Perplexity Sonar Pro Search vs ChatGPT: A 5-5 Split by Task
The ten-prompt comparison split 5-5, with the selections divided across task types.
Perplexity Sonar Pro Search and GPT-6 Astra split ten matched prompts five selections each. The pattern was task-specific: GPT-6 Astra was selected for writing, code, and image-prompt tasks, while Sonar was selected for the bug review, source questions, and numerical tasks. Malte Prüser was the only blind reviewer.
Prompt-by-prompt results
| Prompt | Recorded selection | Observable difference |
|---|---|---|
| 1. Customer email | GPT-6 Astra | GPT explicitly distinguished an expected date from a guarantee. |
| 2. Product update | GPT-6 Astra | Sonar added “no action is required”; GPT kept closer to the supplied note. |
| 3. Python function | GPT-6 Astra | Both sorted tied names; GPT included a negative-score example. |
| 4. Bug review | Sonar Pro Search | Sonar gave explicit empty-input and nonnumeric-value errors. |
| 5. Google AI Pro | Sonar Pro Search | Sonar quoted Google plan wording; GPT said it could not live-check. |
| 6. ChatGPT trial policy | Sonar Pro Search | Sonar supplied the canonical Help Center URL for eligible invitations. |
| 7. Statistics | Sonar Pro Search | Both calculated the values; Sonar showed the mean without 40 as 15. |
| 8. Seat cost | Sonar Pro Search | Both calculated $395 and $39.50 per person and flagged the inputs. |
| 9. Bakery image prompt | GPT-6 Astra | GPT gave a three-meter viewpoint and 35mm lens; both were detailed. |
| 10. Water-cycle prompt | GPT-6 Astra | Both included a clear cycle; GPT specified transpiration and groundwater flow. |
How to read the split
This is a tie by count, with selections divided across task types. GPT-6 Astra received the picks on two writing prompts, the function-writing prompt, and both image-prompt prompts. Sonar Pro Search received the picks on the bug review, both source-based questions, the statistics calculation, and the supplied-input seat arithmetic. The test does not produce a universal ranking, and the per-prompt results are more informative than forcing a tie into a single recommendation.
A tie across selected tasks cannot answer a renewal question; whether Perplexity Pro is worth its annual price estimates whether Perplexity Pro's annual rate fits.
That division is not a claim about all writing, coding, research, or math tasks. There were only ten prompts, each task was short, and the comparison used one exact picker entry for each product. It did not compare Perplexity's entire search experience with every ChatGPT feature, nor did it test long research reports or team workflows. A company considering either tool should reproduce the task it cares about with its own source material and acceptance criteria.
Source prompts require special care. Sonar's answer to the Google question contained a current-looking quote, while GPT explicitly disclosed that it could not check live pages. We checked the current Google plan pages separately and cite them below. Likewise, the OpenAI policy answer is verified against OpenAI's Help Center, not accepted simply because a model included a URL.
How we tested
We used Manual Model Arena on November 8, 2026, sending the same ten prompts to each model with default temperature and no system prompt. The blind sheet randomized answer labels. Malte Prüser was the only blind reviewer. The decoded ranking records one choice per prompt and does not include a written rationale, so the output observations below should not be attributed to him as reasons. Read the Krater methodology for the general process.
| Picker name | Menu ID | Version note |
|---|---|---|
| Sonar Pro Search | perplexity/sonar-pro-search | Exact picker name; no numeric version exposed |
| GPT-6 Astra | openai/gpt-6-astra | Exact picker name; no further upstream version exposed |
The Model Arena is one tool inside Krater, The All-in-One AI Workspace with 500+ AI models. The test used only these two picker entries. It was not a subscription comparison, search-engine benchmark, or audit of every answer. The table below shows paired-battle response time and credits. No per-model timing or credit split was recorded.
