ChatGPT vs Claude
We asked three GPT models and three Claude models 36 harder prompts on 8 November 2026 and 36 everyday prompts the day before, twice each. GPT-6.1 Sol and GPT-6 Astra scored highest of the six on the harder set, and Claude Haiku 5.5 lowest.
Updated 8 November 2026
Compare your own questionWhat the names mean
ChatGPT is OpenAI's chat app, and the models behind it are OpenAI's GPT models. Claude is the name of Anthropic's chat app and of its family of models.
We tested the models themselves through OpenRouter, and whichai.chat sends your messages the same way. Features of the ChatGPT and Claude apps aren't part of the comparison.
How GPT and Claude models scored on harder prompts
The average score across 36 harder prompts asked twice, so 72 answers per model.
| Model | Score | Full marks | Typical reply time | Cost of all answers |
|---|---|---|---|---|
GPT-6.1 Sol
OpenAI
|
99.7% | 71 of 72 | 4.8 s | $0.234 |
|
GPT-6 Astra
OpenAI
|
99.4% | 70 of 72 | 4.3 s | $1.15 |
Claude Sonnet 5.5
Anthropic
|
98.6% | 70 of 72 | 5.5 s | $0.524 |
|
Claude Opus 5.5
Anthropic
|
97.2% | 70 of 72 | 6.8 s | $1.07 |
|
GPT-6 Luna
OpenAI
|
95.6% | 68 of 72 | 3.3 s | $0.020 |
|
Claude Haiku 5.5
Anthropic
|
92.2% | 65 of 72 | 5.3 s | $0.041 |
-
GPT-6.1 Sol
OpenAI
99.7%
- Full marks
- 71 of 72
- Reply time
- 4.8 s
- Cost
- $0.234
-
GPT-6 Astra
OpenAI
99.4%
- Full marks
- 70 of 72
- Reply time
- 4.3 s
- Cost
- $1.15
-
Claude Sonnet 5.5
Anthropic
98.6%
- Full marks
- 70 of 72
- Reply time
- 5.5 s
- Cost
- $0.524
-
Claude Opus 5.5
Anthropic
97.2%
- Full marks
- 70 of 72
- Reply time
- 6.8 s
- Cost
- $1.07
-
GPT-6 Luna
OpenAI
95.6%
- Full marks
- 68 of 72
- Reply time
- 3.3 s
- Cost
- $0.020
-
Claude Haiku 5.5
Anthropic
92.2%
- Full marks
- 65 of 72
- Reply time
- 5.3 s
- Cost
- $0.041
Claude Opus 5.5 returned no answer twice, which counted as 0.
Scored with Raq.com's AI Benchmarker on 8 November 2026. One answer is worth about 1.4 points, so GPT-6.1 Sol and Claude Sonnet 5.5, the best of each, are less than one answer apart.
What we found
On average the GPT models scored higher, 98.2% to Claude's 96.0%, and GPT-6.1 Sol and GPT-6 Astra both finished above 99%. Claude Sonnet 5.5 was the best Claude model at 98.6%.
Claude Opus 5.5 got full marks on all 70 answers it gave. It returned nothing on both tries at reversing a string of random characters, and those two count as 0, which is the only reason it finished below Claude Sonnet 5.5.
The cheapest model from each company slipped the most. GPT-6 Luna, which cost about 2 cents for all 72 answers, said £6,250 for statutory redundancy pay on both tries, when the right answer is £6,000, and miscounted a letter once. Claude Haiku 5.5 gave 16 as the minimum age for the National Lottery on both tries, when it's 18, and sorted a word list in the wrong order on both.
Where the best of each lost marks
GPT-6.1 Sol and Claude Sonnet 5.5 scored highest for their companies. These are the prompts where either dropped points, with the reason.
| Prompt |
GPT-6.1 Sol
|
Claude Sonnet 5.5
|
|---|---|---|
| Five bullets that grow by a word Instructions. Right answer: 5 bullets of 3 to 7 words | 100% | 50% Missed: 3, 4, 5, 6 and 7 words |
| Decision record from an email thread Writing | 90% Judge: The four-line record is otherwise accurate, but the Owner line only names Owen Whitehead and does not explicitly state that he leads the stock check and sends the count sheets. | 100% |
-
Five bullets that grow by a word
Instructions. Right answer: 5 bullets of 3 to 7 words
-
GPT-6.1 Sol
- 100%
-
Claude Sonnet 5.5
- 50%
- Missed: 3, 4, 5, 6 and 7 words
-
-
Decision record from an email thread
Writing
-
GPT-6.1 Sol
- 90%
- Judge: The four-line record is otherwise accurate, but the Owner line only names Owen Whitehead and does not explicitly state that he leads the stock check and sends the count sheets.
