If you've been asking ChatGPT questions by talking to your phone — "What's a good restaurant near me," "Remind me what this medication does," "Help me think through this decision" — there's a decent chance the AI answering you stopped keeping up with the world in April 2024.

Meanwhile, if you'd typed that exact question instead of saying it out loud, you'd have gotten an answer from a model that's current through May 2026.

Same app. Same account. Different AI underneath.

The two versions of ChatGPT you didn't know you were switching between

Here's an object lesson, because the technical version is harder to parse than the situation deserves.

Imagine two customer service representatives who work at the same company. Same uniform, same name tag, same phone number to reach them. But one went through the full training program last month. The other completed training in early 2024 and hasn't been updated since. Their answers will usually overlap. On anything that changed in the last year or so, they'll diverge. And the company's phone system doesn't tell you which one picks up.

That's roughly what's happening with ChatGPT's voice and text modes right now.

ChatGPT's Advanced Voice Mode — the feature that lets you have a natural spoken back-and-forth with ChatGPT — runs on a model derived from the GPT-4o era. Developer Simon Willison documented this on April 10, 2026: the voice mode reports an April 2024 knowledge cutoff. The text interface runs on GPT-5.5 Instant, current through May 2026. That's a gap of about 13 months.

Why this is happening and why it matters

Real-time voice requires fast responses. The AI has to hear you, process what you said, and start replying quickly enough that the conversation doesn't feel awkward — ideally under a second. Current frontier models like GPT-5.5 are too slow for that at the scale ChatGPT operates, at least right now. So OpenAI built Advanced Voice Mode on an older, faster model.

That tradeoff is defensible. What's harder to defend is the lack of disclosure.

If you ask the voice mode about something that changed between April 2024 and now — a product recall, a change in medical guidelines, a law that went into effect, a company's recent policy shift — you might get an answer that's confidently wrong. Not because the AI "hallucinated" (made something up), but because that thing simply didn't exist in its training data. It'll answer from what it knows. And what it knows stopped 13 months ago.

For asking about movies or recipes or how to word an email, none of this matters much. For anything where the last year is relevant — medical questions, legal questions, anything you'd actually look up — it matters quite a bit.

OpenAI almost made this worse

Here's where it gets more interesting.

ChatGPT actually has two voice modes. Advanced Voice Mode — the speech-to-speech model described above, with the April 2024 cutoff. And Standard Voice Mode — an older pipeline that works in three steps: it converts your speech to text, calls the same current GPT model as the text interface, then reads the answer back to you. Standard Voice Mode uses the current model. No 13-month gap.

OpenAI announced plans to retire Standard Voice Mode and move all users to Advanced Voice Mode. Users pushed back loudly in the developer forums. Nick Turley, OpenAI's VP and Head of the ChatGPT app, announced they'd keep Standard Voice Mode after hearing the feedback.

That reversal was good. But the fact that the plan was to remove the more accurate mode in the first place is worth noting. Advanced Voice Mode is more natural-feeling in conversation — it hears your tone, responds with more fluency, feels less like talking to a machine. Standard Voice Mode sounds more robotic. OpenAI optimized for how the experience feels, at the cost of what it knows.

Source spread

• Simon Willison — ChatGPT voice mode is a weaker model [skeptic] — documented the April 2024 knowledge cutoff with direct testing; primary source for the capability gap
• TechRadar — OpenAI brings back Standard Voice Mode after user pushback [skeptic] — covers the reversal and Nick Turley's statement
• OpenAI Voice Mode FAQ [hype] — OpenAI's official framing; notably thin on disclosing model generation or knowledge cutoff
• OpenAI Developer Community — Please don't retire standard voice [skeptic] — the user thread that changed the outcome

What's real and what deserves a side-eye

What's real:

• Advanced Voice Mode is genuinely better at the conversational experience. Faster, more natural, hears tone and pacing, feels like you're talking to someone instead of waiting for output. That's a real improvement for casual use.
• Standard Voice Mode still exists and uses the current model. If you want the most up-to-date answers in voice form, you can get them.
• The knowledge gap is a real engineering constraint, not negligence. Running a frontier model in real-time speech at ChatGPT scale is genuinely hard and expensive.

What deserves a side-eye:

• OpenAI doesn't tell you which model you're using when you switch to voice mode. Not a label, not a pop-up, not a note in the flow. You have to go find a developer's test post to learn this.
• The original plan to retire Standard Voice Mode — the more accurate one — in favor of Advanced Voice Mode was the wrong call. It took community pressure to reverse it.
• "This is a technical constraint" is not the same as "therefore no disclosure is needed." These are two separate decisions.

What to do about it

• For general conversation and creative questions, Advanced Voice Mode is fine. The AI behind it is still very capable for anything that didn't change in the last 13 months, which is most of the world.
• For anything time-sensitive, use Standard Voice Mode or just type the question. Medical questions, legal questions, product comparisons, recent news, anything where 2025 or early 2026 might be relevant — either switch modes or type it instead.
• Check which mode you're in. Standard Voice Mode is in the voice settings within ChatGPT. The default in most recent app updates is Advanced Voice Mode. It takes 30 seconds to check and switch.
• You can ask voice mode to search the web. Say "Can you look this up?" or "Search for the current version of that." ChatGPT's voice mode can browse the web — it just doesn't do it automatically. Asking explicitly bypasses the knowledge cutoff problem.
• Treat voice answers the way you'd treat a smart colleague who was on sabbatical for 13 months. Great for most things. Out of date on anything recent. Cross-check anything that matters.

Further reading

• Simon Willison — ChatGPT voice mode is a weaker model — developer's documented testing of the knowledge cutoff gap; the original write-up
• TechRadar — OpenAI brings back Standard Voice Mode — the reversal story and Nick Turley's statement
• OpenAI Voice Mode FAQ — official product documentation; read it and notice what's not there
• OpenAI Developer Community — Please don't retire standard voice — the thread that changed the outcome