Generative AI Comparison
AInformation gave ChatGPT, Gemini, Claude and Copilot the same tasks with the same prompts and compared the results. We put 10 areas from the AI Practical Work Benchmark (scored by a program) and 11 features from the 4-AI Test Lab (measured in each article) into dated tables, one per method. Click a feature name to open the original article with the method and the full answers.
Each row shows the date it was tested. Includes tests up to October 29, 2026
Generative AI accuracy compared across 10 areas
AI Practical Work Benchmark. Accuracy from 6 tasks, each sent 3 times with the same prompt and scored only by a program. Bold blue marks the top AI in each area.
| Area | ChatGPT | Gemini | Claude | Copilot | Test date |
|---|---|---|---|---|---|
| Reading images and photos | 77.8% | 66.7% | 55.6% | 66.7% | |
| Meeting minutes and summaries | 83.3% | 55.6% | 55.6% | 55.6% | |
| Research and fact-checking | 100% | 22.2% | Could not measure | 72.2% | |
| Reading PDFs and financial statements | 100% | 73.3% | 100% | Not measured | |
| Spreadsheets and aggregation | 100% | 50% | 83.3% | 100% | |
| English text and translation | 83.3% | 16.7% | 50% | 55.6% | |
| Document outline creation | 83.3% | 77.8% | 61.1% | 27.8% | |
| Excel/spreadsheet formulas | 100% | 100% | 100% | 94.4% | |
| Email and business documents | 100% | 38.9% | 83.3% | 100% | |
| Data formatting & name matching | 83.3% | 77.8% | 66.7% | 66.7% |
Generative AI tested and compared on 11 features
4-AI Test Lab. Three tasks (easy, medium, hard) sent once each, with the answers checked in the article. ○△× show correctness per task, and bold blue marks the winner named in the article.
| Feature | ChatGPT | Gemini | Claude | Copilot | Test date |
|---|---|---|---|---|---|
| Image generation | Could not test | ○3 | ○1 △2 | ○1 △2 | |
| Sending a photo or image and asking about it (image recognition) | Could not test | ○3 | ○2 △1 | Could not test | |
| Creating meeting minutes from audio | Could not test | ○1 | ×1 | Could not test | |
| File (PDF) reading | ○3 | ○3 | ○3 | ○3 | |
| Code execution (data analysis and charts) | ○3 | ○2 ×1 | ○3 | ○1 △1 | |
| Creating and editing documents together (Canvas, Docs, Pages) | Could not test | Slides only | All three tasks | Tasks 2 and 3 | |
| Long-text summarization | Pass · 8 of 8 · 7 of 9 | Pass · 8 of 8 · 9 of 9 | Pass · 8 of 8 · 9 of 9 | Pass · 8 of 8 · 8 of 9 | |
| Web search (answers with sources) | ○2 △1 | ○2 ×1 | ○1 | ○2 △1 | |
| Memory (remembering across conversations) | ○1 | ×1 | ×1 | ○1 | |
| Reading spreadsheets (Excel) | ○2 | ○3 | ○2 △1 | ○2 ×1 | |
| Translation (English to Japanese, Japanese to English) | 14 of 14, 5 of 5 | ×1 | 14 of 14, 5 of 5 | 14 of 14, 5 of 5 |
No compared features match these filters yet.
ChatGPT, Gemini, Claude and Copilot: results compared
Cumulative totals from 10 areas and 784 runs of the AI Practical Work Benchmark. 4-AI Test Lab rows aren't included because they're measured differently. Only the far-right column counts the features where the AI came first in the tables above.
| AI | Accuracy | Time taken in seconds (median) | Same answer all 3 times | Followed the format | Source URLs opened | First place in this table |
|---|---|---|---|---|---|---|
| ChatGPT207 trials · 12 fields | 92.3% | 21.9 s | 100% | 96.1% | 100% | 12 features4 tied |
| Claude190 trials · 12 fields | 76.8% | 19.3 s | 95% | 53.8% | — | 6 features2 tied |
| Copilot180 trials · 11 fields | 72.2% | 16.6 s | 87.2% | 76.8% | 93.3% | 2 features2 tied |
| Gemini207 trials · 12 fields | 61.4% | 49 s | 86.1% | 83.1% | 83.3% | 6 features1 tied |
How we made this page
This table shows how AInformation compared ChatGPT, Gemini, Claude, and Copilot on the same tasks. The data comes from 10 fields in the AI Practical Work Benchmark (675 trials in total) and 11 comparison articles from the 4-AI test lab (132 submissions that produced an answer).
The AI Practical Work Benchmark recreates a fictional company's documents each time and sends the same task to all four AIs. A program checks whether each answer is correct by matching numbers, checking for specific words, checking the format, and confirming URLs open. The AIs never grade their own or each other's answers.
The 4-AI Verification Lab tests one feature on four AIs using three tasks: easy, medium, and hard. Each article checks the time taken and the content of every answer one by one, and its verification conditions describe the criteria used.
In the benchmark, the win mark goes to the AI with the highest accuracy rate (ties are marked as equal). In the lab, it goes to whichever AI the editorial team concluded was best in the article. Since each result comes from a single round of testing, rankings can change with a different day or task.
The data in this table is also available as JSON (CC BY 4.0).
FAQ
Which AI is the most accurate when comparing generative AI tools?
In the cumulative results of the AI Practical Work Benchmark (10 fields, 784 trials in total), ChatGPT had the highest accuracy rate at 92.3% (191 of 207 trials). Claude came in second at 76.8%, followed by Copilot at 72.2% and Gemini at 61.4%. That said, the rankings change by feature, so choose based on the row for the feature you want to use.
How did you compare ChatGPT, Gemini, Claude, and Copilot?
AInformation gave the same text for the same tasks to the paid-plan screens of all 4 AI tools. The AI Practical Work Benchmark runs 6 tasks per category, 3 times each, and a program alone judges whether each answer is correct. The 4-AI Verification Lab runs 3 tasks (easy, medium, hard) once each, and we check each answer ourselves in the article. Both start a new conversation every time.
Will I get the same results with the free versions?
Not necessarily. We used ChatGPT Pro, Google AI Pro, Claude's Max plan, and Microsoft 365 Personal, and recorded the model names shown on screen in the scoring table. Free versions offer different models and usage limits, so both response speed and accuracy may differ.
Why do some entries say “could not test” or “could not measure”?
There are 3 reasons we could not measure a result: the editorial team's operation was interrupted, the plan's usage limit was reached, or a human verification check appeared and we stopped. These are not cases where the AI gave a wrong answer, so we do not include them in the accuracy rate's denominator. Cases where the AI refused or timed out are counted as AI results, so we include them in the denominator as incorrect answers.
When is this comparison table updated?
Comparison articles from the 4-AI test lab are added to the table on the day they are published. AI Practical Work Benchmark results are added the day after testing. When we retest the same feature, we replace the old result with the new one. The table currently covers 21 features, and the most recent test was run on October 29, 2026. The data is published under CC BY 4.0, so you can quote it as long as you cite the source.