-
Claude Sonnet 5.5
- 100%
-
On everyday prompts
Our first round, on 7 November 2026: 36 everyday prompts asked twice, with the same models and settings.
| Model | Score | Full marks | Typical reply time | Cost of all answers |
|---|---|---|---|---|
|
GPT-6 Astra
OpenAI
|
99.2% | 69 of 72 | 2.6 s | $0.370 |
|
GPT-6.1 Sol
OpenAI
|
98.9% | 68 of 72 | 2.6 s | $0.076 |
|
Claude Sonnet 5.5
Anthropic
|
98.6% | 68 of 72 | 2.0 s | $0.176 |
|
GPT-6 Luna
OpenAI
|
98.5% | 67 of 72 | 2.3 s | $0.005 |
|
Claude Haiku 5.5
Anthropic
|
97.7% | 68 of 72 | 2.4 s | $0.013 |
|
Claude Opus 5.5
Anthropic
|
97.6% | 64 of 72 | 4.0 s | $0.481 |
-
GPT-6 Astra
OpenAI
99.2%
- Full marks
- 69 of 72
- Reply time
- 2.6 s
- Cost
- $0.370
-
GPT-6.1 Sol
OpenAI
98.9%
- Full marks
- 68 of 72
- Reply time
- 2.6 s
- Cost
- $0.076
-
Claude Sonnet 5.5
Anthropic
98.6%
- Full marks
- 68 of 72
- Reply time
- 2.0 s
- Cost
- $0.176
-
GPT-6 Luna
OpenAI
98.5%
- Full marks
- 67 of 72
- Reply time
- 2.3 s
- Cost
- $0.005
-
Claude Haiku 5.5
Anthropic
97.7%
- Full marks
- 68 of 72
- Reply time
- 2.4 s
- Cost
- $0.013
-
Claude Opus 5.5
Anthropic
97.6%
- Full marks
- 64 of 72
- Reply time
- 4.0 s
- Cost
- $0.481
All six scored between 97.6% and 99.2%. The one wrong checked answer came from Claude Haiku 5.5, and every other lost mark was on writing and summaries, mostly for details the prompt never gave.
The 36 harder prompts
Six kinds of task. Checked answers were marked automatically, coding answers were run against hidden tests, and the judge marked the writing against a written scheme.
Maths and logic
- Cheapest way to fill a café order Checked
- Loan balance after 18 payments Checked
- Days between two dates Checked
- Multiply a 13-digit number by a 10-digit one Checked
- Five-digit numbers with no repeated neighbours Checked
- Four dice that total 14 Checked
- Count the letter r in a passage Checked
Instructions
- Filter, sort and format a word list Checked
- A sentence from A to O Checked
- A product update with six rules Checked
- Five bullets that grow by a word Checked
- A paragraph of exactly 47 words Checked
- Reverse a string of random characters Checked
Long documents
- Lease: the terms after the variation letter Checked
- Lease: rent, guarantee and limits Checked
- Supplier account: VAT and balances Checked
- Meeting minutes: who does what Checked
- Timesheet: hours, overtime and pay Checked
- Bank statement: balance and totals Checked
UK questions
- Christmas bank holidays in 2027 Checked
- Belfast or Newcastle Checked
- Minimum age for the National Lottery Checked
- Statutory redundancy pay Checked
- Registering a birth in Scotland Checked
- Default speed limit in Welsh towns Checked
- Saturday hours for a 15-year-old Checked
Coding
- Parse a duration such as 1h 30m Tests
- Evaluate an arithmetic expression Tests
- Compare two version numbers Tests
- Parse CSV with quoted fields Tests
- Word-wrap text to a width Tests
Writing
- Reply to a refund request with limits Judge
- Explain interest rates to a 12-year-old Judge
- Job advert from rough notes Judge
- Decision record from an email thread Judge
- Fix six errors and change nothing else Judge
GPT and Claude models on whichai.chat
The models whichai.chat recommends from each company today, with figures from the live model catalogue.
| Model | Company | Context window | Reads images |
|---|---|---|---|
|
GPT-5.6 Luna
|
OpenAI | 1.1M tokens | Yes |
|
GPT-6 Astra
|
OpenAI | 1.1M tokens | Yes |
|
GPT-6.1 Sol
|
OpenAI | 1.1M tokens | Yes |
|
Claude Sonnet 5.5
|
Anthropic | 1M tokens | Yes |
|
Claude Opus 5.5
|
Anthropic | 1M tokens | Yes |
The list follows the recommended models in the app and changes when new ones are released.
Run the comparison yourself
-
Pick one model from each
Start a chat and add a GPT model and a Claude model from the model picker. A chat takes up to ten models, so Gemini and others can join too.
-
Ask something from your own work
Use a real question, draft or file. Attach PDFs, text files or images, and tick Web Search if the answer needs current information.
-
Read the answers side by side
Each model's answer has its own tab. Tick Consensus Summary and, once every model has replied, one model reads them all and writes a single summary comparing them.
-
Carry on with the better answer
Choose which answers your next message builds on and keep going with the one that suits the job.
Questions
Is ChatGPT or Claude better?
In our harder test on 8 November 2026, GPT-6.1 Sol scored 99.7% and the best Claude model, Claude Sonnet 5.5, scored 98.6%, which is less than one answer apart. OpenAI's three models averaged 98.2% and Anthropic's 96.0%, pulled down by Claude Haiku 5.5. On the everyday prompts the day before they were closer, so try both on your own questions.
Do I need ChatGPT and Claude subscriptions?
No. whichai.chat runs on a Raq.com account, and answers from every model use the AI allowance in your free trial or plan.
Can I bring my ChatGPT or Claude conversations?
Yes. Upload the data export from ChatGPT or Claude under Import chats, and your past conversations appear beside the new ones.
Can I compare Gemini as well?
Yes. Add a Gemini model to the same chat and its answer gets its own tab next to the GPT and Claude answers.
Got a question you can’t settle?
Ask it once and see what comes back.
Sign-up goes through Raq.com.