Can ChatGPT, Claude or Gemini mark your GAMSAT essay?
Yes, it will give you a number. Which number depends on the chatbot. We sent our three published sample essays to 8 chatbot models, five times each: our strongest sample drew anything from 75 to 90, and our Judges marked it 74. One model’s five tries on our middle sample spread 17 points. Our two AI Judges give you a mark on our 40–84 scale with the reasons, sentence by sentence, and your first one is free.
Free · no card · marked within minutes by two AI Judges
A chat window gives you
- One number, on whatever scale you ask for
- A different number from a different chatbot
- One reading, from one reader
Our Judges give you
- A mark on our 40–84 scale
- Two Judges on every essay, and a third when they disagree widely and time allows
- The reasons, sentence by sentence
One essay, 8 chatbot models: 75–90. Our Judges: 74.
Each dot is one chatbot reply (8 models × 5 tries), on all three essays. The violet line is our Judges’ published mark; the shaded band is the range an earlier release of our Judges gave when it marked the essay again, five times.
Our Judges: 74
Chatbots: 75–90
Our Judges: 63
Chatbots: 63–84
Our Judges: 54
Chatbots: 42–62
Chatbots: 11 October 2026, one plain prompt through each model’s API, scored out of 100 as asked. Our Judges: published marks from an earlier release, re-marked August 2026. Small sample; our marks are ours, not ACER’s. Every score and the method
What you get back
A real report, from a real mark
Sample essay written for demonstration; AI-marked — every score and comment below is the grading pipeline's own output, marked by two AI Judges.
74
scaled score
Line-level annotation
“What carried the work across a dozen generations was not a relay of unusually saintly individuals but an institution”
A highly sophisticated historical insight. Shifting the focus from individual morality to institutional continuity is the core strength of this essay's thesis.
And what to fix next: Address the epistemological challenge of 'pruning': While the call to reform or 'prune' inherited institutions is compelling, the essay could acknowledge the paradox of how we ensure our current 'pruning' does not itself become a flawed inheritance for the next generation.
Three sample essays, three score bands — same prompt, fully annotated.
Read the full marked essayThe test
Three essays, 8 chatbot models, five runs each
On 11 October 2026 we sent our three published sample essays to 8 AI models, five times each, the way a student might: one plain prompt asking for a score out of 100, with no marking guide. We used each model directly, which isn’t quite the chat apps; they add their own settings. Our Judges’ marks are the ones published on each sample, from an earlier release of our Judges. We have not re-marked them on today’s release.
Our strongest sample: our Judges marked it 74
Their five re-marks, Aug 2026: 72–74
ChatGPTGPT-5.6 Luna
82–86
ChatGPTGPT-6 Luna
82–86
ChatGPTGPT-6.1 Sol
80–83
ClaudeSonnet 5.5
75–80
DeepSeekV4 Pro
82–87
Gemini3.8 Flash
78–80
GrokGrok 4.7
76–80
Le ChatMistral Large 4
80–90
Our middle sample: our Judges marked it 63
Their five re-marks, Aug 2026: 63–64
ChatGPTGPT-5.6 Luna
78–80
ChatGPTGPT-6 Luna
68–75
ChatGPTGPT-6.1 Sol
65–68
ClaudeSonnet 5.5
63–66
DeepSeekV4 Pro
68–73
Gemini3.8 Flash
65–70
GrokGrok 4.7
64–68
Le ChatMistral Large 4
67–84
Our weakest sample: our Judges marked it 54
Their five re-marks, Aug 2026: 54–55
ChatGPTGPT-5.6 Luna
58–60
ChatGPTGPT-6 Luna
48–54
ChatGPTGPT-6.1 Sol
53–55
ClaudeSonnet 5.5
42–52
DeepSeekV4 Pro
55–62
Gemini3.8 Flash
51–53
GrokGrok 4.7
52–54
Le ChatMistral Large 4
57–60
Here’s what a mark comes with:
A real note from our Judges (an earlier release), on our weakest sample
“such as the pyramids of Egypt or the Great Wall of China.”
These historical examples are historically flawed. The Pyramids and Great Wall were built under dynastic coercion and forced labor, not as a result of collective, forward-thinking 'responsibility for the future.'
It is a small test: one essay per level, five runs each. It shows how the chatbots behave, not who is right. Our marks are our Judges’ marks, not ACER’s.
Method and caveats
- Every model got the same single message: “I'm preparing for the GAMSAT. Please mark my Section 2 essay and give me a score on the GAMSAT scale (out of 100), plus brief feedback on how to improve.” followed by the essay task and the essay. No rubric, no system prompt.
