ChatGPT vs Claude, in one line: pick ChatGPT (GPT-5.5) for code and for summaries where a wrong detail would hurt, and pick Claude (Sonnet 4.6) if you mostly write emails and explanations for other people. In our six-prompt test GPT-5.5 won three, Claude won one, one was a tie, and both failed the budget task. Claude's answers also cost us about 43% less.
Neither one should be trusted with money totals yet. Asked to plan a ₹15,000 trip to Jaipur, GPT-5.5 came back ₹2,000 over budget, and Claude's day-by-day costs didn't add up to its own summary table. Check the sums yourself.
Who should pick ChatGPT vs Claude
Pick ChatGPT (GPT-5.5) if you paste in code, contracts or reports and need the answer to stick to what is there. Its memo summary added nothing that wasn't in the memo, its code fix raised a clear error instead of failing quietly, and it answered the trick puzzle in two short sentences. It is also the faster of the two in our run: 4.0 seconds on average against 5.1 for Claude.
Pick Claude (Sonnet 4.6) if most of what you send is writing that a person will read. Its invoice email was the only one that named a consequence, and its explanation of compound interest read like a patient tutor rather than a worksheet. Per answer it cost $0.0035 against $0.0061 for GPT-5.5.
If you can't decide, that's a fair outcome: on half of these tasks a reader would struggle to tell the two apart. Claude's bigger model, Opus 4.8, did better than Sonnet on several of these tasks, but at about twice the price. We compared those two in Claude Opus vs Sonnet.
How we ran the test
We sent each prompt once to openai/gpt-5.5 and anthropic/claude-sonnet-4.6 through OpenRouter's API on 7 November 2026, between 16:34 and 16:39 UTC. There was no system prompt, no web search and no chat history, so what you see is the model, not an app's extra instructions. None of these answers came through AI4Chat's own chat. Each answer was capped at 700 tokens. GPT-5.5 thinks before it answers, so we set its reasoning to low; Claude ran with extended thinking off, which is its default.
Six prompts and one run each is a small sample. Run the same prompt twice and you will get different wording, sometimes a different verdict. Treat this as a careful spot check, not a benchmark. We trimmed long answers (marked [...]) and swapped long dashes in replies for commas (or "to" in number ranges), because this site does not print them. Nothing else in the quotes was changed.
Newer versions of both models are out (OpenAI's GPT-6 and 6.1, Anthropic's Claude Sonnet 5 and 5.5). They are not part of this test.
Source: Our OpenRouter test, 7 November 2026 (billed cost and wall-clock time per answer)
The six tests, one by one
1. A polite but firm email about an unpaid invoice
We gave both a blunt three-sentence note to a client called Megan about an $850 invoice that was three weeks late, and asked for something polite but firm, under 120 words.
GPT-5.5 kept it short (71 words) and ended with “Please confirm when payment will be made. If there’s an issue with the invoice, let me know by the end of the day so we can resolve it promptly.” That deadline is for raising problems, though, not for paying.
Claude (100 words) was the only one to state what happens next: “If I don't hear back by the end of this week, I'll need to explore further steps to settle the outstanding balance.” It also closed with the stock line about prompt attention, which we would delete.
Our call: Claude. Firm means a consequence, and only Claude wrote one.
2. Explain compound interest to a 15-year-old
The prompt asked for under 150 words and a worked example: Rs 10,000 at 8% for 3 years. The right answers are Rs 12,400 with simple interest and Rs 12,597.12 compounded yearly.
Both got both numbers. GPT-5.5 showed the three compounding steps and finished on “So, compound interest gives more: Rs 12,597.12 vs Rs 12,400.” Claude rounded to Rs 12,597 and added a line a teenager might remember: “Compound interest gives you Rs 197 extra here, sounds small, but over 20 to 30 years, this difference becomes enormous.”
Our call: a tie. GPT-5.5 is tidier; Claude is friendlier. Both are correct.
