ChatGPT, Claude or Gemini for IELTS Speaking: What a Purpose-Built Examiner App Does Better

You can practise IELTS Speaking with ChatGPT, Claude or Gemini, and for ideas, explanations and unlimited talking time they are genuinely excellent. What they are not built for is assessment. None of their makers’ own pages describe marking speech against the IELTS band descriptors, grading pronunciation against an exam criterion, or keeping a running count of the mistakes you repeat week after week. For that specific job — scored, timed, tracked speaking practice — an examiner-built tool is doing something a general assistant was never designed to do.

We make IELTS Mentor AI, one of the apps discussed here. We’ve tried to be fair; every statement about ChatGPT, Claude and Gemini below comes from their makers’ own pages, checked on 18 September 2026.

What ChatGPT, Claude and Gemini are actually built for

OpenAI’s help page for ChatGPT Voice says it “lets you speak with ChatGPT and hear a spoken response,” and that the Live option “is designed for natural back-and-forth conversation.” Anthropic’s help page says voice mode “allows you to have complete spoken conversations with Claude.” Google’s page for Gemini Live describes “a feature that lets you talk with AI from Google using just your voice.”

Three companies, one description: a spoken interface to a general assistant. None of those pages mentions IELTS, band scores or pronunciation assessment — not because anything is hidden, but because that is not what these products are. Anthropic’s own headline for Claude is “The AI for Problem Solvers.” Judged as general assistants, all three are remarkable. Judged as IELTS examiners, they are not making the claim at all. Which voice options you get depends on your plan, region and app version, so check the maker’s page.

Four things a general assistant does genuinely well for IELTS

Most comparison articles skip this part, and it is what makes the rest worth trusting. Several jobs in Speaking preparation genuinely suit a general assistant better.

Generating practice questions and ideas. The bottleneck for candidates practising alone is often not feedback but material — running out of questions, or of things to say about them.

Give me 10 IELTS Speaking Part 3 questions about technology and society, at the difficulty an examiner would actually use, then three possible angles I could argue for each.

Explaining a grammar point until it makes sense. If you don’t see why “I have been to London last year” is wrong, an assistant will explain it five different ways, in your own language, without getting impatient.

Here is my Part 2 answer: [paste your transcript]. For each grammar mistake, explain in my own language why it is wrong, then show the corrected sentence next to my original one.

Rephrasing an answer at a higher level. Seeing your own idea expressed the way a Band 8 candidate would, with every change named, shows the gap between what you meant and what you said.

Rewrite this answer at Band 8 level, keeping my ideas and my story exactly the same. Then list every word and structure you changed, and say what each change is doing.

Being an unlimited conversation partner. It will talk to you at midnight, switch to your first language to explain something, and switch back — at whatever length you want.

Notice what all three prompts have in common: none asks for a band score.

What none of the makers claim: a score anchored to the descriptors

IELTS Speaking is scored by a certificated examiner against four equally weighted criteria — Fluency and Coherence, Lexical Resource, Grammatical Range and Accuracy, and Pronunciation — and each has published descriptors, band by band, describing what an examiner is listening for.

Ask any of the three for a band score and you will get one. It will be well written, confident, and often not unreasonable. But nothing on any maker’s page describes their model as calibrated to those descriptors, checked against examiner marking, or evaluated on IELTS at all. The number is produced the way any other plausible sentence is produced, not measured against a rubric.

You shouldn’t take that on faith. Read the public band descriptors yourself, then hold any score — from an assistant, from an app, from anywhere — against the descriptor wording for the band it claims.

Conversation is not an exam: format, timing and pronunciation

The exam’s shape doesn’t exist in a chat. IELTS Speaking is Part 1, then a cue card with one minute of preparation and up to two minutes of talking, then Part 3 — a fixed sequence under fixed timing. An assistant will run something like it if you prompt it to, then drift within a few turns unless you keep steering. You end up invigilating yourself, which removes the pressure the format exists to create. A full timed mock exists because sitting the shape of the exam is a separate skill from answering its questions.

Pronunciation is a scored criterion, not a transcription problem. OpenAI’s page suggests setting the language you speak most often because it “can help ChatGPT understand your speech more accurately” — understanding you, not assessing you. That is the right goal for an assistant and the wrong one for an examiner. IELTS marks pronunciation on stress, rhythm, intonation and listener effort, and no maker describes their voice features as assessing any of it.

Memory remembers you; it is not an error log

This is the point most easily got wrong. All three assistants do have memory, and it is real.

OpenAI says that when Memory is enabled, ChatGPT “can remember relevant preferences and details from your chats,” while also stating that it “does not retain every detail from every conversation” and that “ChatGPT decides which available information is relevant to a response.” Anthropic says Claude “saves memory as a set of individual topics as you chat, rather than summarizing conversations after they end.” Google describes personalisation in which Gemini can “learn from your chats to understand more about you and your world” — and notes that memory of past chats is “not available in certain features, like Gems or Live chats,” which is precisely where a spoken practice session would happen.

All of that is memory of you: your preferences, your context, the topics you return to. What none of it is described as being is a cumulative record of one learner’s recurring errors — the kind that says you dropped the third-person “-s” in nine of your last twelve answers, that this is holding Grammatical Range and Accuracy down, and that it is what today’s practice should target. Remembering that you are preparing for IELTS is not the same as counting what keeps going wrong — and an app that explains your mistakes is built around the counting.

A fair test you can run in about ten minutes

Don’t decide this from an article. Transcribe one of your own recorded answers and run three checks.

Consistency. Paste the same transcript into two brand-new chats and ask each for a band score with a criterion breakdown. An anchored measurement should land in the same place twice; if the numbers differ, you have learned how much weight to give either.

Justification. Ask it to quote the exact band-descriptor wording behind each criterion’s number, then check that wording against the published descriptors. A score that can’t point at the descriptor language behind it is a guess wearing a uniform.

The record. A week and several sessions later, ask which errors you have repeated most often, and how many times — then check that against what you actually said. This is why automated scoring is worth judging on its consistency over time rather than on any single confident answer.

The comparison, row by row

Checked 18 September 2026 against the makers’ own pages: OpenAI’s ChatGPT Voice and Memory help articles; Anthropic’s voice-mode and memory help articles and its Claude overview; Google’s Gemini Live overview and its help page on memory of past chats. Voice availability varies by plan, region and app version.

The honest division of labour

The mistake isn’t using ChatGPT, Claude or Gemini for IELTS. It is asking one tool to do two jobs that have almost nothing in common.

Use a general assistant for the open-ended half: unlimited questions, ideas you wouldn’t have thought of, grammar explained until it clicks, your answer rewritten a level up with the changes named, in your own language whenever that is faster. No examiner app will out-argue it.

Use an examiner-style tool for the measured half: an answer scored on all four criteria, a band range instead of one falsely precise number, the exam’s own timing across Parts 1 to 3, and a record of which mistakes keep reappearing, so today’s practice targets the ones actually costing you. That is the narrow job IELTS Mentor AI is built for, narrow on purpose. The same split works if you are preparing entirely on your own: the assistant is your study partner, the scored practice your reality check.

Both halves matter. Only one of them is being claimed by the companies whose products you would be trusting with it.