Gemini vs ChatGPT came out closer than most online arguments suggest. On six everyday prompts, Gemini 3.5 Flash and GPT-5.5 won one task each, tied three, and both broke a ₹15,000 trip budget. Pick ChatGPT if you paste in code; pick Gemini if you summarise documents or want the lower bill, since its answers cost us about 40% less.
One catch decides whether Gemini works for you at all: it thinks before it writes, and that thinking counts against the answer's length limit. At the low thinking setting, four of its six answers ran out of room before they finished. At minimal thinking, all of them came back complete.
Source: Our OpenRouter test of GPT-5.5 and Gemini 3.5 Flash, 7 November 2026
Gemini vs ChatGPT, task by task
| Task | Call | GPT-5.5 | Gemini |
|---|---|---|---|
| Invoice email | Tie | Firm, 71 words | Firm, adds a subject |
| Rs 10,000 interest | Tie | Exact | Rounded |
| Python bug fix | ChatGPT | Raises an error | Returns None |
| Memo summary | Gemini | Weak question | Sharp question |
| ₹15,000 trip | Neither | ₹17,000 | ₹16,000, says ₹14,800 |
| Jug puzzle | Tie | Correct | Correct |
GPT-5.5 vs Gemini 3.5 Flash, one run per prompt. The details are below; full prompts are in the copy box.
How we tested, and the thinking setting that mattered
Each prompt went once to openai/gpt-5.5 and google/gemini-3.5-flash through OpenRouter's API on 7 November 2026, with no system prompt, no web search and a 700-token cap on each answer. Both models think before answering, and we started both at the low reasoning setting.
GPT-5.5 used very little of its budget on thinking (186 tokens across all six answers). Gemini spent between 592 and 669 tokens thinking on four prompts and was cut off mid-answer each time. A second try with thinking capped at 200 tokens made no difference (the model thought for 577 to 671 tokens anyway). Setting thinking to minimal fixed it; the email and the puzzle answers below are from the low setting, the other four from minimal. Quotes below are trimmed where marked [...], with the models' dashes printed as commas. Nothing else was edited, and none of these answers came through AI4Chat's own chat.
This isn't a quirk of our test. Any app or API call that limits answer length will hit it. Our eight cut-off tries cost $0.051 between them and returned nothing usable, which is a few pennies here and real money at volume. If you use Gemini through the API, set the thinking level on purpose, and raise the length limit when you raise the thinking.
Rupee maths and a Jaipur trip: the India-specific prompts
Two of our six prompts were written with Indian readers in mind. The first asked for simple and compound interest on Rs 10,000 at 8% for 3 years. Both models got Rs 12,400 and about Rs 12,597. Gemini's version suits a 15-year-old better: it showed the year-by-year balance for both methods, then signed off with “With compound interest, you make an extra Rs 197 for doing absolutely nothing!”
The second asked for a 3-day Jaipur plan for two on ₹15,000, train tickets excluded. Gemini's plan was the richer read, with Amer Fort, Nahargarh at sunset and a Chokhi Dhani dinner priced at ₹1,000 a head. Then it summed up: “Total Estimated Cost: ₹14,800 (Accommodation: ₹3,000 + Transport: ₹1,800 + Tickets: ₹1,200 + Day 1: ₹2,500 + Day 2: ₹4,000 + Day 3: ₹3,500)” Add those six numbers and you get ₹16,000. GPT-5.5's sum was correct, at ₹17,000, which is also over budget; it then suggested trimming shopping and transport to get back to ₹15,000.
Each plan's total, as the model stated it and as its own line items add up
| The budget we set | ₹15,000 |
| GPT-5.5, stated and added up | ₹17,000 |
| Gemini 3.5 Flash, as stated | ₹14,800 |
| Gemini 3.5 Flash, items added up | ₹16,000 |
We checked only whether each plan agrees with itself, not whether its hotel or ticket prices are right. Gemini named restaurants we did not check.
Source: Our OpenRouter test, 7 November 2026
The other four tests
The invoice email was a draw. GPT-5.5 wrote 71 words and asked Megan to flag any problem with the invoice by the end of the day. Gemini added a subject line and closed with “I appreciate your prompt attention to resolving this matter today.” Both are firm enough to send.
On the code fix, both found the same three bugs: repeated top values, short lists and sorting the caller's list in place. GPT-5.5 raised a clear error with a message. Gemini returned None with a comment that reads “Or raise ValueError”, which leaves the important decision to you. ChatGPT takes this one.
