DeepSeek vs Gemini: Gemini wins 41 to 39 on the Intelligence Index, at 4.6x the cost per task

Which is smarter per dollar, DeepSeek or Gemini?

DeepSeek is, by a wide margin. Artificial Analysis runs each model through the same ten tests and works out what one task costs at list price. That cost counts the tokens a model burns while it thinks, as well as what it charges for each one. DeepSeek's Flash model scores 39 for $0.27 a task. Gemini's Flash scores 41 for $1.24. So Google's two extra points cost 4.6 times as much.

Put it per dollar and DeepSeek gets about 144 index points where Gemini Flash gets 33. The dotted line on the chart is the frontier: the best score you can buy at each price. Both labs sit on it, which is why neither one wins outright. The Pro models flip the order: DeepSeek V4 Pro scores 36 for $0.67 a task, and Gemini 3.1 Pro scores 30 for $1.30.

All of the scores, the cost per task and the speeds come from each model's page on Artificial Analysis (Intelligence Index v4.3.2). The blended price weights input three to one over output, which is the ratio Artificial Analysis uses, and it's built from the list rates on DeepSeek's and Google's API pages that same day.

Two rows need a caveat. Gemini 4 Argon tops the chart, but Google's 30 September post says it's "rolling out to a set of trusted cyber defenders through our Fairwind Program." Paid API customers and Google AI Ultra subscribers come next, with no date. You can't buy it yet. Gemini 3.5 Flash-Lite is the cheapest point at $0.19 a task, but it scores 22, so it's a model for short, simple jobs.

Is DeepSeek better than Gemini for coding?

For agent work in a terminal, yes. On Terminal-Bench 4.0, which is one of the ten tests in the index, DeepSeek V4.1 Flash scores 26.8% and Gemini 3.8 Flash scores 19.7%. Gemini 3.1 Pro, one of the two models that come with Google AI Pro, scores just 4.0% on it.

Gemini wins when the job is code written from a spec. On SciCode, which asks for scientific functions from a written description, Gemini Flash scores 56.6% to DeepSeek's 51.9%. Gemini 3.1 Pro does better still at 58.7%. Argon beats all of them on both tests, with 57.1% on Terminal-Bench, so that's the bar to watch when Google opens it up.

DeepSeek guesses far more often when it doesn't know

DeepSeek's Flash model nearly always answers, even when it's wrong. The AA-Omniscience test asks hard questions of fact and counts how often a model gives a wrong answer when it could have said it doesn't know. Artificial Analysis defines this hallucination rate as "the proportion of incorrect answers out of all non-correct responses".

On that test DeepSeek V4.1 Flash guesses wrong on 96.5% of the questions it misses, and its Pro model does so on 94.8%. For Gemini 3.8 Flash it's 55.2%, for Gemini 3.1 Pro it's 50.9%, and for Argon it's only 15.1%. Lower is better.

That gap matters more than the two index points. If you'll quote the answer, send it to a client or cite it, Gemini is the safer pick, and search should be on either way. Code is the easy case. A test suite catches most of DeepSeek's guesses before they ship.

DeepSeek vs Gemini API pricing in October 2026

DeepSeek charges $0.30 for each million tokens you send to V4.1 Flash and $1.20 for each million it writes, and V4 Pro costs $1.32 and $3.96. Its pricing page cuts both in half off peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, except on Chinese public holidays.

Gemini 3.8 Flash costs $0.75 and $3.75 until 31 December 2026. Then it doubles to $1.50 and $7.50 on 1 January 2027, as Google's Gemini API pricing page shows. Gemini 3.1 Pro Preview is $2 and $12 when the prompt is 200K tokens or less. Argon starts at $2 and $10, and Google says $4 and $20 will apply once the intro period is over.

So the cost gap above widens in January. Every Gemini Flash token will cost twice what it does now, while DeepSeek's list rate stays put. Google does run a free API tier on 3.8 Flash and 3.5 Flash-Lite, which DeepSeek's pricing page doesn't list. The catch is in Google's own table: free tier traffic is marked as used to improve Google's products, and paid traffic isn't.

Is Gemini free like the DeepSeek app?

Yes, both have a free app, but the limits are very different. The DeepSeek app is listed as Free on the App Store, and it's published by Hangzhou DeepSeek Artificial Intelligence. On Google's US plans page the free tier of Gemini gives you 3.5 Flash-Lite, and Google's limits page caps the free app at a context window of 32K tokens.

Google sells three steps up from that. Google AI Plus is $4.99 a month, and for that you get 3.6 Flash, 3.1 Pro and a 128K window. Google AI Pro is $19.99 a month for 3.8 Flash, 3.1 Pro, a 1M window and 4x the free usage. Ultra starts at $99.99. Is Google AI Pro worth it does the math on the middle plan, and Gemini usage limit reached is for when you hit the cap.

Gemini's models can also take in more kinds of input. Gemini 3.8 Flash reads text, images, speech and video. DeepSeek V4.1 Flash reads text and images, and V4 Pro reads only text. The DeepSeek app has a failure mode of its own as well: the server busy message.

Where does each one keep your chats?

DeepSeek keeps them in China. Its privacy policy, last updated 10 February 2026, says: "we directly collect, process and store your Personal Data in People's Republic of China." It uses that data "to train and improve our technology" and gives you a right to opt out. It names no settings switch for that, only the [email protected] address.

Google doesn't name one country. Its privacy policy, in effect from 1 October 2026, says your information "may be processed on servers located outside of the country where you live." In the Gemini app, Keep Activity is on by default for anyone 18 or over. Google keeps that activity for 18 months by default and uses it to improve its services, training its generative AI models included. Some chats are read by human reviewers.

The app is one door into DeepSeek, not the only one. DeepSeek vs ChatGPT explains how the app, the API and the open weights differ, and the weights are where the China question goes away.

Can you download DeepSeek or Gemini and run it yourself?

DeepSeek, yes. Gemini, no. DeepSeek publishes V4.1 Flash on Hugging Face, and the model card says the weights are licensed under MIT. It's a 552B parameter model that uses 16B per token, so it needs a server, not a laptop. Any host you trust can run it, under that host's privacy terms.

Gemini's weights stay with Google. Its open family is Gemma, and Gemma 4 31B ships under Apache 2.0 with a 256K window. It scores 15 on the same index, which is 24 points under DeepSeek's open model. If you want open weights, DeepSeek is the stronger download by a long way.

Pick DeepSeek if, pick Gemini if

Whizi carries both of these families on one bill, but neither of them on Starter. On Pro, which is $29.99 a month with 2,500 credits, you get DeepSeek V4.1 Flash at 1 credit a message, Gemini 3.8 Flash at 2 and Gemini 3.1 Pro at 10. Pro's side by side view sends one prompt to both of them. Powerhouse at $49.99 adds DeepSeek V4 Pro and Gemma 4. Gemini 4 Argon isn't in Whizi, because Google hasn't released it outside of its test program. The DeepSeek and Gemini docs list every row.