Claude vs Gemini: the cheaper model that wasn't

Pick Gemini when the job is to report what is there. Its memo summary added nothing that wasn't in the memo, and it answered the trick puzzle cleanly the first time. It is also the more flexible model on paper: OpenRouter lists it as taking video and audio as well as text, images and files.

Pick Claude when the job is to persuade or reassure someone. Its invoice email was the only one of the two that set a deadline and said what happens after it, and its explanations read like a person talking.

Both returned the same quiet code fix, and both explained compound interest correctly. If you mostly do those things, choose on price and speed instead, and read the next section first.

The memo prompt asked for three bullets on a four-day-week pilot and one question, with no facts added. Gemini kept to the memo, overtime and school pickups included (it did leave out the comparison group's 93%), and asked “How will the four-day work week impact our ticket response rates and staff workload when we transition out of our quietest season and into a busier period?”

Claude's question was sharper, but its bullets wandered from the source. It filed the results under its own heading, “Productivity mixed”, turned "school pickups" into "childcare difficulties" and dropped the line about overtime. For a summary that people will act on, Gemini's is the one to trust.

The jug puzzle asks whether you can measure 4 litres with a 6 and a 9-litre jug. You can't. Gemini said so in two sentences, starting “No, it is impossible to measure exactly 4 litres”. Claude's answer opened with “Solution: Yes, 4 litres can be measured.” and then, two lines later, “Wait, let me reconsider before listing steps.” The revised answer was right, but someone skimming would take away the wrong one.

We asked both to turn a blunt note about an $850 invoice, three weeks late, into something polite but firm. Claude (100 words) wrote “If I don't hear back by the end of this week, I'll need to explore further steps to settle the outstanding balance.” That is what firm means: a date and a consequence.

Gemini (92 words) added a subject line, which is useful, and asked for “your prompt attention to resolving this matter today”. "Today" with no consequence reads as pressure without a plan. Its email also cost $0.0058, more than twice Claude's $0.0022, because 512 tokens of thinking sat behind it.

Both explained simple and compound interest to a 15-year-old in under 150 words and both reached Rs 12,597. Gemini was chattier (“Compound interest is a superpower.”) and showed the simple-interest balance year by year; Claude called it “interest on interest” and pointed out the Rs 197 difference.

On the buggy Python function both found the same three problems and wrote nearly the same fix, one that returns None for a list that is too short. That fails quietly: the caller gets no error and may not notice. Gemini at least left a comment, “Or raise ValueError”.

We asked for a 3-day Jaipur plan for two on ₹15,000, with a cost per day and a total. Claude's day costs were ₹1,800, ₹3,500 and ₹2,500, which come to ₹7,800, but its summary table said ₹11,000. Gemini wrote out its sum in full: “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)” Those six numbers add up to ₹16,000.

Both plans look finished, and both are wrong in a way you only catch with a calculator. Two models in our wider test did keep the sums straight, Claude Opus 4.8 and Grok 4.3; see Claude Opus vs Sonnet and Grok vs ChatGPT.

Each prompt went once to anthropic/claude-sonnet-4.6 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. Claude ran with extended thinking off. Gemini started at the low thinking setting; on four prompts it spent almost the whole 700 tokens thinking and was cut off. A retry with thinking capped at 200 tokens was cut off too, so we re-ran those four at the minimal setting. The email and puzzle answers here are from the low setting, the other four from minimal. The full story of that is on Gemini vs ChatGPT.

With one run per prompt, read the verdicts as a spot check. Long answers are trimmed at [...], and dashes in the replies are shown as commas; the rest of each quote is exactly as the model returned it. These calls went to the model makers through OpenRouter, not through AI4Chat's chat.

All 934 of Gemini's thinking tokens came from just two prompts, the email (512) and the puzzle (422), the two we ran at the low setting. At $9.00 per million that thinking was about 38% of Gemini's bill. If the answers had come out the same without it, Gemini's six would have cost $0.0134, about 36% less than Claude's. So the list price is real, but only if you keep the thinking setting low and check that answers still finish.

These are API prices. The Claude and Gemini apps sell monthly plans, which we track on AI subscription prices; if you want to try several models without a subscription to each, see free AI models and our list of ChatGPT alternatives.

Both families have newer models on OpenRouter: Claude Sonnet 5.5 (28 Sep 2026, $2.00 in and $10.00 out per million tokens), Gemini 4 Argon (13 Oct 2026, $2.00 and $10.00) and Gemini 3.8 Flash (2 Sep 2026, $0.83 and $4.13). We haven't tested them, so we make no claims about how they compare.