On the memo summary, both kept every number straight. The difference was the closing question. GPT-5.5 asked “Should the team collect three more months of data before extending the pilot, given that it ran during the quietest season?”, which is the finance lead's own point handed back. Gemini asked how ticket response rates and workload will hold up once the team moves into a busier period, which is the question a manager would actually need answered. Gemini takes this one.
The jug puzzle (measure 4 litres with a 6 and a 9) has no solution, and both said so straight away, with the same reason: anything you can measure is a multiple of 3.
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.Run them in any chat app. Add up the Jaipur answer before you trust it.
Published specs and API prices
| GPT-5.5 | Gemini 3.5 Flash | |
|---|---|---|
| Context window | 1,050,000 tokens | 1,048,576 tokens |
| Input, per million tokens | $5.00 | $1.50 |
| Output, per million tokens | $30.00 | $9.00 (thinking billed the same) |
| Longest answer | 128,000 tokens | 65,536 tokens |
| Released on OpenRouter | 24 Apr 2026 | 19 May 2026 |
| Accepts | Text, images, files | Text, images, video, audio, files |
| Model page | openrouter.ai | openrouter.ai |
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 |
| Gemini 3.5 Flash | $3.64 |
Hidden thinking tokens are included for both. The eight cut-off tries are not.
Source: Our OpenRouter test, 7 November 2026: average billed cost per answer, times 1,000
The list prices say Gemini should be far cheaper: $9.00 against $30.00 per million output tokens, less than a third. Per answer the gap was smaller, because Gemini spent 934 tokens thinking across the six final answers and wrote longer replies. For app subscriptions rather than API use, see AI subscription prices, which tracks ChatGPT and Google's plans in rupees and dollars.
The misses, side by side
| Check | GPT-5.5 | Gemini |
|---|---|---|
| Trip plan stayed within ₹15,000 | No | No |
| Trip total matched its own items | Yes | No |
| Every answer finished at low thinking | Yes | No |
| Code fix fails loudly on bad input | Yes | No |
| Memo question added something new | No | Yes |
| Cheaper of the two per answer | No | Yes |
From the six test answers above, one run each.
Gemini's misses were about trust. A total that disagrees with its own items by ₹1,200 is worse than an honest overspend, because it looks right. And at the low thinking setting it could cost more than GPT-5.5 for a short answer: its 91-word invoice email, with 512 tokens of thinking behind it, cost almost twice what GPT-5.5 paid for a 71-word one.
GPT-5.5's misses were about effort. It noticed it was ₹2,000 over budget and proposed two cuts of about ₹1,000 each instead of rewriting the plan, and its memo question handed the finance lead's point straight back.
Both models lost the same task in the same way: they are good at setting up a sum and unreliable at keeping a running total inside a long answer. Claude did no better on the Jaipur prompt, as you can see in Claude vs Gemini and ChatGPT vs Claude.
Newer Gemini and GPT versions
Both families have moved on since these versions. OpenRouter lists Gemini 4 Argon (13 Oct 2026, $2.00 in and $10.00 out per million tokens), Gemini 3.8 Flash (2 Sep 2026) and GPT-6.1 Terra (26 Oct 2026, $1.00 and $5.00). We haven't tested them, so this page makes no claims about them.
Questions people ask
Is Gemini better than ChatGPT for coding?
Not in our test. Both found the same three bugs, but GPT-5.5's fix raised a clear error where Gemini's returned None and left the choice to you. One short function is a small sample, so try your own code on both.
Is Gemini better than ChatGPT at maths?
They were level. Both did the compound interest example correctly, and both got the Jaipur budget wrong: GPT-5.5 went over, and Gemini misadded its own line items. Use either to set up a calculation, then check the total.
Is Gemini cheaper than ChatGPT?
On the API, yes. Gemini 3.5 Flash lists at $1.50 in and $9.00 out per million tokens, against $5.00 and $30.00 for GPT-5.5, and our answers cost about 40% less. Its hidden thinking is billed as output, so long thinking narrows the gap.
Why did you test Gemini 3.5 Flash and not a Gemini Pro model?
Gemini 3.5 Flash (19 May 2026) is newer than Gemini 3.1 Pro, and OpenRouter's listing describes it as offering "near-Pro level coding and reasoning at Flash-tier cost and speed". We wanted the newer model, not the bigger label. A Pro model may do better on hard reasoning.
Is Gemini better than ChatGPT for research?
This test can't tell you. Our runs had no web access, and research quality in the Gemini and ChatGPT apps depends mostly on their search tools. That needs its own test with citations checked by hand.
Which one handles Indian prompts better?
Both handled rupee amounts and Jaipur sights without trouble. Gemini gave the more detailed local plan; GPT-5.5 kept its arithmetic straight. Neither stayed within the ₹15,000 budget.