Known Issues
This page lists current limitations you may run into when using the models TrustedRails serves. Entries are removed as issues get fixed.
Protocol-level gaps (unsupported endpoints and parameters) are documented separately: see OpenAI API Compatibility and Anthropic API Compatibility.
Very short answers get padded
Section titled “Very short answers get padded”Models: all chat models; most visible on deepseek-ai/DeepSeek-V4-Flash-0731
Surfaces: OpenAI and Anthropic APIs
What happens: Completions have a minimum length of roughly 64 output tokens. When the natural answer is shorter (for example, a one-word reply), the answer arrives followed by repeated or unrelated text. Occasionally the repetition runs all the way to your max_tokens, and the response ends with finish_reason: "length" (stop_reason: "max_tokens" on the Anthropic API) even on a trivial prompt. The padding counts as output tokens and is billed. With response_format set to json_object or json_schema (OpenAI API) the content is not affected: you get clean, parseable JSON; reported token usage may still include the padding.
The minimum applies to all models. MiniMax and GLM think by default, and their reasoning usually fills the minimum before the answer starts, so they hit this far less often than DeepSeek. It can still happen when the reasoning is very short.
What you can do:
- Avoid prompts that demand one-word answers; letting the model reply with a full sentence usually clears the threshold.
- Set an explicit, moderate
max_tokensso a padded response cannot run far. - On DeepSeek, enabling thinking also takes the output past the threshold.
MiniMax may return an empty answer
Section titled “MiniMax may return an empty answer”Models: MiniMaxAI/MiniMax-M2.7
Surfaces: OpenAI and Anthropic APIs
What happens: Occasionally the response finishes normally (finish_reason: "stop", stop_reason: "end_turn" on the Anthropic API) but the message content is empty: the whole answer ends up inside the reasoning output (the reasoning field, or a thinking block on the Anthropic API).
What you can do:
- Retry the request.
- An empty content with
finish_reason: "stop"and a populated reasoning output is this issue, not a token-budget problem: budget exhaustion ends withfinish_reason: "length"instead.
Forced tool calls are not strictly enforced
Section titled “Forced tool calls are not strictly enforced”Models: all chat models, varying by model
Surfaces: OpenAI API (tool_choice: "required") and Anthropic API (tool_choice: {"type": "any"})
What happens: Asking for “any tool” does not guarantee a tool call: the model may still reply with plain text. Forcing a specific named tool works reliably.
What you can do:
- Check the response for
tool_calls(or atool_useblock) and retry if the call is missing. - Force the specific tool by name when you know which one you need.
Temperature 0 is not fully deterministic
Section titled “Temperature 0 is not fully deterministic”Models: all
Surfaces: OpenAI and Anthropic APIs
What happens: temperature: 0 makes outputs more stable, as expected, but does not guarantee exact reproducibility: identical requests can still return different outputs. The seed parameter is accepted for compatibility but does not make outputs reproducible.
What you can do:
- Do not build logic that depends on exact output reproducibility.
GLM thinking can consume small token budgets
Section titled “GLM thinking can consume small token budgets”Models: zai-org/GLM-5.3-Flash
Surfaces: OpenAI and Anthropic APIs
What happens: GLM thinks by default and its reasoning is often long. With a small max_tokens the entire budget can go into thinking: you get an empty answer with finish_reason: "length" (stop_reason: "max_tokens" on the Anthropic API), and the thinking is billed as output tokens. With json_schema structured output (OpenAI API), reasoning tends to grow further and responses can take noticeably longer than usual.
What you can do:
- Give GLM a generous
max_tokenswhen thinking is on. - Disable thinking when you need short, fast answers:
reasoning_effort: "none"on the OpenAI API,thinking: {"type": "disabled"}on the Anthropic API.