Skip to main content

Message is Empty when reasoning is long

Bug Reports

When using T3 Chat with the reasoning model enabled, assistant replies sometimes show only the hidden “thinking” section and the UI displays “Message is empty” instead of the actual assistant message. This seems to happen when the internal reasoning payload is very long (e.g., for complex SQL analysis). If the assistant is asked to “keep thinking to a minimum” and answer briefly, messages appear normally again. Noticeable on models such as Model: GPT‑5.1 (Reasoning). Roughly what’s probably happening:

  1. The model sends back:

    • A big hidden “reasoning” section.

    • A shorter “final message” that should be shown to you.

  2. When the reasoning section is very long, the T3 Chat front‑end or API:

    • Either hits a size/serialization limit, or

    • Mis-parses the response so that the visible message field ends up null/empty.

  3. The UI then displays “Message is empty”, even though the model did send an answer.

Status: In Progress23 comments

This request was merged into another request

Comments23

  • Mark Florkowski

    Team

    This appears to be a case of hitting the output token limit while reasoning. Will look into rendering a better state for this.

    • Copper Supernova

      Thanks for the update. One idea for improving the UX: Could T3 Chat enforce a hard token budget for the final, user‑visible answer?

      For example, reserve a minimum number of X tokens for the answer itself.

      • Copper Supernova

        Just to follow up on my earlier comment:

        Would it also make sense to put an explicit cap on the reasoning budget itself? For example, given a total output limit, you could let the hidden reasoning use at most limit - min_answer_tokens, so there are always at least min_answer_tokens left for the user‑visible answer.

        That way the model can’t burn the entire quota on internal reasoning, and in the worst case the user just gets a shorter, more high‑level answer instead of an empty message.
        Curious if that fits with how your reasoning vs. visible streams are wired up, or if there are constraints that make it tricky?

        • Mark Florkowski

          Team

          It would make a lot of sense for the API to be designed this way, but unfortunately most if not all are not

  • Aqua Chameleon

    For GPT-5.1-Thinking specifically, maybe set reasoning effort lower? Does it still do that on lower effort settings?

    I’ve only seen Qwen3-235b-a22b-thinking do this often (or at least, I assume that’s what happening when the message is empty after a loooong CoT). When it runs out of tokens mid-thought, I prompt it to continue. Whenever I want Qwen3-235b to think as much as it wants on a question, I am prepared to spend several credits on the answer XD (I haven’t used it in a while.)

    It’s annoying that even if you want to limit tokens spent on thinking in the settings, you’d have to set a different limit per model, since they have different total token maximums, right? -_- …assuming the model has such a setting; I was surprised to learn that not all “reasoning” models have one!

  • Teal Ox

    I am getting this often with Gemini 3 Pro High Reasoning. The prompt and reasoning didn’t seem that big but maybe there was a lot of web search results?

  • Copper Supernova

  • Red Pudding

    I think I’ve gotten like 4 or more of these over the last two days on queries using via Gemini 3 Pro with high reasoning and search enabled.

  • Brown Mole

    Is there any progress on this? After the new message-counting system, this happens frequently with random models. One of the examples below.

    https://t3.chat/share/984t4fvbyd

  • Salmon Jaguar

    As of today (2026-02-18) I am getting this on almost every invocation of Sonnet 4.6, Sonnet 4.6 (Thinking), and Opus 4.6, regardless of the complexity of the request, AFAICT.

    • Mark Florkowski

      Team

      Can you provide the urls of the chats where you encoutnered this so that I can checl logs?

      • Salmon Jaguar

        So this is the one that made me hit up the forum:
        https://t3.chat/chat/2f3362df-074e-4290-b35e-47e4fca312b4
        BUT: I did reruns, so I don’t know if you can see half the empty AI messages. I assume you can see the chat w/o me giving perms?

        • Mark Florkowski

          Team

          I can’t see the chat itself, but I can see logs. Will look into this

  • Copper Supernova

    You can check the logs for this chat URL where the bug occurred:

    https://t3.chat/chat/ffa36b2b-f08f-47b6-b160-483c62f54ca2

  • Brown Meteorite

    getting same issue on Minimax M2 and GPT 5 mini high

  • Scarlet Highlighter

    ran into this with DeepSeek v3.2 (Thinking) search enabled, it first reasoned then searched the web then reasoned then stopped and didn’t write the message.