FIELD NOTES / 4 MIN READ

Gemini 3 Can Mix Built-In Tools With Your Functions. GPT-5+ Still Can't

Gemini 3's API now supports calling a built-in tool like Google Search or code execution in the same turn as your own custom function. OpenAI's current function-calling docs explicitly block built-in tools from sharing a parallel function-call batch. If your bot's task needs grounding plus a private action in one decision step, that difference changes which model fits, and there are known failure modes on both sides worth checking before you build around either.

AI-assisted editorial: researched and drafted with AI using the primary sources linked below. Check the cited evidence and any stated test scope. Illustrative examples are not measured product results. Editorial policy.

Key takeaways

  • Gemini 3 supports combining built-in tools (Search, Maps grounding, Code Execution, URL Context) with custom function calling in a single call; the 2.5 series required separate calls for this.
  • OpenAI's current function-calling docs state built-in tools cannot be included in a parallel function-call batch, so a GPT-5+ bot still needs a second turn to mix a hosted tool with a custom function.
  • If gemini-3.1-pro-preview starts ignoring your custom tool in favor of a bash command, Google's own guidance is to switch to the gemini-3.1-pro-preview-customtools variant, not to assume the model is broken.

What Gemini 3 actually added

Google's Gemini 3 developer guide states that the model supports built-in tools alongside custom function calling in the same call. The changelog describes this as a dedicated feature: the ability to use Gemini's built-in tools alongside custom function calling tools in a single API call, and notes that combining built-in tools with function calling is now supported for Gemini 3 models specifically.

This matters because the built-in tools are hosted by Google and can't be rewritten by you: Google Search grounding, Grounding with Google Maps, File Search, Code Execution, and URL Context. Custom function calling is the mechanism for your own actions, like checking an internal database or sending a notification. Gemini 3 also allows Structured Outputs to be combined with built-in tools, including grounding, URL context, code execution, and function calling, in the same response.

OpenAI's current function-calling guide draws a firmer line. Functions can be called in parallel when built-in tools are also available starting with GPT-5, but built-in tools cannot be included in a parallel function-call batch. In practice, a GPT-5+ bot that wants to run OpenAI's web search and your own custom function in the same turn still needs to split that into separate round trips; Gemini 3 does not require the split for its own built-in tools.

Sources: Gemini 3 developer guide | Gemini API | Google AI for Developers, Release notes | Gemini API | Google AI for Developers, Function calling | OpenAI API.

An illustrative example

Illustrative only, not a BotBento test: imagine a small-team bot that answers "Are we still open for walk-ins today, and do we have the espresso machine part in stock?" The first half needs current information the model doesn't have in its weights; the second half needs a call to your own inventory function.

On Gemini 3, you can hand the model both a grounding tool and your check_inventory function in one request. The Gemini 3 developer guide confirms structured outputs can be combined with built-in tools including grounding with Google Search and function calling in the same response, so the model can decide, in one turn, to ground the hours question and call your function for the part lookup.

On GPT-5 or later, you would hand the model your check_inventory function and OpenAI's hosted web search tool, but because built-in tools cannot be included in a parallel function-call batch, the two can't resolve in the same batched response. You end up either running the built-in tool first and the custom function in a follow-up turn, or splitting the user's question into two separate requests in your own orchestration code.

Sources: Gemini 3 developer guide | Gemini API | Google AI for Developers, Function calling | OpenAI API.

Where this breaks in practice

Gemini 3's own documentation flags a specific failure mode: if you are using gemini-3.1-pro-preview and the model ignores your custom tools in favor of bash commands, the fix Google documents is switching to the gemini-3.1-pro-preview-customtools model variant, not adjusting your prompt. If your bot's tool calls silently stop firing after a Gemini model bump, check which model ID you're actually pointed at before assuming a config problem.

Gemini 3 also changed how reasoning state is carried across calls. The models use thought signatures, encrypted representations of the model's internal reasoning, to maintain context across calls. In stateful mode, the server manages this automatically; in stateless mode, where you manage conversation history yourself, you must include thought blocks with their signatures in subsequent requests to validate authenticity. If you're mixing a built-in tool call with a custom function across stateless requests, dropping that signature is a plausible reason a multi-tool turn quietly loses context partway through.

On the OpenAI side, the failure mode is different: you can set parallel_tool_calls to false to force exactly zero or one tool call per turn, which is a deliberate safety lever, not a bug, but it means a bot relying on multiple custom functions firing together in one response needs that flag left on its default. Separately, OpenAI's guide notes that for fine-tuned models, if the model calls multiple functions in one turn, strict mode is disabled for those calls, so a fine-tuned bot that also needs strict schema guarantees should test whether it is actually issuing parallel calls before trusting strict validation.

Sources: Gemini 3 developer guide | Gemini API | Google AI for Developers, Function calling | OpenAI API.

The decision

If your bot's task genuinely needs a hosted tool (search, grounding, code execution) and a private function to resolve in the same decision step, only Gemini 3's current API supports that without a second round trip, per Google's developer guide. If your bot's tasks never need both kinds of tool in the same turn, this difference doesn't affect your model choice, and you should decide on other grounds like cost, latency, or existing SDK investment.

Before switching a production bot to lean on this, pin an exact model ID rather than a dateless alias. Gemini's own docs show the family moving fast within 2026, including a documented variant swap for a tool-ignoring bug, so a dateless pointer can change behavior under you without a code change on your end. If you're already running GPT-5 or later with a built-in tool and a custom function, plan for two turns rather than one, and don't assume a future SDK release will merge them until OpenAI's own function-calling guide documents it.

Either way, test the exact combination you need with the exact model ID you plan to ship, not a general claim about the provider. A model that supports mixed tool calls in principle can still ignore your schema, drop a signature, or silently disable strict mode under specific conditions documented above.

Sources: Gemini 3 developer guide | Gemini API | Google AI for Developers, Release notes | Gemini API | Google AI for Developers, Function calling | OpenAI API.

Primary sources

Sources checked 2026-11-03. Standards and product documentation can change; follow the linked version when implementing.

  1. Gemini 3 developer guide | Gemini API | Google AI for Developers — Google AI for Developers
  2. Release notes | Gemini API | Google AI for Developers — Google AI for Developers
  3. Function calling | OpenAI API — OpenAI

BotBento is in development. Suggest a correction.