Do you have any experience with ChatGPT 5.2 instant deciding to ignore tool call results?

[finite wraith]

Anyone have any experience with ChatGPT 5.2 instant deciding to ignore tool call results? I've found a few threads on this topic, but no one there got any responses: https://community.openai.com/t/apps-sdk-it-is-showing-the-ui-for-a-tool-call-but-not-actually-calling-the-tool/1365018
OpenAI Developer Community
Apps SDK - it is showing the UI for a tool call, but not actually c...
I’m seeing an issue where ChatGPT (via Apps SDK) will look like it’s calling a specific tool, but not actually call it. It’s reading the tool _meta to get the invoking / invoked text as well as the outputTemplate, but not anything else. This is also a read only tool and it doesn’t ask for my permission to call it at all. I’m seeing ze...

🔥
Click to react
❤️
Click to react
👍
Click to react
Add Reaction
Edit
Forward
More
[15:25]Friday, January 9, 2026 at 15:25!Artifact
OAI
: Is this just a skill issue on my part with the tool results? I get a very clear UI popup confirming tool use, but the model's response might as well just be ignoring my tool call results. (edited)Friday, January 9, 2026 at 15:26
🔥
Click to react
❤️
Click to react
👍
Click to react
Add Reaction
Edit
Forward
More
[15:26]Friday, January 9, 2026 at 15:26!Artifact
OAI
: I even went so far as to return context that asked it to talk like a pirate to no avail...

[worn niche]

Hi @finite wraith. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

---

There’s no direct mention in the docs of a bug specific to “GPT‑5.2 Instant ignoring tool call results,” but there are a few relevant pieces of information and known issues that might explain what you’re seeing.

1. Tool UI vs. actual tool execution (Apps SDK / MCP issues)
   Several recent forum threads describe situations where:

   • The UI for a tool or widget appears in ChatGPT Apps,
   • Metadata like openai/toolInvocation/invoking, openai/toolInvocation/invoked, and openai/outputTemplate is clearly being read,
   • But the actual behavior (tool execution or widget rendering) is broken or inconsistent. [Ui not rendering; Metadata regression]

   In one case, a regression caused ChatGPT to “reject or lose the message metadata that MCP tools rely on,” which made tools/widgets appear but not behave correctly, and this had to be fixed on OpenAI’s side. [Metadata regression]

   There are also multiple reports of:

   • “Error in message stream” when invoking custom connectors or widgets, often after a platform change, with no code changes on the user side. [Error in message stream]
   • Confirmation UI for actions disappearing or being flaky, blocking interaction with widgets or tools. [Confirm button disappeared]

   These show that, at least historically, there have been intermittent platform‑side issues where the UI suggests a tool is being used, but the underlying behavior is buggy or incomplete.

2. Model behavior vs. platform plumbing
   The GPT‑5.2 docs and troubleshooting guides focus on how to get the model to use tools correctly (overthinking, underthinking, malformed tool calls, calling too many tools, etc.), but they don’t describe a specific, expected behavior where GPT‑5.2 Instant intentionally ignores tool outputs after a successful call. [GPT‑5 troubleshooting; Custom tools]

   When the model misuses tools, the guidance is usually:

   • Tighten prompts and tool descriptions,
   • Use policies like “select one tool or none; cap tool calls,”
   • Or debug malformed calls. [Calling too many tools; Malformed tool calling]

   None of that explicitly covers “tool call succeeds, but the model’s final message ignores the result,” so if your tool output is clearly correct and visible in Developer Mode / inspector, this starts to look more like a platform or routing quirk than intended model behavior.

3. What you can do to distinguish “skill issue” vs. platform issue
   Based on the docs and forum patterns, a few concrete checks:

   • Verify the tool output is actually present in the message stream.
     In ChatGPT Developer Mode, expand the tool call and confirm:
     • The tool input is what you expect, and
     • The tool output is present and well‑formed JSON or text. [Developer mode]
       If the output isn’t there, the problem is upstream (tool not really being called, or failing silently).

• Check for metadata / widget regressions.
  There was a known regression where ChatGPT “rejected or lost” MCP metadata, breaking widget behavior until engineering pushed a fix. [Metadata regression]
  If your experience started suddenly without code changes (like the “Error in message stream” and confirmation‑button threads), it may be a similar platform‑side issue rather than your prompt or tool design. [Error in message stream; Confirm button disappeared]

  • Prompt the model to explicitly ground its answer in the tool output.
    The GPT‑5 troubleshooting guide suggests using meta‑prompting and explicit policies when tool use is flaky. For example, you can add instructions like:

    “After any tool call, you must base your final answer solely on the tool’s result. Quote or summarize the tool output explicitly.” [GPT‑5 troubleshooting]
    If, even with that, GPT‑5.2 Instant still ignores the tool output that you can see in the inspector, that’s strong evidence of a bug rather than a “skill issue.”