3. Fix a buggy Python function
The function was meant to return the second-largest number by sorting the list and taking nums[-2]. It has three faults: it returns the top number again when that number is repeated, it crashes on short lists, and it reorders the caller's list.
Both found all three. The difference is what happens on bad input. Claude's fix returns None, which can slip through and break something later. GPT-5.5 raised an error with a message and flagged a real ambiguity: “returns a duplicate maximum ([3, 3, 2] -> 3) if “second-largest” means second-largest distinct value”. It also found the answer in one pass instead of sorting.
Our call: ChatGPT. Loud errors beat silent ones, and asking what "second-largest" means is what a good reviewer does.
4. Summarise a memo without adding facts
We wrote a short memo (about 140 words) about a four-day-week pilot: 38 of 52 support staff, satisfaction flat at 4.4, tickets answered within 24 hours down from 94% to 91%, sick days down 22%, and a finance lead who wants three more months of data because the pilot ran in the quietest season. Three bullets, one question, no added facts.
GPT-5.5's bullets were faithful to the line. Its question, though, just repeated the finance lead's point back. Claude asked the sharper question, “what is the minimum acceptable threshold”, but took small liberties on the way: it headed a bullet “Productivity mixed”, turned "answered" into "resolution", and turned school pickups into childcare.
Our call: ChatGPT. The brief said no added facts, and "resolution" is a different metric from "answered". If you only care about the follow-up question, Claude's was better.
5. Plan a ₹15,000 Jaipur trip for two
Three days, two people, ₹15,000 total excluding train tickets, a cost for each day and a total. This was the task both models failed, in different ways.
GPT-5.5's days added up correctly, to ₹17,000. Then it wrote “To keep it within ₹15,000, reduce shopping by ₹1,000 and choose shared autos/public transport to save around ₹1,000.”
Claude stayed under budget on paper. Its summary table came to ₹11,000, but its own day totals were ₹1,800, ₹3,500 and ₹2,500, which make ₹7,800. One of the two is wrong, and the plan doesn't say which.
Our call: neither. We are not judging whether their hotel or entry prices are right, only whether each plan agrees with itself. Neither did.
6. A logic puzzle with no solution
Can you measure exactly 4 litres with only a 6-litre and a 9-litre jug? You can't: anything you measure is a multiple of 3. The prompt invited a list of steps, which is the trap.
GPT-5.5 answered in two short sentences and got it right. Claude opened with “Solution: Yes, 4 litres can be measured.” A line later it wrote “Wait, let me reconsider before listing steps.” Its revised answer was correct and clearly explained.
Our call: ChatGPT. Claude got there, but an answer that opens with the wrong conclusion is a problem if you skim.
1. Rewrite this email so it is polite but firm, and keep it under 120 words. Return only the email.
Hi Megan, the invoice for the September logo work (#1042, $850) was due three weeks ago. I've emailed twice and heard nothing. When will it be paid?
2. Explain the difference between simple and compound interest to a 15-year-old in under 150 words. Include one worked example: Rs 10,000 at 8% a year for 3 years, showing both final amounts.
3. This Python function should return the second-largest number in a list, but it gives wrong answers for some inputs. Find every problem and give a fixed version. Keep the explanation under 80 words.
def second_largest(nums):
nums.sort()
return nums[-2]
4. Summarise this memo in exactly 3 bullet points, then add one question the team should ask before deciding. Do not add facts that are not in the memo.
MEMO: Four-day week pilot, support team
From March to May, 38 of our 52 support staff worked four 9-hour days instead of five 8-hour days. The other 14 stayed on the normal schedule as a comparison group. Customer satisfaction scores stayed flat at 4.4 out of 5 for both groups. The share of tickets answered within 24 hours fell from 94% to 91% in the pilot group and stayed at 93% in the comparison group. Sick days in the pilot group fell by 22% compared with the same months last year. Nine pilot staff said the longer days made school pickups harder. Overtime costs did not change. HR recommends extending the pilot to the whole team from July, but the finance lead wants three more months of data first because the pilot ran during our quietest season.