| Prompt | Paired response time | Paired credits |
|---|---|---|
| 1. Customer email | 14.6416 seconds | 11 |
| 2. Product update | 16.3660 seconds | 11 |
| 3. Python function | 10.8292 seconds | 9 |
| 4. Bug review | 9.6534 seconds | 7 |
| 5. Google AI Pro | 20.9621 seconds | 11 |
| 6. Trial policy | 21.4920 seconds | 13 |
| 7. Statistics | 7.4295 seconds | 6 |
| 8. Seat cost | 6.0307 seconds | 6 |
| 9. Bakery prompt | 9.8110 seconds | 7 |
| 10. Water-cycle prompt | 9.9170 seconds | 7 |
Task 1: Customer email
Prompt: Draft an email about order 8421, delayed two days by a carrier disruption, with Friday as the new expected date and a refund option.
Sonar gives the email a subject line, states the delay, and offers a refund request by reply. GPT adds that Friday is an expected date rather than a guaranteed arrival. Both retain the supplied cause and timing. The prompt does not confirm what a support team will do after a request, so the drafts are examples rather than live refund policy. GPT-6 Astra received the recorded selection.
Task 2: Product update
Prompt: Explain two retries for failed uploads, storing an error code, an unchanged data model, a 20 percent rollout, and a stop condition if failures rise.
Sonar explains the retry process in nontechnical language, but adds “No action is required from users,” which was not in the supplied note. GPT describes the background sync process, its two retries, and the unchanged data model. Both mention the 20 percent rollout and stopping condition. The saved selection was GPT-6 Astra. We do not treat any additional operational instruction as a fact about a real product.
Task 3: Python function and tests
Prompt: Return all highest-scoring names alphabetically, handle empty input, include three tests, and explain complexity.
Both functions locate the maximum score and sort the names tied at that score. Sonar's examples include a tie and a single winner. GPT includes a tie, a negative-score case, and empty input. Their stated runtime descriptions account for scanning the records and sorting the tied names. GPT's answer also states assumptions about strings and comparable numeric scores. The recorded selection was GPT-6 Astra. The code was reviewed statically, not executed.
Task 4: Bug review
Prompt: Fix empty input in average(values) and state what happens with nonnumeric values.
Sonar materializes the iterable, raises ValueError for an empty list, and raises a clearer TypeError if a value is nonnumeric. GPT raises ValueError when the collection is empty and leaves nonnumeric handling to Python's sum. For ordinary lists, the original divides by zero on empty input and raises TypeError when adding a string to numbers. The recorded selection was Sonar Pro Search.
Task 5: Google AI Pro price and sharing
Prompt: Give Google's current U.S. monthly price, whether the plan page says family sharing is available, and exact sources.
Sonar says the U.S. page lists $19.99 per month and quotes family sharing with up to five others. GPT gives the same high-level information but says it cannot verify live wording. On November 8, Google's U.S. plan page lists AI Pro at $19.99 per month with 5 TB; the AI plans page states family sharing with up to five others. We checked those pages separately. Sonar Pro Search received the recorded selection.
Task 6: ChatGPT Plus trial policy
Prompt: Explain whether OpenAI describes a universal trial or only limited promotions, using official URLs and no unsupported eligibility assumptions.
Sonar cites OpenAI's current promotional subscriptions and free-trial invitations FAQ and describes campaign-level eligibility. GPT gives the limited-promotion conclusion and separates it from the regular paid Plus plan. OpenAI's Help Center independently supports the limited, eligible-invitation framing. The separate 2026 student offer had a claim deadline of October 31, 2026, so it was no longer open on the check date. The recorded selection was Sonar Pro Search.
Task 7: Mean, median, and range
Prompt: Calculate the mean, median, and range for 12, 14, 14, 20, and 40.
Both give mean 20, median 14, and range 28. Sonar says removing 40 makes the mean 15, and describes the median as less affected by a large value. GPT also shows the without-40 calculation. The arithmetic is verified directly from the supplied values. Sonar Pro Search received the recorded selection.