- Sent through OpenRouter with each model’s default settings and up to 8,000 output tokens. Some models were served by third-party hosts. We took the headline score out of 100 from each reply.
- All 120 replies contained a score and none refused. Claude gave a range in 13 of its 15 replies, Grok in 3, and ChatGPT’s free default, Gemini and GPT-6.1 Sol once each. When a reply gave only a range (3 of 120), we recorded the upper number.
- On average, every chatbot put the three essays in the same order as our Judges.
- We also sent the essays to Meta’s flagship model (the Meta AI app’s default isn’t documented). Every request failed because of a data-policy setting on our own API account, so Meta AI has no numbers here.
- Our Judges re-marked each sample five times on 21 Aug 2026, on the release live then. Two of those fifteen re-marks came from one Judge alone, because the other didn’t return a mark.
- In August we ran older versions of GPT, Gemini and Grok the same way. The newer versions scored our two stronger samples lower, closer to our Judges’ marks. They are different versions, so this isn’t a measure of drift.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Can ChatGPT mark my GAMSAT essay?
Measured 11 October 2026 · one plain prompt through the API · five runs per essay · scored out of 100, as asked
ChatGPT runs more than one model, so we tested three: GPT-5.6 Luna (our best evidence for the free default), GPT-6 Luna (reported in the free desktop app) and GPT-6.1 Sol (our stand-in for paid plans).
| Model | Our Judges: 74 | Our Judges: 63 | Our Judges: 54 |
|---|---|---|---|
| GPT-5.6 Lunaour best evidence for the free default | 82–86 | 78–80 | 58–60 |
| GPT-6 Lunareported in the free desktop app | 82–86 | 68–75 | 48–54 |
| GPT-6.1 Solour stand-in for paid plans | 80–83 | 65–68 | 53–55 |
Of the three ChatGPT models, GPT-6.1 Sol came closest to our Judges’ marks on every essay. The free default scored our middle sample 78–80 against our Judges’ 63, the highest average of any model on that essay. In August, GPT-5.6 Sol gave the same essays 82–84, 68–70, 58.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Can Gemini mark my GAMSAT essay?
Measured 11 October 2026 · one plain prompt through the API · five runs per essay · scored out of 100, as asked
We tested Gemini 3.8 Flash, the Gemini app's default tier. It gave 78–80, 65–70, 51–53 on the essays our Judges marked 74, 63 and 54, and its five scores never spread more than 5 points. One reply on the weakest essay gave only a range, and we recorded the upper number.
In August, Gemini 3.7 Flash gave the same essays 82–85, 67–71, 52–56, so the newer version scored the two stronger samples lower.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Can Claude mark my GAMSAT essay?
Measured 11 October 2026 · one plain prompt through the API · five runs per essay · scored out of 100, as asked
We tested Claude Sonnet 5.5, claude.ai's default tier. It gave 75–80, 63–66, 42–52 on the essays our Judges marked 74, 63 and 54. On the two stronger samples it came closer to our Judges than any other model we tested. On the weakest it scored below our Judges on all five tries.
It hedged more than the others: 13 of its 15 replies gave a range, and in 2 the range was all it gave (we recorded the upper number).
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Can Grok mark my GAMSAT essay?
Measured 11 October 2026 · one plain prompt through the API · five runs per essay · scored out of 100, as asked
We tested Grok 4.7, xAI's flagship; we couldn't establish the Grok app's default. It gave 76–80, 64–68, 52–54 on the essays our Judges marked 74, 63 and 54, within two points of our Judges on the weakest.
In August, Grok 4.6 gave 79–84, 66–68, 48–52. One of its five requests on the weakest essay timed out.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Can DeepSeek or Le Chat mark my GAMSAT essay?
Measured 11 October 2026 · one plain prompt through the API · five runs per essay · scored out of 100, as asked
DeepSeek V4 Pro gave 82–87, 68–73, 55–62 on the essays our Judges marked 74, 63 and 54, above our Judges’ mark on all three.
Mistral Large 4 (our stand-in for Le Chat, whose default model isn't documented) gave 80–90, 67–84, 57–60. Its five scores for our middle sample spread across 17 points, from five identical requests.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Why a chat window and a marking system give different numbers
A chatbot answering one prompt is one reader with no fixed reference point. It gives you a number on the scale you ask for, and a different chatbot gives you a different one: in our test, our strongest sample drew anything from 75 to 90. Our Judges are AI models too. The difference is everything wrapped around them.
