DeepSeek vs Grok: Cost, Limits and What Each One Refuses

Almost every DeepSeek vs Grok comparison online repeats the same wrong number. They quote DeepSeek at "$0.14 per month" against Grok at around $16, and neither figure describes a real product. DeepSeek's consumer chat is free, and $0.14 is a garbled per-million-token API rate, not a subscription. The short version: DeepSeek is dramatically cheaper and stronger on structured reasoning, maths and code, while Grok costs real money and buys you live information, image and video generation, and far fewer topic refusals.

This comparison uses prices taken from each vendor's own documentation on 14 September 2026, names the model versions actually shipping rather than the ones that were current a year ago, and treats the awkward part honestly: these two models refuse very different things, and that difference matters more than any benchmark on this page.

The Key Takeaways

• Price reality: DeepSeek Flash is $0.30 in / $1.20 out per million tokens at peak and half that off-peak. Grok 4.6 is $2 in / $6 out. That is roughly a sixfold gap, not the hundredfold gap the comparison farms imply.
• The $0.14 myth: no DeepSeek plan has ever cost $0.14 a month. The nearest real figure is $0.15 per million off-peak input tokens.
• Off-peak is a real lever: DeepSeek halves its rates outside 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Batch work overnight and you pay half.
• Refusals are the real difference: independent testing in August 2026 scored DeepSeek's V4-Flash-0731 build 63.8 on China-sensitive prompts against 19.8 on matched controls, a gap of 44 points.
• Hosting does not fix that: the widely repeated idea that a US host makes DeepSeek uncensored does not survive the evidence. Data residency and topic refusals are two different problems.

DeepSeek vs Grok at a Glance

Both are positioned as the model you add alongside a mainstream assistant rather than the one you replace it with. They pull in opposite directions: DeepSeek competes on cost and openness, Grok competes on recency and range.

What DeepSeek vs Grok Actually Cost

This is where most comparisons fall apart, because they mix two incompatible units. An API rate is charged per million tokens consumed. A consumer subscription is charged per month regardless of use. Putting "$0.14" next to "$16" implies a comparison that does not exist.

The API rates, from the vendor docs

According to DeepSeek's own API pricing documentation, the company lists two models. Flash costs $0.30 per million input tokens and $1.20 per million output at peak, with cache hits at $0.006. V4 Pro costs $1.32 input and $3.96 output at peak, with cache hits at $0.044. Grok 4.6 is $2 input and $6 output, per xAI's own launch post. That rate is not flat. xAI's model documentation sets it as the price for prompts under 200,000 tokens, and at or above that threshold the entire request bills at $4 and $12. Long-context work on Grok costs double the headline price.

Against Flash, Grok is about six times more expensive on input and five times on output. Against V4 Pro, the gap narrows to roughly 1.5x. The headline "DeepSeek is a hundred times cheaper" only works if you compare DeepSeek's cached-input rate against Grok's uncached output rate, which is not a comparison anyone actually faces. For the wider picture across every vendor, our full AI pricing comparison puts both of these in context.

Off-peak pricing is the lever nobody mentions

DeepSeek charges half price outside peak hours, and defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. That weekday qualifier is missing from nearly every summary of DeepSeek's pricing, and it matters: weekend work bills at the off-peak rate regardless of the clock. If your usage is batch processing rather than live chat, scheduling it outside those windows genuinely halves the bill. Grok publishes no equivalent discount.

What the consumer plans cost

xAI publishes three consumer prices on its own pricing page: a free tier at $0, SuperGrok at $30 a month and SuperGrok Plus at $100 a month. It names further tiers, including SuperGrok Lite, SuperGrok Heavy, Business and Enterprise, without attaching a price to any of them. That matters, because the figures repeated across comparison sites, typically a $10 Lite plan and a $300 Heavy plan, do not appear on xAI's own page at all. Only the $30 agrees. DeepSeek, by contrast, charges nothing for its consumer chat and bills only for API use. We track both in detail in our Grok pricing guide and our DeepSeek pricing guide, and if you only want to know whether you can avoid paying at all, the free-tier breakdown covers what the no-cost plan still includes.

What Each One Refuses to Answer

Price is the question people search for. Refusals are the question that actually decides which model you keep. These two sit at opposite ends of that spectrum, and the gap is measurable rather than anecdotal.

DeepSeek and China-sensitive topics

DeepSeek systematically avoids a recognisable set of subjects: the events of Tiananmen Square, Taiwan's political status, the treatment of Uyghurs, criticism of Xi Jinping, and the Cultural Revolution. This is not occasional caution. Research published by CTGT in August 2026 scored the V4-Flash-0731 build at 63.8 on China-sensitive prompts against 19.8 on matched non-China control prompts, a gap of 44 points across 152 matched pairs rated by four judges. One caveat the figure deserves: that exact build has since been retired, and DeepSeek now serves those calls with V4.1-Flash, which has not been measured the same way.