5. Plan a 3-day trip to Jaipur for two people on a total budget of Rs 15,000, not counting train tickets. Give a day-by-day plan with an estimated cost for each day and a total. Keep it under 200 words.
6. Using only a 6-litre jug and a 9-litre jug (no markings, unlimited water), can you measure exactly 4 litres? If yes, list the steps. If no, explain why in two sentences.Paste one into any chat app to run your own comparison.
Specs and prices from the published model pages
| GPT-5.5 | Claude Sonnet 4.6 | |
|---|---|---|
| Context window | 1,050,000 tokens | 1,000,000 tokens |
| Input, per million tokens | $5.00 | $3.00 |
| Output, per million tokens | $30.00 | $15.00 |
| Longest answer | 128,000 tokens | 128,000 tokens |
| Released on OpenRouter | 24 Apr 2026 | 17 Feb 2026 |
| Model page | openrouter.ai | openrouter.ai |
API list prices, which are what we paid in this test. App subscriptions are priced separately.
Source: OpenRouter model list, checked 7 November 2026
What 1,000 answers like the six in this test would cost, in US dollars
| GPT-5.5 | $6.12 |
| Claude Sonnet 4.6 | $3.50 |
| Claude Opus 4.8, for reference | $7.43 |
GPT-5.5 charges twice Claude's output rate but wrote shorter answers, so the gap per answer is smaller than the price list suggests. Its hidden reasoning tokens are included.
Source: Our OpenRouter test, 7 November 2026: average billed cost per answer, times 1,000
On price lists, GPT-5.5 is the expensive one: $30.00 per million output tokens against $15.00. In practice the gap shrinks because GPT-5.5 is terse. Across the six answers it wrote 967 visible tokens to Claude's 1,304, plus 186 tokens of thinking you never see but still pay for. If you pay a flat monthly subscription instead of per token, this matters less; we track those plans on AI subscription prices.
Where each one fails
One pattern held for both: they are better at reasoning about a problem than at keeping a running total. If an answer contains a sum you will act on, add it up yourself. We saw the same thing when we ran these prompts through Gemini in Gemini vs ChatGPT.
What about the newer versions?
OpenRouter lists newer models from both companies, among them GPT-6.1 Terra (released 26 Oct 2026, $1.00 in and $5.00 out per million tokens) and Claude Sonnet 5.5 (28 Sep 2026, $2.00 and $10.00). Both are listed cheaper than the versions we tested. We haven't run our prompts through them, so we make no claim about how they compare.
Questions people ask
Is Claude better than ChatGPT for coding?
Not in our one coding test. Both found all three bugs in the same function, but GPT-5.5 wrote the safer fix: it raised a clear error on bad input and asked what "second-largest" should mean. A single short bug is a thin basis for a coding verdict, so test your own code before you switch.
Is Claude better than ChatGPT for writing?
For writing that a person will read, we preferred Claude. Its invoice email was the only one with a real consequence in it, and its explanation was warmer. GPT-5.5 was shorter and stuck closer to the facts, which is what you want in a summary.
Which is cheaper, ChatGPT or Claude?
On the API, Claude Sonnet 4.6 lists at $3.00 in and $15.00 out per million tokens, against $5.00 and $30.00 for GPT-5.5. Our six answers cost $0.021 on Claude and $0.037 on GPT-5.5. Monthly app plans are priced differently and change by country.
Can ChatGPT and Claude do maths reliably?
Simple, stated maths, yes: both got the compound interest example right to the paisa or the rupee. Running totals inside a long plan, no: both budgets went wrong. Use them to set up the sum, then check it.
Why did you test through the API instead of the ChatGPT and Claude apps?
The apps add their own instructions, memory and web search, and change them often, so two people can get different answers on the same day. The API gives the model on its own, which makes a fairer side-by-side. The trade-off is that the apps may answer a little differently from what you see here.
Is ChatGPT or Claude better for research?
We didn't test that. These runs had no web access, so they say nothing about search or citations. Research quality depends heavily on each app's search tools, which deserve their own test.