Hosted gpt-oss endpoint system prompts contain "You have no access to tools"

[satusky]

We are using the foundation model endpoints (provisioned and pay-per-token) for the gpt-oss models for a research project. We have been experiencing consistent tool call failures: gpt-oss-20b was failing ~1 out of every 12, while gpt-oss-120b was failing close to 1/4 of the calls.

I noticed that the system prompt length for those models were 10x the length of the other models we were comparing them to (claude-sonnet-4-5, gemma-3-12b, meta-llama-3-1-8b-instruct), so I made gpt-oss-20b leak its system prompt (on the first try) and it contains the phrase "You have no access to tools..." (see attached screenshot).

Why is that included in the system prompt? I haven't seen that in other cloud provider endpoints. We would like to reliably use those models for agentic loops, but obviously we can't do that when they fail to call tools 1/4 of the time. 

[Ashwin_DSA]

Hi @satusky,

I'm not an expert in this area, but after some internal research, I wouldn't treat a leaked system-prompt snippet as a product issue related to tool use. I can’t comment on the exact internal prompt template for a specific invocation, but the public documentation is the source of truth here. Databricks documents that Foundation Model APIs support function calls, and both OpenAI GPT OSS 20B and GPT OSS 120B are listed as supported models. It also documents the function-calling behaviour and request controls in the Function calling on Databricks and Foundation model API reference pages.

The presence of a phrase like "you have no access to tools" in a leaked prompt does not appear to indicate that GPT OSS endpoints are categorically incapable of using tools on Databricks. Publicly, these models are documented as supporting function calling.

That said, if you are seeing a materially higher tool-call failure rate with GPT OSS than with other models, that is still a valid concern. The public API docs note that when tools are provided, tool_choice defaults to auto, and if your workflow requires a tool call on that turn, setting tool_choice to required is a more deterministic option.

As a practical debugging step, I would recommend running A/B tests of the same workflow against another model that supports function calls, so you can determine whether the issue is specific to GPT OSS behaviour in your setup or something broader in the agent loop implementation. You may have already done this based on your original post but just pointing it out in case you haven't.

I would also recommend opening a Databricks support ticket with concrete repro details, such as the endpoint type, model name, sample request payloads, timestamps or request IDs, and the observed failure rate, so the team can investigate whether this is an endpoint-specific reliability issue.

If this answer resolves your question, could you mark it as “Accept as Solution”? That helps other users quickly find the correct fix.