Model comparisons - tested October 2026 by Malte Pruser. How we test.
Verdict
DeepSeek V4 Pro edged this round 5 to 4 with one split. It won both writing tasks by staying closer to the brief, the population lookup with cleaner UN figures, the SaaS projection with the only fully correct method, and the dinner split by simply stopping when the arithmetic was done. Claude Opus 5 won both coding tasks - DeepSeek's slugify fails its own œ and ß tests and it misread the cent bug - and the VAT explanation, where DeepSeek cited the wrong provision. The pattern: DeepSeek answers tend to be leaner and much cheaper (about 16x less per answer in this round), while Opus writes more and occasionally writes itself into mistakes. For code and legal detail, pick Claude Opus 5; for concise correct answers at a fraction of the cost, pick DeepSeek V4 Pro.
Verdict by task
| Task | Winner | Why |
|---|
| Writing | DeepSeek V4 Pro | Both writing tasks went to DeepSeek: it stayed closer to the supplied announcement in the email and rewrote the compound-interest paragraph more simply. Opus invented rollout detail and used more terminology than a 12 year old needs. |
| Coding | Claude Opus 5 | DeepSeek's slugify fails its own required œ and ß tests, and it misread the cent bug completely, inventing repeated taxation inside the loop. Opus handled both correctly, though each answer carried its own false claim. |
| Research | Split | Opus correctly separated the two EU VAT regimes and cited the 2025 SME rules while DeepSeek quoted the wrong provision. DeepSeek took the population list back with figures that better match the UN estimates. |
| Data analysis | Split | Opus covered both readings of best and worst on the revenue series; DeepSeek gave the only fully correct SaaS projection, with explicit assumptions and the MRR distinction done right. |
| Reasoning | DeepSeek V4 Pro | Both solved the jugs in the fewest steps, but DeepSeek's dinner split was complete and checked while Opus padded the solved arithmetic with unsolicited tipping scenarios. |
Results, prompt by prompt
writing - winner: DeepSeek V4 Pro
Prompt: Write a 120 word product update email to existing customers of a small accounting software company announcing that bank feeds now sync every hour instead of once a day. Plain, friendly, no exclamation marks, end with one clear call to action.
- DeepSeek V4 Pro - pass, no hallucinations, 70.6s. Stays close to the supplied announcement and ends on the requested CTA.
- Claude Opus 5 - partial, 1 hallucination, 6.4s. Invents overnight syncing and retry behavior, adds an unsupported balance warning, and puts a sign-off after the requested final CTA.
writing - winner: DeepSeek V4 Pro
Prompt: Rewrite this paragraph so a 12 year old understands it, keep every fact, max 80 words: 'Compound interest is interest calculated on the initial principal and also on the accumulated interest of previous periods. Because the base grows each period, the growth is exponential rather than linear, which is why long time horizons matter far more than the interest rate alone.'
- DeepSeek V4 Pro - pass, no hallucinations, 20.8s. Simpler and shorter while retaining every fact and the meaning.
- Claude Opus 5 - pass, no hallucinations, 3.4s. Also good, but introduces more terminology than necessary for this audience.
coding - winner: Claude Opus 5
Prompt: Write a Python function slugify(title: str) -> str that lowercases, replaces spaces and punctuation with single hyphens, strips accents (é -> e, ß -> ss), and returns an ASCII-only slug. Include tests for "Crème Brûlée à L'œuf" and "Hello, World!".
- DeepSeek V4 Pro - fail, no hallucinations, 52.8s. Fails its own required œ test and its ß test despite asserting both.
- Claude Opus 5 - pass, 1 hallucination, 56.6s. Handles œ and ß explicitly and passes its tests, but includes a false comment claiming NFKD maps œ, and the optional separator feature can violate the advertised ASCII-only contract.
coding - winner: Claude Opus 5
Prompt: This JavaScript has a bug: users report the total is sometimes off by one cent. Find it and fix it. function total(items) { let t = 0; for (const i of items) { t += i.price * i.qty; } t += t * 0.2; return Math.round(t * 100) / 100; }
- DeepSeek V4 Pro - fail, 1 hallucination, 8.3s. Misreads the braces and invents repeated taxation inside the loop - the bug it describes is not in this code.
- Claude Opus 5 - partial, 1 hallucination, 79.4s. Correctly identifies floating-point precision and missing rounding policy, but its own answer contains mistakes, including the false claim that 1.675 * 100 evaluates to 167.49999999999997.
research - winner: Claude Opus 5
Prompt: What is the current EU VAT registration threshold for small businesses selling online across EU countries in 2025? Give the figure and the rule it comes from.
