DeepSeek vs Grok: Grok is smarter by 7 points, DeepSeek costs 10 times less per task

Which is better, DeepSeek or Grok?

Grok is better on measured quality. DeepSeek wins on anything that costs money or time. On the Artificial Analysis Intelligence Index (v4.3.2, read 12 October 2026), Grok 4.7 at its top effort setting scores 46 and DeepSeek V4.1 Flash scores 39. Both came out in September 2026.

Each dot is one model at one effort setting. The cost on the bottom axis is what one task in the index costs at list price, and it counts the tokens a model spends thinking as well as the price of each token. Grok's 7 extra points cost $2.73 a task at high effort against DeepSeek's $0.27. That's 10 times as much. The xhigh setting costs $3.74 and scores no higher.

Turning Grok down doesn't close the gap. At low effort Grok 4.7 scores 42, still three points over DeepSeek, and a task costs $1.25. So the cheapest Grok that beats DeepSeek costs 4.6 times as much. Per dollar, DeepSeek gets 144 index points, and Grok gets between 12 and 34 depending on the setting.

DeepSeek also answers sooner. In Artificial Analysis's runs its first token arrived in 0.97 seconds, while Grok 4.7 at its top setting took 88.14 seconds, because it thinks before it writes. That's a long wait. Once they start, DeepSeek streams 211 tokens a second and Grok streams 83.

How do the current DeepSeek and Grok models compare?

DeepSeek's API lists two models, and Grok's docs point to one flagship, tested here at three effort settings. Every figure below is from Artificial Analysis, read 12 October 2026, with list prices from DeepSeek's pricing page and Grok's model docs the same day:

Grok's docs leave no doubt about which model to use: "For everything else, including code, use Grok 4.7. It is the most capable model we've built." DeepSeek's choice is less obvious. Its Pro model scores 36, three points under Flash, and it costs 2.5 times as much per task. It reads only text, too, while Flash and Grok 4.7 both take text and images.

One name has changed. Artificial Analysis lists Grok's maker as SpaceXAI, the name xAI took in July 2026 after its merger with SpaceX. The app, the API and the Grok name are the same.

Is DeepSeek or Grok better for coding?

For agent coding in a terminal it's a tie, so DeepSeek wins on price. On Terminal-Bench 4.0, one of the ten tests in the index, DeepSeek V4.1 Flash scores 27% and Grok 4.7 scores 26%. DeepSeek does that for $0.27 a task, and Grok scored its 26% at the xhigh setting, which costs $3.74.

Grok wins when the code comes from a written spec. On SciCode, which asks for scientific functions from a description, Grok 4.7 scores 57% to DeepSeek's 52%. DeepSeek's Pro model is far behind on agent work, at 14% on Terminal-Bench.

Big codebases favor DeepSeek. Its window is 1M tokens to Grok's 500K, and on AA-LCR, the long context reasoning test, Flash scores 84% to Grok's 77%. Grok also gets pricier past 200K: its docs say a prompt that reaches that threshold is "billed at the higher rate for all tokens in the request", which is $4 in and $12 out per million.

Which one makes things up more often?

DeepSeek does, by a wide margin. The AA-Omniscience test asks hard questions of fact. It rewards right answers, takes points off for wrong ones, and doesn't penalize a model that says it doesn't know. Scores run from -100 to 100, and Artificial Analysis spells out the floor: "negative scores mean more incorrect than correct."

Grok 4.7 scores 32 on it. DeepSeek V4.1 Flash scores -5, and V4 Pro scores 1. So when DeepSeek's Flash model is unsure, it guesses, and it guesses wrong more often than right.

Grok leads on the reasoning tests as well, with 43% to 39% on Humanity's Last Exam and 18% to 14% on CritPt, a physics research test. If you'll quote the answer or send it to a client, Grok is the safer of the two, and you should keep web search on in either. Code is the easier case, because a test suite catches a wrong guess before it ships.

Writing style is the one thing none of these tests measure. The reviews on page one that judge tone, like Tom's Guide's seven prompt test, put Grok 3 up against DeepSeek R1 in February 2025, and both of those models are two generations old now. In that test Grok won five of the seven prompts, the short story among them, and DeepSeek won the joke and the one about recent AI news. If voice is what you care about, try your own prompt in both of them before you pick one.

How much do DeepSeek and Grok cost on the API?

DeepSeek charges a fifth of Grok's output price, and much less than that off peak. As of October 2026, DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens at peak. DeepSeek halves both outside its peak hours, which are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and it charges the off peak rate all day on weekends and Chinese public holidays.

Grok 4.7 costs $2 in and $6 out, with cached input at $0.50. Grok 4.3 is the cheaper option in Grok's own list, at $1.25 in and $2.50 out with a 1M window. Artificial Analysis gives it a score of 25, which is the same as DeepSeek Flash with its reasoning off, and it marks the model as deprecated.

Input is where Grok's bill grows. Its 81,000 output tokens per index task come to about $0.49 at $6 per million, out of a $3.74 total, so most of the cost is the context it reads. Cached input shows the widest gap in the two price lists: DeepSeek charges $0.006 per million cached tokens at peak and Grok charges $0.50, which is 83 times as much. Agents re-read the same context on every step, so that rate adds up fast.

Are DeepSeek and Grok free?

Both apps are free, and only Grok sells a paid plan. The DeepSeek app is listed as Free on the App Store with no in-app purchases. Grok's free plan on its pricing page lists "Real-time web and X search" and voice mode, and the paid tiers run from $10 to $300 a month.

DeepSeek has no paid tier to buy your way out of a busy server. When it's full, you get the server busy message and wait. Grok sells extra usage credits from $5 to $100 in the App Store as well.

DeepSeek can't match X search at any price. When a launch or a crisis is an hour old, Grok can read the posts before most articles about it exist. Is SuperGrok worth it works through whether the $30 plan pays off.

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, by email to privacy@deepseek.com rather than through a switch in the app.

Grok's privacy policy, updated 11 October 2026, calls its maker "a US-based company" but doesn't say where the data sits. It may use your information "to train our models", and its consumer FAQ puts the off switch in the app.

The DeepSeek app isn't the only way to run its model. The weights are published under MIT on Hugging Face, so any host you trust can serve them under its own privacy terms. Grok 4.7's weights aren't published. Is DeepSeek safe goes further into the China question.

Pick DeepSeek for volume and long input, Grok for facts and X

Whizi carries both of these on one bill. On Pro, which is $29.99 a month with 2,500 credits, DeepSeek V4.1 Flash costs 1 credit a message and Grok 4.7 costs 6, and side by side model chats send one prompt to both. Starter has neither. If live X posts or Grok's video are why you want Grok, its own app is the place for those. The DeepSeek and Grok docs list every row, and DeepSeek vs Gemini runs the same numbers against Google.