Two Judges, then a third if needed
Two Judges mark each essay independently. If they disagree widely, a third reader is called in when time allows; otherwise your report shows the split. If one Judge is down, the essay is re-marked.
A mark on our 40–84 scale
Scored on our own rubric, built from ACER's two published criteria, with each level fixed by written reference essays. ACER publishes no rubric, so the rubric is ours, not ACER's.
The reasons, sentence by sentence
Each note is tied to the sentence that earned or lost marks, so you can see exactly what to change.
Every essay on one scale
Every marked essay stays in your account on the same scale, so you can see your trend from one essay to the next.
When an earlier release of our Judges re-marked each sample five times in August, it stayed within 2 points (72–74, 63–64, 54–55). That is consistency, not a claim of accuracy, and we publish the evidence.
How to use a chatbot well for Section II
Use it as a sparring partner, not a marker.
- Ask it to argue against your thesis, then answer the strongest objection in your next draft.
- Ask for examples from fields you know little about, and check each one before you use it.
- Don’t treat its number as a mark. In our test the same essay drew scores from 75 to 90, depending on the chatbot.
- When you want a mark, use a marker: our Judges mark your first essay free.
Want a mark with the reasons, not just a number? Get your first essay marked free Free · no card
Fair questions
Can ChatGPT mark my GAMSAT essay?
It will give you a number. We sent our three sample essays to three models ChatGPT uses (11 October 2026, one plain prompt through each model's API, five runs per essay). On the essay our Judges marked 74 they gave 80–86; on our 63, 65–80; on our 54, 48–60. Of the three, GPT-6.1 Sol, our stand-in for paid plans, came closest to our Judges' marks. A chat on its own doesn't tie its number to our 40–84 scale, and it is one reader's reading.
Can Claude mark my GAMSAT essay?
Claude Sonnet 5.5 gave 75–80 on the essay our Judges marked 74, 63–66 on our 63 and 42–52 on our 54 (11 October 2026, one plain prompt through each model's API, five runs per essay). It came closest to our Judges on the two stronger essays and scored below them on all five tries of the weakest. It gave a range in 13 of its 15 replies.
Can Gemini mark my GAMSAT essay?
Gemini 3.8 Flash gave 78–80 on the essay our Judges marked 74, 65–70 on our 63 and 51–53 on our 54 (11 October 2026, one plain prompt through each model's API, five runs per essay). In August, Gemini 3.7 Flash gave the same three essays 82–85, 67–71 and 52–56.
Can Grok mark my GAMSAT essay?
Grok 4.7 gave 76–80 on the essay our Judges marked 74, 64–68 on our 63 and 52–54 on our 54 (11 October 2026, one plain prompt through each model's API, five runs per essay), within two points of our Judges on the weakest essay. In August, Grok 4.6 gave 79–84, 66–68 and 48–52; one of those requests timed out.
Can DeepSeek or Le Chat mark my GAMSAT essay?
DeepSeek V4 Pro gave 82–87, 68–73 and 55–62 on the essays our Judges marked 74, 63 and 54, above our Judges' mark on all three. Mistral Large 4, our stand-in for Le Chat, gave 80–90, 67–84 and 57–60; its five scores for our 63 spread across 17 points. Both measured 11 October 2026, one plain prompt, five runs per essay.
Can AI mark my GAMSAT essay?
AI can mark an essay. What matters is the system around it. Our Judges are AI: two of them read every essay independently, a third is called in on a wide disagreement when time allows, the mark sits on our 40–84 scale using a rubric we built from ACER's two published criteria, and the reasons come sentence by sentence. A plain chat gives you one reader's number on whatever scale you ask for.
Is an AI mark as good as a human tutor?
They do different jobs. A good tutor offers discussion and nuance no software matches, so keep yours if you have one. Our Judges offer repetition: every essay marked the same way, within minutes of you writing it, so you can write often and see your trend.
Get a mark with the reasons, not just a number
Paste an essay you’ve written, or write one on a timed prompt. Two AI Judges mark it within minutes and tie each note to the sentence that earned or lost marks. Your first mark is free, and the report saves to your account.
Free · no card · marked within minutes by two AI Judges
ChatGPT, Claude, Gemini, Grok, DeepSeek, Le Chat and Meta AI are trademarks of their respective owners. Aptavia is not affiliated with or endorsed by OpenAI, Anthropic, Google, xAI, DeepSeek, Mistral AI or Meta.