Skip to content
llmwise
Sign inStart free

DeepSeek V4 Pro · Writing

DeepSeek V4 Pro for writing: our test runs

DeepSeek V4 Pro ran our 5 writing prompts in llmwise, through the same pipeline your messages take. Here's every reply as it came, whether it passed, what it cost and what it counts as on Pro, next to the other models.

Start freeSee pricing

Based on 5 of our test runs on , through OpenRouter with the app's own prompt and settings. Updated .

Short answer

In our test runs on October 8, 2026, DeepSeek V4 Pro passed 5 of 5 of our writing prompts (both hard ones): tied 1st of 20 models. Each reply counted as 1 message on Pro, where it has 250 a month, and cost $0.0011 to run on average.

Prompt by prompt

Each of the 5 writing prompts, and how DeepSeek V4 Pro's reply did.

DeepSeek V4 Pro on each writing prompt
PromptResultCostTime
Announce a second bakery shop on LinkedInEverydayPassed: Graded 4.5 of 5 on average (lowest 4).$0.00103.9 s
Rewrite corporate jargon in plain wordsEverydayPassed: Graded 4.3 of 5 on average (lowest 4).$0.00245.8 s
Decline a meeting and offer two timesEverydayPassed: Graded 4.7 of 5 on average (lowest 4).$0.00063.3 s
A product announcement with five rulesHardPassed: Graded 4.3 of 5 on average (lowest 4).$0.00051.8 s
Argue both sides of free busesHardPassed: Graded 4.7 of 5 on average (lowest 4).$0.00094.1 s

Every writing reply from DeepSeek V4 Pro

  1. Announce a second bakery shop on LinkedIn

    An everyday prompt: “Announce a second bakery shop on LinkedIn”, in full.

    Passed: Graded 4.5 of 5 on average (lowest 4).

    Big news from Crumb & Co. — we're opening a second shop, and this time we're heading to Leeds.

Our new bakery will open its doors on Thursday 14 November, and we'd love you to be among the first through them. Come and see the new space, smell the bread coming out of the oven, and say hello to the team.

To mark the day, we're giving away free coffee to everyone who visits — our way of saying thank you for supporting a small, independent bakery.

We've spent months getting this little shop ready, and we're equal parts nervous and excited. If you're in Leeds, we'd be glad to see you. Bring a friend, grab a warm loaf, and help us make it feel like home.

See you on the 14th.

    400 tokens in, 199 out (35 of them reasoning) · 3.9 s · $0.0010 · 1 message on Pro · answered by deepseek/deepseek-v4-pro-0813 via Wafer ·

  2. Rewrite corporate jargon in plain words

    An everyday prompt: “Rewrite corporate jargon in plain words”, in full.

    Passed: Graded 4.3 of 5 on average (lowest 4).

    In the third quarter, our teams worked together to focus more on customers. This produced a 12% increase in customer retention and a 0.30 percentage point drop in churn among business customers. We also fine-tuned our various customer contact channels to use resources more efficiently.

    431 tokens in, 576 out (518 of them reasoning) · 5.8 s · $0.0024 · 1 message on Pro · answered by deepseek/deepseek-v4-pro-0813 via Wafer ·

  3. Decline a meeting and offer two times

    An everyday prompt: “Decline a meeting and offer two times”, in full.

    Passed: Graded 4.7 of 5 on average (lowest 4).

    Subject: Vendor demo – alternative times

Hi Tom,

Thanks for the invitation to the vendor demo on Wednesday at 11:00. Unfortunately, I’m presenting to the board at that time, so I won’t be able to join.

Would Thursday at 10:00 or Friday at 14:00 work for you instead? I’m happy to make either time work.

Best regards

    406 tokens in, 120 out (38 of them reasoning) · 3.3 s · $0.0006 · 1 message on Pro · answered by deepseek/deepseek-v4-pro-0813 via Wafer ·

  4. A product announcement with five rules

    A hard prompt: “A product announcement with five rules”, in full.

