Why doesn't GPT ever ask me anything?

How this got made: I noticed a pattern, asked Claude to research it, we went back and forth, I wrote this up. Just so u know.

🚀 TL;DR
Elicitation works, but only when the model knows when to shut up and when to ask.
OpenAI bet "just act," Anthropic bet "ask first," and both are now converging
to the same middle: act by default, ask only when it actually earns its keep.

The real feature isn't elicitation. It's calibration.

Table of contents

🤔 The thing I noticed

I use both Claude Code and Codex. A lot. And at some point I realized something:

Codex never asks me anything.

Not on simple tasks. Not on complex ones. Not even when the spec is clearly ambiguous and any reasonable human would go “hey, did u mean X or Y?” Nope. It just picks one and ships it.

Claude Code, on the other hand, used to stop me every five minutes. “Did you mean this?” “Which approach do you prefer?” “Should I also update the tests?” Sometimes helpful. Sometimes I wanted to throw my keyboard.

So I started digging. Is this a bug? A feature? A product philosophy? Turns out it’s all three, depending on who u ask.

🔍 What is elicitation anyway?

Quick primer. Elicitation is when the model pauses mid-turn and asks u a structured clarifying question (with selectable options, not just freeform text) before continuing.

In Claude Code, this is the AskUserQuestion tool. It pops up with radio buttons, u pick one, the model continues with ur answer baked in. It’s part of the MCP spec too, so technically any model can use it.

The keyword is technically. Because GPT models don’t. At all.

🧠 Why OpenAI said no

This isn’t a technical limitation. OpenAI’s official prompting guides instruct developers to tell their models not to ask.

From the GPT-5.2 prompting guide:

“Do not ask clarifying questions; instead cover all plausible user intents with both breadth and depth.”

From the Codex prompting guide:

“Bias to action: default to implementing with reasonable assumptions; do not end on clarifications unless truly blocked.”

This isn’t just one team’s opinion either. OpenAI’s own GPT-5 prompting guide documents how Cursor found that GPT-5 “occasionally deferred to the user for clarification or next steps before taking action, which created unnecessary friction.” Their fix was prompt engineering to suppress it.

During Codex alpha testing, OpenAI reported that “people building with Codex were happier as the agent was better understanding their intent and made more progress per turn, with fewer clarifying questions.” The design goal for Codex is “assign, supervise, review.” Like handing a task to a colleague who delivers results, not one who peppers u with questions.

Cade Russell, one of our PMs at Casper, nailed it: “I used to call Codex T-1000 for this exact reason.” Zero hesitation, zero questions, just execution.

Fair enough. But does it actually work better?

📊 The research: someone actually measured this

Turns out, yes, people have run the numbers.

“Ask or Assume?” (2026), the most direct eval. They ran coding agents on an underspecified variant of SWE-bench:

• Ask-then-act: 69.4% resolve rate
• Assume-and-execute: 61.2%

That’s an 8 percentage point gap. On hard, ambiguous tasks, asking first wins.

But here’s the nuance. AskBench (2026) found the exact opposite for clear tasks: unnecessary questions actively hurt performance. Zero accuracy gain, just overhead.

So the answer isn’t “always ask” or “never ask.” It’s: know which regime ur in.

Sounds easy. It’s not. CLAMBER (2024) showed that current LLMs are bad at assessing their own knowledge boundaries. They don’t reliably know when they should ask.

Which means if u can’t calibrate well, “never ask” might genuinely be the safer default.

That’s OpenAI’s bet in a nutshell.

🗳️ Tiny Casper poll interlude

So I threw a Slack poll at Casper. 37 people in the channel, 7 actually voted.

“elicitation: model asking clarifying questions. ur reaction?”

Six people also hit 🚀. Not counting those. Emoji votes are not real votes.

n=7, very unscientific, but the vibe was clear: ask, but earn the interruption.

📉 Plot twist: Claude is asking less too

Here’s the part that surprised me. Recent Claude models are also trending toward less asking.

This is intentional. From Claude’s constitution:

“Checking in or asking clarifying questions more than necessary for simple agentic tasks makes Claude more annoying and less useful, and reflects poorly on Anthropic.”

Anthropic’s autonomy research backs this up: Claude Code asks for clarification more as task complexity rises, and on the hardest tasks it asks more than twice as often as on simple ones. But the overall trend across model versions is clear. Less asking, more acting.

Anthropic’s early access partners described Opus 4.6 as having “a propensity to work autonomously without hand-holding.” If u’ve used it recently, u’ve felt this. It just… does things now.

🔀 The convergence

So we’ve got two companies that started on opposite ends:

They’re meeting in the middle. Both independently arrived at the same UX insight: users want agents that act, not agents that interview.

And ironically, both have their own failure modes:

• Codex over-assumes on scope. Does too much without checking.
• Claude over-assumes on intent. Interprets questions as criticism and starts editing code.

HN users literally prefix Claude messages with “THIS IS JUST A QUESTION. DO NOT EDIT CODE.” lmao. Both sides have work to do.

🧩 The deeper pattern

This isn’t just about elicitation. It’s the same architectural tradeoff that shows up everywhere in the agent stack:

• MCP vs direct file access. Protocol overhead vs tight coupling.
• Human-in-the-loop vs full autonomy. Safety vs speed.
• Elicitation vs assume-and-act. Correctness vs throughput.

The pattern is always: intermediation trades throughput for correctness.

Add a layer. A protocol, a question, a checkpoint. More accuracy on hard tasks. More friction on easy ones.

The crossover point is always the same: task complexity and specification quality.

Well-specified simple tasks → just do the thing, any intermediation is waste. Underspecified complex tasks → pause and check, the cost of redoing is way higher than the cost of asking.

The real feature isn’t “can the model ask questions.” It’s “does the model know when to.”

💡 My 2 cents

Both bets are defensible. Neither is right yet. The winner is whoever nails calibration first.

As a user who lives in both tools daily, I’m just glad they’re converging. Because getting interrogated on obvious tasks is annoying, and getting surprised by wrong assumptions on complex ones is worse.

📖 Papers & sources

• Ask or Assume? Uncertainty-Aware Clarification-Seeking in Coding Agents (2026)
• AskBench: When and What to Ask (2026)
• CLAMBER: Clarifying Ambiguous Information Needs (2024)
• Measuring AI Agent Autonomy, Anthropic (2026)
• Introducing GPT-5.3-Codex, OpenAI (2026)
• Claude Opus 4.6 Announcement, Anthropic
• MCP Elicitation Specification
• Claude’s Constitution
• GPT-5 Prompting Guide, incl. Cursor case study
• GPT-5.2 Prompting Guide
• Codex Prompting Guide

Want to chat? LinkedIn, X DMs, or [email protected].