The same research found something more pointed. Between the preview build and the official release, the score on China-sensitive prompts rose from 57.4 to 63.8 while the score on control prompts fell. The model became more permissive in general and less permissive on that specific set of topics. An older and more widely quoted figure, that DeepSeek refused 85% of sensitive prompts, comes from Promptfoo in January 2025 and describes the R1 model, not anything currently shipping. Treat it as history.

Why a US host does not fix it

There is a popular belief that running DeepSeek through an American provider removes the censorship. The evidence does not support it. Promptfoo's testing was conducted through OpenRouter, a third-party router rather than DeepSeek's own app, and still produced those refusal rates. The behaviour appears to sit in the model rather than in a filter bolted onto the chat interface.

This is worth separating cleanly, because two different problems get blurred together. Where a model is hosted determines data residency, which is a genuine and important privacy question, and our guide to running DeepSeek through a US host on a Mac covers that side properly. What a model refuses is a property of the model itself, and changing the server it runs on does not reliably change it. If your concern is that your prompts stay out of China, hosting solves it. If your concern is getting a straight answer about Taiwan, it does not.

Grok and the unfiltered pitch

Grok is marketed on the opposite premise, as the assistant willing to engage with edgy, political and culturally charged material that competitors decline. In practice it does answer a far wider range of questions than DeepSeek, and it has a permissive image and video mode that no mainstream rival matches.

Two caveats. First, "unfiltered" is xAI's positioning rather than a measured property, and we found no independent study quantifying Grok's refusal rate the way CTGT quantified DeepSeek's. Second, that permissiveness has produced real incidents, including well-documented failures that are worth understanding before you hand it to a team. Fewer refusals is a feature and a liability at the same time.

Which Is Better for Real Work

Strip out the positioning and the two models separate cleanly by task, because they were optimised for different things.

Maths, code and structured output

DeepSeek's reputation was built here, and it holds. Both V4.1 Flash and V4 Pro are strong on multi-step reasoning, produce reliable structured output, and ship a one million token context window at no extra charge, with no length-tiered pricing, which makes long document and large codebase work practical at a price that would be uncomfortable on any frontier model. If your work is analysis, extraction or code, this is the cheaper tool and frequently the better one. Our DeepSeek versus ChatGPT comparison goes deeper on where that reasoning advantage shows up.

News, research and anything time-sensitive

Grok wins this outright, and it is not close. It reads live web and X data, which means it can answer questions about a developing story, and about what people are posting on X right now, that DeepSeek has no route to. DeepSeek does ship a Search mode in its chat app, so the gap is narrower than most comparisons claim, but it has no access to X and exposes no web-search tool through its API. For monitoring a developing story, checking what people are actually saying about a product, or research where recency is the whole point, Grok is the only one of the two that functions. Benchmarks here need care. At launch in August 2026 xAI reported Grok 4.6 scoring 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol, but xAI never names which version of that index it used. On Artificial Analysis' own current v4.3 index the numbers are different: Grok 4.6 scores 44 against GPT-5.6 Sol's 47. Treat the launch figure as a dated vendor claim, not a standing result.

Openness and running it yourself

DeepSeek publishes its weights and Grok does not. If you want to run a model on your own hardware, audit it, or build on it without depending on a vendor's uptime, DeepSeek is the only option here. That is a real structural advantage, and it is why DeepSeek keeps appearing in our roundup of the best open source AI models. Just do not assume open weights means unfiltered, as covered above.

DeepSeek vs Grok on a Mac

Neither company treats macOS as a priority, which shapes how you actually use them day to day.

Grok is reachable through the web app, through X, and through a workable desktop route covered in our walkthrough on running Grok on a Mac as an app. DeepSeek has no first-party Mac client at all, so you are using the browser, calling the API, or running the weights locally on Apple Silicon, which is realistic only on a machine with substantial memory.

This is the practical case for not choosing. These two models are complementary rather than competing: one is cheap and deep, the other is current and permissive, and the tasks they are each bad at are the tasks the other handles well. Fello AI puts DeepSeek, Grok, ChatGPT, Claude and Gemini behind one keystroke in a single native Mac app, with no API keys to manage and no separate subscription for each, which is the cheapest way to keep both available without committing to either. If Claude is the other name on your shortlist, our Grok vs Claude comparison weighs that pair on the benchmarks both vendors actually publish.

The Verdict

Choose DeepSeek if cost is the binding constraint, your work is analysis, maths or code, you need a very large context window, or you want weights you can run yourself. Schedule batch work off-peak and the bill halves again.

Choose Grok if you need information from this week rather than last year, you want image and video generation in the same tool, or you have hit the wall of a model declining to discuss something ordinary. Pay attention to what you are buying: Grok's price is the cost of recency and range, not of raw intelligence.

The honest answer for most people is that the deciding factor is neither price nor benchmarks. It is whether the questions you ask fall inside the set DeepSeek will not discuss. If they do, the cheaper model is not cheaper, it is unusable, and the comparison is over before it starts.

FAQ