    Passed: Graded 4.3 of 5 on average (lowest 4).

    Introducing the Nordvik 2 desk lamp. It dims in five steps, so you can find the right light for any task. It charges over USB-C, keeping your desk free of extra cables. The arm turns 270 degrees for precise positioning. Available for €79, with shipping on 1 December. Have you reserved yours yet?

    418 tokens in, 96 out (25 of them reasoning) · 1.8 s · $0.0005 · 1 message on Pro · answered by deepseek/deepseek-v4-pro-0813 via Wafer ·

  5. Argue both sides of free buses

    A hard prompt: “Argue both sides of free buses”, in full.

    Passed: Graded 4.7 of 5 on average (lowest 4).

    Making buses free turns public transit into a true public good, like sidewalks or libraries. It removes fares as a barrier for low-income riders, boosts ridership, cuts traffic congestion and emissions, and speeds boarding by eliminating payment delays. Cities already subsidize most transit costs; dropping the small fare share is a modest investment with outsized returns in mobility, equity, and cleaner air.

The strongest case against free buses is that fares, however small, are the only direct signal of demand. Removing them invites overcrowding on popular routes while doing nothing to expand service where it is sparse. Scarce transit dollars are better spent on frequency, reliability, and coverage—improvements that actually attract riders—than on subsidizing trips for people who can already afford to pay.

    395 tokens in, 193 out (40 of them reasoning) · 4.1 s · $0.0009 · 1 message on Pro · answered by deepseek/deepseek-v4-pro-0813 via Wafer ·

How it compares on writing

DeepSeek V4 Pro beside the job's picks among every model. 3 models share the hard-problems pick, DeepSeek V4 Pro among them.

DeepSeek V4 Pro beside other models on writing
ModelPassedOn ProCost per reply
DeepSeek V4 ProOur pick: hard problems (shared), best value5 of 5250 a month on Pro$0.0011
GPT-6 LunaOur pick: hard problems (shared), everyday5 of 560 a day on Pro$0.0001
On Pro: Pro's count on each model. Cost: what OpenRouter charged us per reply, on average.

Every limit is published. Paid plans also have a monthly fair-use limit on AI cost: Plus $3.75, Pro $7.50, Max $20, Ultra $42, Studio $85. Using every message on your plan at typical sizes stays under it; very large messages and heavy research use it faster. Every limit, explained.

DeepSeek V4 Pro next to each model on writing

DeepSeek V4 Pro against each model on the same 5 writing prompts. One run each, through OpenRouter: a wait depends on the provider and the load that day, so a lead under 10% counts as close.

  • Against Claude Fable 5.1: DeepSeek V4 Pro answered 1.5× sooner (3.9 s to 6.0 s), cost 16.1× less, and passed 5 of 5 to its 3. Only DeepSeek V4 Pro passed A product announcement with five rules; Argue both sides of free buses.

  • Against Claude Opus 5.5: DeepSeek V4 Pro answered 2.1× sooner (3.9 s to 8.2 s), cost 17.2× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Argue both sides of free buses.

  • Against Claude Sonnet 5.5: DeepSeek V4 Pro answered 1.7× later (3.9 s to 2.3 s), cost 3.2× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Argue both sides of free buses.

  • Against Claude Sonnet 5: DeepSeek V4 Pro answered 1.1× sooner (3.9 s to 4.5 s), cost 3.2× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Announce a second bakery shop on LinkedIn.

  • Against Claude Haiku 5.5: DeepSeek V4 Pro answered 2.6× later (3.9 s to 1.5 s), cost 6.7× more, and passed 5 of 5 to its 3. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words; Argue both sides of free buses.

  • Against Claude Haiku 4.5: DeepSeek V4 Pro answered 1.8× later (3.9 s to 2.2 s), cost within 10% of it, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed A product announcement with five rules.

