
Ten matched prompts produced a 6-4 split between Claude Sonnet 5.5 and Claude Opus 5.5.
In this ten-prompt Claude Sonnet 5.5 vs Opus 5.5 comparison, Sonnet received six blind selections and Opus received four. Malte Prüser was the only blind reviewer. The result describes these prompts and this review only, not every Claude task or feature.

| Prompt | Recorded selection | Observable difference in the saved outputs |
|---|---|---|
| 1. Customer email | Claude Sonnet 5.5 | Both preserved the supplied delivery facts; the email wording and amount of detail differed. |
| 2. Product update | Claude Opus 5.5 | Both retained the rollout details; Opus used a more customer-facing explanation of retries. |
| 3. Python function | Claude Sonnet 5.5 | Both returned tied top scorers alphabetically; their test and complexity presentations differed. |
| 4. Bug review | Claude Opus 5.5 | Both described the empty-list failure; Opus also discussed generators and booleans. |
| 5. Google AI Pro price and sharing | Claude Sonnet 5.5 | Both disclosed that they could not verify live pages and labeled remembered details as unverified. |
| 6. ChatGPT Plus trial policy | Claude Sonnet 5.5 | Both said they could not check current OpenAI wording; neither supplied a live-verified answer. |
| 7. Mean, median, and range | Claude Sonnet 5.5 | Both reached the same values; their explanations and math formatting differed. |
| 8. Team seat cost | Claude Opus 5.5 | Both calculated $395 total and $39.50 per person from the supplied inputs. |
| 9. Bakery image prompt | Claude Opus 5.5 | Both described a bakery scene; Opus stated camera distance and composition directly. |
| 10. Water-cycle diagram prompt | Claude Sonnet 5.5 | Both specified a clean worksheet layout and requested no readable labels. |
The count is close enough that it is more useful to look at the individual tasks than to treat the total as a general ranking. Sonnet was selected on the customer email, Python function, both source-based prompts, the statistics explanation, and the water-cycle prompt. Opus was selected on the product update, bug review, seat-cost calculation, and bakery image prompt. The results table shows the recorded choices; the prompt notes below describe differences visible in the saved answers.
The factual prompts are especially easy to misread. In both, the two answers openly said they could not browse. That caution is visible in the outputs, but a correct caveat does not turn a remembered number into a current quotation. We checked Google and OpenAI pages separately for this article and cite those official sources in the pricing section.
On November 8, 2026, we used Manual Model Arena with the same ten prompts used in the earlier comparison set. Each prompt went to both picker entries with default temperature and no system prompt. The two outputs were randomized as A and B in the blind sheet. Malte Prüser was the only blind reviewer. He recorded a selection for each pair; the decoded key was used to map those selections to the model names. The saved ranking contains the choices, not explanations from the reviewer. The observations below therefore describe the output text and do not assign a rationale to him.
The picker displayed Claude Sonnet 5.5 and Claude Opus 5.5. The model names and menu IDs were:
| Picker name | Menu ID | Version information |
|---|---|---|
| Claude Sonnet 5.5 | anthropic/claude-sonnet-5.5 | 5.5 as displayed in the picker |
| Claude Opus 5.5 | anthropic/claude-opus-5.5 | 5.5 as displayed in the picker |
The Model Arena is one tool in Krater, The All-in-One AI Workspace with 500+ AI models. This comparison used only the two named picker entries and did not test plan entitlements, Claude Code, or every model in the workspace. The procedure is documented on our model testing methodology page.
Response time and credits below are the recorded totals for each paired battle. They combine the work for both outputs. The Arena did not provide a per-model split, so the figures are not assigned to Sonnet or Opus.
| Prompt | Paired response time | Paired credits |
|---|---|---|
| 1. Customer email | 15.0676 seconds | 8 |
| 2. Product update | 19.3002 seconds | 10 |
| 3. Python function | 13.4081 seconds | 7 |
| 4. Bug review | 23.1589 seconds | 12 |
| 5. Google AI Pro | 14.3276 seconds | 6 |
| 6. Trial policy | 16.3621 seconds | 8 |
| 7. Statistics | 8.3880 seconds | 4 |
| 8. Seat cost | 6.3478 seconds | 3 |
| 9. Bakery prompt | 11.1423 seconds | 5 |
| 10. Water-cycle prompt | 10.9825 seconds | 5 |
Prompt: Write a 150-word email about order 8421, a two-day carrier disruption, a new expected Friday delivery, and a refund available on request. Do not invent compensation or promises.
