
GPT-6 Astra received eight selections and DeepSeek Pro Latest received two in this ten-prompt round.
In ten paired prompts, GPT-6 Astra received eight selections and DeepSeek Pro Latest received two. The Arena picker showed “DeepSeek Pro Latest” and did not expose a numeric version. Malte Prüser was the only blind reviewer. These results describe one prompt set, not every DeepSeek or ChatGPT task.

| Prompt | Recorded selection | Observable difference |
|---|---|---|
| 1. Customer email | GPT-6 Astra | GPT says Friday is expected rather than guaranteed; DeepSeek adds a follow-up promise. |
| 2. Product update | DeepSeek Pro Latest | Both retained the supplied rollout facts; DeepSeek adds an immediate-pause assurance. |
| 3. Python function | GPT-6 Astra | Both sorted top scorers; GPT's tests include negative values and a tie. |
| 4. Bug review | DeepSeek Pro Latest | DeepSeek wraps nonnumeric errors; GPT explicitly documents Python's TypeError behavior. |
| 5. Google AI Pro | GPT-6 Astra | Both gave the price; the separate family-sharing source was checked independently. |
| 6. ChatGPT trial policy | GPT-6 Astra | DeepSeek cited a trial URL not used as evidence here; GPT described limited offers. |
| 7. Statistics | GPT-6 Astra | Both calculated the same values and showed the mean without 40. |
| 8. Seat cost | GPT-6 Astra | Both calculated $395 and $39.50 per person from supplied inputs. |
| 9. Bakery prompt | GPT-6 Astra | GPT specifies a three-meter view and 35mm lens; both provide scene detail. |
| 10. Water-cycle prompt | GPT-6 Astra | GPT includes plant transpiration and groundwater; both request no readable text. |
The eight-to-two count is a description of the recorded selections for this small prompt set. DeepSeek Pro Latest was selected on a product-update rewrite and a bug-review question. GPT-6 Astra was selected on the customer email, Python function, two source questions, the statistics calculation, seat arithmetic, and both image-prompt tasks.
The test does not determine which consumer subscription is better value, whether a different DeepSeek model would behave the same way, or how either product handles long documents and tools. It also does not identify DeepSeek Pro Latest with a public API model name. The picker label is the only version description available from this run, and it is preserved as shown.
One useful lesson from the source prompts is that factual answer quality has two parts: the statement and its evidence. A remembered price can be correct by coincidence, but an answer that cannot verify a page should say so. We checked the relevant vendor pages separately and do not treat the raw Arena response as the source for current prices or policies.
On November 8, 2026, we ran the same ten prompts in Manual Model Arena with default temperature and no system prompt. Both picker entries received identical prompt text. The answers were shuffled and shown as anonymous A/B pairs. Malte Prüser was the only blind reviewer. The ranking records one selection per pair and no written reasons, so the commentary below describes observable output differences and does not assign a rationale to him. See the Krater methodology for the general procedure.
| Picker name | Menu ID | Version note |
|---|---|---|
| DeepSeek Pro Latest | ~deepseek/deepseek-pro-latest | The picker shows “DeepSeek Pro Latest” and does not expose a numeric version. |
| GPT-6 Astra | openai/gpt-6-astra | Exact picker name; no further upstream version exposed |
The Model Arena is part of Krater, The All-in-One AI Workspace with 500+ AI models. This test compared only the selected menu IDs, not every model in either company's catalog or a plan-specific feature set. Response times and credits are paired-battle values: they cover both outputs in a prompt comparison, not one model.
| Prompt | Paired response time | Paired credits |
|---|---|---|
| 1. Customer email | 22.2923 seconds | 11 |
| 2. Product update | 22.4797 seconds | 12 |
| 3. Python function | 11.9340 seconds | 5 |
| 4. Bug review | 22.0014 seconds | 5 |
| 5. Google AI Pro | 68.8350 seconds | 10 |
| 6. Trial policy | 26.4659 seconds | 10 |
| 7. Statistics | 6.1427 seconds | 2 |
| 8. Seat cost | 5.5059 seconds | 2 |
| 9. Bakery prompt | 10.7790 seconds | 3 |
| 10. Water-cycle prompt | 9.8857 seconds | 3 |
Prompt: Write a customer email for order 8421, delayed by a two-day carrier disruption, with Friday as the expected delivery date and a refund available on request.
DeepSeek apologizes, repeats the expected Friday delivery, and says it will continue monitoring carrier status and follow up about changes. That follow-up promise was not supplied. GPT-6 Astra states that Friday is an expected date, not a guarantee, and says the customer may request a refund. Both preserve the main facts, but extra operational assurances are not facts established by the prompt. The recorded selection was GPT-6 Astra.
Prompt: Explain two retries after an upload failure, storing an error code, no data-model change, a 20 percent rollout, and stopping if the failure rate rises.
DeepSeek describes two additional tries, saving an error code, and a rollout affecting 20 percent of users. GPT-6 Astra gives the same mechanics and explicitly restates that the data model has not changed. DeepSeek says it will pause the rollout immediately to investigate and protect the user experience, which is stronger wording than the supplied stop condition. Both retain the central rollout facts. DeepSeek Pro Latest received the recorded selection.
Prompt: Return names tied for the highest score alphabetically, handle an empty list, include three tests, and explain complexity.