  • Against GPT-6 Astra: DeepSeek V4 Pro answered 1.2× sooner (3.9 s to 4.8 s), cost 10.4× less, and passed 5 of 5 to its 5.

  • Against GPT-6.1 Sol: DeepSeek V4 Pro answered 1.2× later (3.9 s to 3.3 s), cost within 10% of it, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words.

  • Against GPT-6 Sol: DeepSeek V4 Pro answered 1.1× later (3.9 s to 3.5 s), cost 2.1× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words.

  • Against GPT-6 Luna: DeepSeek V4 Pro answered 1.6× later (3.9 s to 2.4 s), cost 10.8× more, and passed 5 of 5 to its 5.

  • Against Gemini 4 Argon: DeepSeek V4 Pro answered 1.5× sooner (3.9 s to 5.7 s), cost 1.3× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Argue both sides of free buses.

  • Against Gemini 3.1 Pro (preview): DeepSeek V4 Pro answered 2.3× sooner (3.9 s to 9.1 s), cost 8.6× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Argue both sides of free buses.

  • Against Gemini 3.8 Flash: DeepSeek V4 Pro answered within 10% of its time (3.9 s to 3.9 s), cost 1.2× less, and passed 5 of 5 to its 3. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words; A product announcement with five rules.

  • Against DeepSeek V4.1 Flash: DeepSeek V4 Pro answered 2.9× later (3.9 s to 1.3 s), cost 2.4× more, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words.

  • Against Grok 4.7: DeepSeek V4 Pro answered 1.5× sooner (3.9 s to 6.0 s), cost 5.4× less, and passed 5 of 5 to its 3. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words; Argue both sides of free buses.

  • Against Kimi K3: DeepSeek V4 Pro answered 1.3× later (3.9 s to 3.0 s), cost 3.6× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed A product announcement with five rules.

  • Against GLM 5.3: DeepSeek V4 Pro answered 2.2× later (3.9 s to 1.8 s), cost 1.9× more, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed Argue both sides of free buses.

  • Against GLM 5.3 Flash: DeepSeek V4 Pro answered 2.2× sooner (3.9 s to 8.5 s), cost 3.3× more, and passed 5 of 5 to its 2. Only DeepSeek V4 Pro passed Rewrite corporate jargon in plain words; A product announcement with five rules; Argue both sides of free buses.

  • Against Mistral Large 4: DeepSeek V4 Pro answered 2.5× sooner (3.9 s to 9.8 s), cost 2.2× less, and passed 5 of 5 to its 4. Only DeepSeek V4 Pro passed A product announcement with five rules.

How these runs were done

Rubric (graded). The grader model scores the reply from 1 to 5 on each published criterion. It passes with an average of 4 or more and no criterion under 3, and only if it also meets the prompt's automatic rules (length, words it must or mustn't use).

The grader is Claude Opus 5.5 at low reasoning effort; its own replies are graded by GPT-6 Astra, so no model grades itself. Its prompt and every rubric are on the methods page.

How the runs were done, and every writing prompt.

DeepSeek V4 Pro, and writing, elsewhere

Questions

Is DeepSeek V4 Pro good for writing?

In our test runs it passed 5 of 5 writing prompts, tied 1st of the 20 models in llmwise. Every reply is on this page, so you can judge them yourself.

How many of my messages does a writing reply from DeepSeek V4 Pro use?

1 message each on Pro, where it has 250 a month on Pro. The price of a message is fixed and shown before you send it, however long the reply.

How were these runs done?

The same way for every model: each prompt sent through llmwise's own pipeline, each reply checked the same way. The methods page has every prompt and how each is scored.

Claude, GPT, Gemini, DeepSeek, Grok, Kimi, GLM, and Mistral, in one chat.

See what a message costs before you send it. Free is 5 messages to try; sign in with an email link, no password or card.

Start freeSee pricing