Task 8: Team seat cost
Prompt: Calculate eight $20 standard seats, two $100 premium seats, and a fixed $35 fee, then give the average for ten people.
Both outputs calculate $160 plus $200 plus $35, for a total of $395 and $39.50 per person. Each flags that the numbers are supplied inputs, not a vendor quote. Sonar shows a short arithmetic sequence; GPT gives the monthly subtotals and average. The recorded selection was Sonar Pro Search.
Task 9: Bakery image prompt
Prompt: Write a photorealistic editorial bakery prompt with composition, light, camera distance, and no brand names or readable text.
Sonar uses a medium-wide horizontal frame from about three meters away and adds a negative-space note for an editorial layout. GPT also specifies a three-meter camera distance and 35mm lens, with the counter as a leading line. Both describe a small bakery without brands. The Arena did not generate an image from the prompts. GPT-6 Astra received the recorded selection.
Task 10: Water-cycle diagram prompt
Prompt: Write a clean middle-school water-cycle diagram prompt with a specified layout and no readable labels.
Sonar includes evaporation, transpiration, condensation, precipitation, runoff, and infiltration or percolation. GPT gives a left-to-right land-and-water layout and adds groundwater moving toward a lake. USGS supports the water-cycle terms used in the prompts, including transpiration and infiltration. The task evaluated text prompts only, not the output of an image model. GPT-6 Astra received the recorded selection.
Perplexity and ChatGPT plans
Current plan information was checked on November 8, 2026. Perplexity plans describe access to search, models, research, and Computer usage. ChatGPT plans describe a separate set of models and limits. These subscription pages do not identify the exact picker entry used in the Arena and do not explain the 5-5 split.
| Product | Published individual price | Plan detail |
|---|---|---|
| Perplexity Pro | $20/month or $200/year (about $17/month billed annually) | Paid consumer plan with expanded research and model access. |
| Perplexity Max | $200/month or $2,000/year (about $167/month billed annually) | Highest consumer tier with higher access and monthly Computer credits. |
| ChatGPT Go | $8/month in the U.S. | Individual plan with expanded access compared with Free. |
| ChatGPT Plus | $20/month | Individual plan with advanced reasoning and expanded features. |
| ChatGPT Pro | $100, $200, or $500/month | Three monthly usage tiers listed by OpenAI. |
Perplexity's Help Center lists Max at $200 monthly or $2,000 annually and describes Pro at $20 monthly or $200 annually. OpenAI's current pricing and Pro help pages show the ChatGPT plans. Check the vendor page for regional availability, tax, and the exact checkout price. Sources: Perplexity Pro, Perplexity Max, ChatGPT pricing, and ChatGPT Pro tiers.
Frequently Asked Questions
Did either model win overall?
No overall winner is presented. The recorded selection count was five each, and the choices divided across task types.
Which tasks did Sonar receive?
Sonar Pro Search received the selections on the bug review, Google pricing prompt, OpenAI policy prompt, statistics prompt, and seat-cost arithmetic.
Was this a two-reviewer test?
No. Malte Prüser was the only blind reviewer. The ranking contains one choice for each of the ten prompt pairs and no second reviewer's vote.
Does the result compare Perplexity Pro with ChatGPT Plus?
No. The Arena compared picker entries. The pricing section explains consumer plans but did not test plan-level usage, search interfaces, or checkout entitlements.
Are the source answers fully fact-checked?
The Google price and family-sharing statement and OpenAI's promotion policy were checked against official pages. The remaining output examples are shown as responses, not as general factual advice.
Related reads
Bottom line
The recorded selections split evenly. GPT-6 Astra received five, including both writing prompts, the Python prompt, and the two image-prompt tasks. Sonar received five, including the bug review, the source questions, and the numerical prompts. Review the task-level output and choose evaluation prompts that match your own work. One round cannot establish a universal winner.