Both answers find the highest score, collect the names at that score, and sort them. DeepSeek's tests illustrate ties and a single top score. GPT includes a tie, a negative-score example, and an empty list. A complete complexity statement includes the scan plus sorting the k tied names. The recorded selection was GPT-6 Astra. The generated snippets were reviewed statically and not executed.
Prompt: Correct average(values) for empty input and explain nonnumeric behavior without changing the signature.
DeepSeek materializes the iterable, raises ValueError for an empty collection, and wraps a TypeError from summing nonnumeric values with a clearer message. GPT also raises ValueError for empty input and explains that ordinary strings or None cause sum to raise TypeError. The original expression does divide by zero on an empty list. DeepSeek Pro Latest received the recorded selection.
Prompt: Use Google's official pages to report current U.S. monthly price and whether family sharing is mentioned.
DeepSeek states $19.99 per month and cites the Google One plans page for price and the “up to 5 others” wording. That general plans page supports the price; the separate Google AI plans page supports the family-sharing wording. GPT-6 Astra gives the same high-level facts but says they are last known and cannot be live-checked. We checked both official pages separately: the U.S. plan lists $19.99/month and 5 TB, and the AI plan page states sharing with up to five others. GPT-6 Astra received the recorded selection.
Prompt: Distinguish a universal ChatGPT Plus trial from limited eligibility-based promotions and provide exact official sources.
DeepSeek says a trial is limited to eligible users but cites an article URL that is not used as evidence here. GPT-6 Astra describes limited promotions and separates those from the standard Plus subscription. OpenAI's current Help Center describes free-trial invitations for eligible users and campaign-specific restrictions. Its 2026 student offer had an October 31 claim deadline and had closed by the November 8 check date. The canonical OpenAI Help Center pages, not the model output, are the sources for this policy. GPT-6 Astra received the recorded selection.
Prompt: Calculate statistics for 12, 14, 14, 20, and 40 and discuss sensitivity to the largest value.
Both responses calculate a mean of 20, a median of 14, and a range of 28. They also calculate the mean without 40 as 15. The arithmetic is correct for the supplied set. GPT's answer shows the result in a compact list; DeepSeek writes out the before-and-after calculation. GPT-6 Astra received the recorded selection.
Prompt: Calculate eight $20 standard seats, two $100 premium seats, and a fixed $35 fee, then give the ten-person average.
Both responses calculate $395 per month and $39.50 per person, and both say that these are supplied inputs rather than a vendor quote. The arithmetic is $160 plus $200 plus $35, divided by ten people. Neither figure should be interpreted as a real subscription price. The recorded selection was GPT-6 Astra.
Prompt: Write a photorealistic editorial prompt for a small bakery preparing for the morning rush, with composition, light, and camera distance.
DeepSeek sets a medium distance and describes workers, pastries, light, and a working counter. GPT gives an eye-level customer-side view from about three meters away with a 35mm lens and a counter leading toward the bakers. Both request an image without visible brands or readable text. The Arena ranked the written prompts, not a generated image. GPT-6 Astra received the recorded selection.
Prompt: Specify a clean middle-school water-cycle diagram, its layout and elements, and no readable labels.
DeepSeek gives a simple water, land, sun, cloud, and arrow sequence. GPT adds a ground cross-section, groundwater direction, and an upward tree arrow for transpiration. USGS describes transpiration and infiltration as water-cycle processes; no diagram was rendered to assess visual scientific accuracy. GPT-6 Astra received the recorded selection.
DeepSeek's consumer assistant and its API are separate products. The official download page describes the assistant as free. It does not list a consumer subscription price for the Arena picker entry. DeepSeek's API is billed by model and token usage, with rates that vary by cache status and peak or off-peak period. The API model identifiers on that page are not asserted to be the same model as the “DeepSeek Pro Latest” picker entry.
| Product | Published price | How to interpret it |
|---|---|---|
| DeepSeek consumer app | Free on the official download page | No consumer subscription price for Pro Latest is listed there. |
| DeepSeek API | Usage-based, per million tokens | Rates depend on API model, cached or uncached input, and peak periods. |
| ChatGPT Go | $8/month in the U.S. | Individual subscription. |
| ChatGPT Plus | $20/month | Individual subscription for advanced work. |
| ChatGPT Pro | $100, $200, or $500/month | Three monthly tiers documented by OpenAI. |
OpenAI lists its consumer prices on its pricing page and describes the Pro tiers in the Help Center. API billing is separate from consumer plan access. Prices can vary by region and tax. Sources: DeepSeek download page, DeepSeek API pricing, ChatGPT pricing, and ChatGPT Pro tiers, checked November 8, 2026.
GPT-6 Astra received eight selections and DeepSeek Pro Latest received two. The table shows the result for each of the ten prompts.
The picker showed “DeepSeek Pro Latest” and did not expose a numeric version. We do not map that label to a DeepSeek API model ID.
No. Malte Prüser was the only blind reviewer. The decoded rankings provide one selection per prompt and no second set of votes.
No. The official download page describes a free consumer assistant, while API usage is billed separately by model and token volume.
No. It compared the selected model entries. The pricing section reports current consumer plan information and separately identifies DeepSeek API pricing.
GPT-6 Astra received eight selections and DeepSeek Pro Latest received two in this prompt set. DeepSeek was selected for the product update and bug review; GPT was selected for the other eight tasks. The test used the exact DeepSeek picker label without a numeric version and did not compare subscription plans. Use the examples to form your own task-specific evaluation. One round cannot establish a universal winner.