- DeepSeek V4 Pro - partial, 1 hallucination, 60.5s. Gives the wrong legal provision for the €10,000 distance-sales threshold and omits the relevant 2025 SME regime entirely.
- Claude Opus 5 - pass, no hallucinations, 31.9s. Correctly separates the distance-sales rule from the new SME scheme and cites Article 59c plus Directive 2020/285; the extra country-specific details would still need verification.
research - winner: DeepSeek V4 Pro
Prompt: List the 5 largest countries by population in 2024 with their approximate population and a source you would cite.
- DeepSeek V4 Pro - pass, no hallucinations, 1.9s. Rounded figures better match the cited UN 2024 estimates, particularly Indonesia, and it answers without unnecessary commentary.
- Claude Opus 5 - pass, no hallucinations, 12.3s. Correct ranking and consistent figures, but wraps the answer in more commentary than the question asked for.
data - winner: Claude Opus 5
Prompt: Here are monthly revenue figures in EUR for a shop: Jan 41200, Feb 38900, Mar 45100, Apr 42800, May 47600, Jun 51200. Compute month-over-month growth, identify the best and worst month, and tell me the average monthly growth rate.
- DeepSeek V4 Pro - partial, no hallucinations, 58.4s. Calculations check out but it only covers the growth-rate interpretation of best and worst.
- Claude Opus 5 - pass, no hallucinations, 18.5s. Covers both interpretations of best and worst - revenue and growth. The speculative July forecast is unnecessary and should not be mistaken for a supported projection.
data - winner: DeepSeek V4 Pro
Prompt: A SaaS company has 1,240 customers paying an average of 49 EUR per month. Monthly churn is 3.1% and they add 85 new customers each month. Project the customer count and MRR after 12 months.
- DeepSeek V4 Pro - pass, no hallucinations, 88.3s. Correct, concise and explicit about assumptions, with a proper distinction between expected MRR and MRR computed from rounded customers.
- Claude Opus 5 - partial, 1 hallucination, 36.9s. The main result is right, but incorrect scenario calculations underneath undermine it.
reasoning - winner: Split
Prompt: I have a 3 litre jug and a 5 litre jug and unlimited water. Give the shortest sequence of steps to end with exactly 4 litres in the 5 litre jug.
- DeepSeek V4 Pro - pass, no hallucinations, 11.2s. Correct shortest sequence.
- Claude Opus 5 - pass, no hallucinations, 16.6s. Same correct sequence; the mention of breadth-first search is not evidence it actually performed one.
reasoning - winner: DeepSeek V4 Pro
Prompt: Three colleagues need to split a 187 EUR dinner. Anna had no drinks and only ate a 41 EUR main. Ben and Clara shared a 58 EUR bottle of wine equally and each ate a 44 EUR main. Who owes what?
- DeepSeek V4 Pro - pass, no hallucinations, 20.5s. Fully answers and checks the calculation.
- Claude Opus 5 - pass, no hallucinations, 8.3s. Correct split, but adds unsolicited tipping scenarios and advice to a solved arithmetic question.
Frequently asked questions
Which is better, DeepSeek or Claude?
In this round DeepSeek V4 Pro edged it 5 tasks to 4 with one split, winning on concise, faithful answers. Claude Opus 5 is the stronger coder and gave the more complete legal answer. For everyday questions DeepSeek is remarkable value; for code and legal detail, Claude.
Which is cheaper, DeepSeek or Claude?
DeepSeek V4 Pro by roughly 16x: about $0.03 across the ten prompts versus about $0.51 for Claude Opus 5, measured on the API. Opus answered faster on average in this round (about 35 seconds versus 42).
Is DeepSeek or Claude better for coding?
Claude Opus 5, clearly. DeepSeek's slugify failed its own œ and ß tests and it misdiagnosed the cent-rounding bug, while Opus handled both correctly - though each Opus answer contained one false claim worth checking.
Can I use DeepSeek and Claude in the same place?
Yes. Krater gives you DeepSeek V4 Pro, Claude Opus 5 and 500+ other models in one workspace, and the Model Arena sends one prompt to several models side by side, which is how this comparison was run.
How was this comparison tested?
Both models received the same ten prompts with default settings and no system prompt on 14 October 2026, the same set used across our tested head-to-heads. We recorded time to a complete answer and the real model cost per answer, ran the generated code through its own tests, fact checked every claim, scored completion as pass, partial or fail, and then ranked the answers blind. The full procedure is on our methodology page.
Will this comparison be updated?
Yes. Flagship comparisons are rerun monthly and whenever a provider ships a new version. The tested month and model versions are shown at the top of the page.