4. Known instability around Apps SDK / MCP
   Multiple forum posts show that Apps SDK + MCP + widgets have had periods where:
   • Tools in development apps fail intermittently,
   • Widgets don’t render or get stuck behind broken confirmation flows,
   • Behavior changes after platform updates without user code changes. [Error in message stream; Confirm button disappeared; Ui not rendering]

That pattern matches your description of “very clear UI popup confirming tool use” but behavior that doesn’t line up with the tool’s result.

---

Bottom line:

• The official docs do not describe an intended behavior where GPT‑5.2 Instant routinely ignores valid tool results.
• There are documented regressions and bugs around Apps SDK / MCP metadata, widgets, and tool flows that can make it look like tools are being used while the underlying behavior is broken. [Metadata regression; Ui not rendering]
• If you can see correct tool outputs in Developer Mode but the model’s reply ignores them even with explicit instructions to use them, it’s likely closer to those platform issues than to a simple prompting “skill issue.”

If you can reproduce this reliably, the best next step (per the forum patterns) is to file a support ticket or post a minimal repro in the Developer Community, including the raw tool input/output JSON from Developer Mode so engineering can check for another regression.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

@finite wraith Kapa.ai is still learning and improving, please let me know how I did by reacting below.

[finite wraith]

@worn niche tools that generate UI work, but tools that return text don't affect the model's behavior.

@worn niche ^^^ se above

[worn niche]

Let me check the knowledge base for you again. One moment! :books:

---

There’s no explicit documentation in the provided sources about GPT‑5.2 Instant specifically “ignoring” tool call results, but there are closely related issues and behaviors around tools, Apps SDK, and MCP that match parts of what you’re seeing.

1. Similar platform issues (Apps SDK / MCP)

Recent forum posts show that Apps SDK + MCP can get into states where:

• The UI clearly indicates a tool is being used (invoking/invoked text, widget frame, confirmation overlay),
• But the actual tool behavior or result handling is broken or incomplete.

Examples:

• A regression where ChatGPT “rejected or lost the message metadata that MCP tools rely on”, causing tools/widgets to appear but not function correctly until engineering pushed a fix. [ChatGPT not reading metadata]
• Apps SDK dev apps showing “Error in message stream” when invoking custom connectors or widgets, often after a platform change, with no code changes on the user side. [Error in message stream]
• The confirm action button disappearing or being flaky, leaving users stuck at a confirmation overlay and unable to proceed with the tool/widget, even though the UI suggests a tool is in play. [Confirm button disappeared]

These show that the platform has had intermittent bugs where the plumbing around tools and widgets misbehaves, even though the UI suggests everything is fine.

2. Tool calls vs. model behavior

The official docs on tools and GPT‑5 troubleshooting talk about:

• Malformed tool calls,
• Calling too many tools,
• Prompt contradictions causing bad tool behavior. [Malformed tool calling; Calling too many tools]

But they do not describe an intended behavior where the model routinely ignores valid tool outputs once they’re in context.

Instead, they recommend:

• Tight, unambiguous tool descriptions,
• Explicit policies like “select one tool or none; prefer answering from context,”
• Meta‑prompting to debug malformed calls. [GPT‑5 troubleshooting]

So if your tool output is clearly present and correct, and the model’s final message still behaves as if nothing happened (e.g., ignores “talk like a pirate” returned from the tool), that’s not something the docs frame as expected behavior.

3. How to tell “skill issue” vs. platform issue (with what we know)

Given the docs we have:

1. Check the tool output in Developer Mode
   • Expand the tool call and confirm that:
     • The tool was actually invoked, and
     • The output is present and well‑formed. [Developer mode]
   • If the output is missing or malformed, the problem is with the tool call or MCP plumbing, not the model “ignoring” it.

2. Watch for metadata / widget regressions

   • There is a documented case where metadata handling broke and was later fixed. [ChatGPT not reading metadata]
   • Combined with the “Error in message stream” and “confirm button disappeared” threads, this suggests that sudden behavior changes without code changes on your side can be platform regressions, not your prompt design. [Error in message stream; Confirm button disappeared]

3. Force grounding in tool output via prompt

   • The GPT‑5 troubleshooting guide suggests using explicit instructions when tool use is flaky, e.g.:

     “After any tool call, base your final answer solely on the tool’s result; quote or summarize the tool output explicitly.” [GPT‑5 troubleshooting]

   • If, even then, GPT‑5.2 Instant ignores a tool result that you can see in the inspector, that strongly points to a bug / regression rather than a “skill issue.”

4. Direct answer to your question

• The knowledge sources do not explicitly say “GPT‑5.2 Instant sometimes ignores tool call results.”
• They do show multiple recent, platform‑side issues where:
  • Tools and widgets appear to be used,
  • But metadata, execution, or result handling is broken in Apps SDK / MCP flows. [ChatGPT not reading metadata; Ui not rendering; Error in message stream; Confirm button disappeared]

Given that, your experience is plausibly aligned with known regressions/instability around Apps SDK + MCP, rather than something clearly documented as “you’re just using tools wrong.”

If you can capture:

• The raw tool input/output JSON from Developer Mode, and
• A minimal repro (prompt + tool definition),

posting that in the Developer Community or a support ticket is the best way—per the existing threads—to get engineering to confirm whether this is another regression.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon: