini 3.1 Pro Preview; the available documentation excerpt says the original Gemini 3 Pro Preview endpoint was shut down on March 9, 2026. The available lifecycle excerpt, dated February 19, 2026, listed no shutdown date for Gemini 3.1 Pro Preview. Its status and access conditions today are not confirmed here. Do not assume the coding results for Gemini 3.0 predict Gemini 3.1 Pro Preview. (Google Gemini model page, Gemini API deprecation schedule)
For the historical coding-test pair, consider Gemini 3.0 for a narrowly specified coding task and GPT-5.2 for refactoring or extending existing code—but treat both as candidates to test on your own workload. For a production API decision, verify the exact model ID and current status in the provider channel you plan to use.
What exactly is being compared?
The historical coding comparison and the API model records do not describe one interchangeable model pair:
- Historical coding comparison: GPT-5.2 versus Gemini 3.0, as reported by Kilo.
- Google API model record:
gemini-3.1-pro-preview, identified as a Preview model in the available documentation. The originalgemini-3-pro-previewendpoint was listed as shut down on March 9, 2026. - OpenAI API model record: OpenAI’s GPT-5.2 documentation describes it as a previous flagship. Whether
gpt-5.2is currently selectable for new requests through the direct OpenAI API is not confirmed here.
These are not like-for-like coding results: the Kilo tests do not establish how Gemini 3.1 Pro Preview performs against GPT-5.2. Record the exact model ID, API channel, and settings for any evaluation. (OpenAI GPT-5.2 documentation, Google Gemini API deprecation schedule)
What the coding comparison found
A Kilo article dated December 12, 2025 describes three tasks run in Kilo Code. These are the article author’s reported results, not an independent benchmark.
| Task | Reported outcome | What it suggests—and what it doesn’t |
|---|---|---|
| Python rate limiter | The prompt specified 10 requirements, including exact names and method signatures, error-message formatting, time.monotonic(), and threading.Lock(). The article says Gemini 3.0 produced a 90-line implementation that followed the specification. |
A positive result for Gemini on this particular, tightly constrained task—not evidence that it is better at coding overall. |
| TypeScript API-handler refactor | The 365-line handler had reported security, validation, and design issues. The task called for layered architecture, Zod validation, and 10 requirements. The article says Gemini missed rate limiting and database transactions, while GPT-5.2 implemented rate limiting. | The author found GPT-5.2 more complete on this refactor. The result applies to that prompt, codebase, and run. |
| Notification-system extension | Models had to explain a 400-line system and add an email handler matching its patterns. The article describes Gemini’s handler as minimal, with omissions including CC/BCC support and proper recipient handling. | A reported weakness on this extension task, not a general verdict on Gemini’s email or multimodal capabilities. |
The article says each task began in an empty Kilo Code project; it used Code Mode for the first two tasks and Ask Mode followed by Code Mode for the third. GPT-5.2 and GPT-5.2 Pro were tested at the highest available reasoning effort. The article also calls Gemini faster and cheaper and reports a 12% score gap, but does not provide the underlying score table, scoring method, or exact pairwise runtime and price figures in the available text. Treat those claims as qualitative, not as current API cost or latency estimates. Read Kilo’s three-task comparison.
Production API facts to check
The coding results above came from a coding-agent environment. They do not establish API compatibility, production availability, service limits, or operational suitability. The table separates what the available provider-document excerpts identify from what remains unconfirmed as of November 8, 2026.
| API question | OpenAI direct API: GPT-5.2 | Google Gemini Developer API: Gemini 3.1 Pro Preview |
|---|---|---|
| Model ID and status | The documentation identifies GPT-5.2 as a previous flagship. Current selection for new requests and any lifecycle or account-access restriction are not confirmed here. | The documented ID is gemini-3.1-pro-preview; the available model record identifies it as Preview. Current status and project-access conditions are not confirmed here. |
| Context and output limits | Not confirmed here. | The available model-page excerpt lists 1,048,576 input tokens and 65,536 output tokens. |
| Input and output types | Not confirmed here. | The available model-page excerpt lists text, image, video, audio, and PDF input, with text output. |
| Lifecycle qualification | Current lifecycle status is not confirmed here. | A lifecycle excerpt dated February 19, 2026 listed no shutdown date for this Preview model; that does not establish its status today. The original Gemini 3 Pro Preview endpoint was listed as shut down on March 9, 2026. |
These documented Gemini limits and modalities do not predict which model will be more accurate, faster, cheaper, or easier to operate on your workload. Confirm both providers’ current model availability, access terms, and limits for your account and deployment channel before committing. (OpenAI GPT-5.2 documentation, Gemini 3.1 Pro model page, Gemini API deprecation schedule)
Cost and speed: what can be compared
The Kilo article says Gemini 3.0 was faster and cheaper in its tests, but the available text lacks exact GPT-5.2-versus-Gemini measurements. Figures in that article for GPT-5.2 Pro or other models are not a substitute for pairwise numbers.
No current, directly comparable API prices for GPT-5.2 and Gemini 3.1 Pro Preview are established here. Check each provider’s pricing for the exact model and API channel before estimating spend; account for input, output, cached input, and any applicable tools or modality charges. Measure latency on the same prompts and input sizes with consistent settings. Track completion quality and retries as well as response time: cost per accepted result may be more useful than cost per call.
Which one should you choose?
- For a precise, bounded coding task: Gemini 3.0 is the stronger candidate in Kilo’s single rate-limiter example. Test it yourself before relying on that result.
- For a substantial refactor or extension: GPT-5.2 is the more promising of the two in that report, where the author describes it as more complete on the TypeScript refactor and notification-system task.
- For a new API deployment: Do not decide from the coding report alone. The original Gemini 3 Pro Preview endpoint was listed as shut down; Gemini 3.1 Pro Preview is a separately named Preview model. OpenAI’s GPT-5.2 documentation describes it as a previous flagship, but current API selection is not confirmed here. Verify each exact model ID, current lifecycle status, and applicable access terms before committing.
A small, reproducible bake-off
Use representative requests from your actual workload. Include exact-spec coding, changes to existing code, and document or media inputs only if your application needs them. Keep prompts, tools, output limits, and evaluation criteria consistent where the APIs allow it. Score correctness and requirement coverage against a checklist; record latency, token use, failures, and human rework. Then compare cost per accepted result, not just listed cost per token.
The published coding comparison is a reason to shortlist models for testing—not a substitute for testing them in your application.
Sources
We Tested GPT-5.2/Pro vs. Opus 4.5 vs. Gemini 3 on 3 Real-World Coding Tasks
Gemini 3 Pro preview | Gemini API
Gemini 3.1 Pro preview | Gemini API | Google AI for Developers