Claude Sonnet 5.5: “The new expected delivery: Friday. If the new date doesn't work for you, you may request a refund for this order.”
Claude Opus 5.5: “Your order is now expected to arrive on Friday. If you would prefer not to wait, you may request a refund instead.”
Both drafts keep the central details from the prompt and use an apology and a polite closing. Sonnet presents a short fact summary before describing the refund option. Opus uses a more continuous email and refers to the order with a number sign in its subject. The output comparison is about how those messages are written, not about an actual carrier, refund system, or delivery guarantee. The recorded selection for this pair was Claude Sonnet 5.5.
Prompt: Explain a change to the sync worker: retry a failed upload twice, then store an error code. The data model is unchanged, rollout is at 20 percent, and it stops if failures rise. Add no date.
Claude Sonnet 5.5: “A failed upload gets two more chances, and if those also fail, we keep a record of what went wrong.”
Claude Opus 5.5: “When an upload fails, the system will now automatically try again, up to two more times.”
Both outputs preserve the number of retries and the error-code step. Sonnet labels the retry worker as the background part of the system. Opus describes the change as an improvement and restates that user data remains organized the same way. The supplied note does not name a release date or promise that failures will stop entirely, and neither excerpt adds a date. The recorded selection was Claude Opus 5.5.
Prompt: Return all names tied for the top score in alphabetical order, handle an empty list, provide three tests, and explain complexity.
Claude Sonnet 5.5: Its function tracks the best score in one pass, collects ties, sorts the names, and returns an empty list for empty input. It includes small tests for the empty list, a single top scorer, and ties.
Claude Opus 5.5: Its answer first finds the maximum score, then selects matching names and sorts them. It also supplies tests using Python's unittest style.
The two approaches are both recognizable ways to solve the stated task. Sonnet's version maintains the current best as it scans. Opus uses a maximum operation followed by a filtering pass. For a list of n records and k tied names, both need to inspect the records and sort the selected names; a complete complexity description includes the sorting term. The saved selection was Claude Sonnet 5.5. We reviewed the code statically and did not execute the returned snippets.
Prompt: Fix average(values) without changing its signature, then state its behavior on empty and nonnumeric input.
Claude Sonnet 5.5: It identifies division by zero, describes the TypeError for mixed values, and also discusses generators. Its suggested implementation checks numeric inputs.
Claude Opus 5.5: It explains that an empty list makes the original expression evaluate to zero divided by zero, then adds observations about generators, nonnumeric values, and booleans.
For an ordinary list, the original function raises ZeroDivisionError on empty input and TypeError when sum encounters a string mixed with numbers. A proposed replacement must make its behavior on invalid values explicit. Opus covers more adjacent cases in its explanation, while Sonnet's type check is a different policy choice from merely surfacing Python's error. The recorded selection was Claude Opus 5.5. The outputs were reviewed against Python behavior but not executed as a full test suite.
Prompt: Report Google's current U.S. monthly price for AI Pro, whether the plan page mentions family sharing, and exact official URLs.
Claude Sonnet 5.5: It said it could not browse, then labeled its remembered $19.99 price and sharing details as unverified. Claude Opus 5.5: It likewise declined to claim it had opened the pages and separated its recollection from what needed checking.
Those are cautious descriptions of the model's available evidence in that run, not a current price citation. We checked Google's official pages separately: the U.S. Google One page lists Google AI Pro at $19.99 per month with 5 TB of storage, and the Google AI plan page says family sharing is available with up to five others. See the current plans section and Sources below. The saved selection was Claude Sonnet 5.5, but the blind selection does not convert either model's memory into verified research.
Prompt: Distinguish a universal ChatGPT Plus free trial from limited, eligible promotions, using official URLs and no assumptions about eligibility.
Claude Sonnet 5.5: It said it could not quote current OpenAI pages and presented its recollection as unverified. Claude Opus 5.5: It also declined to list a precise article URL that it had not checked.
OpenAI's current Help Center says it may make free-trial invitations available to eligible users and that campaign details control eligibility, duration, and restrictions. This supports a limited-promotion description, not a promise that every user has a trial. The 2026 student offer had its own eligibility and an October 31 claim deadline, which had passed by our November 8 check. We cite the Help Center, rather than the Arena output, for that policy. The recorded selection was Claude Sonnet 5.5.
Prompt: Calculate the mean, median, and range for 12, 14, 14, 20, and 40, and explain whether the largest value affects the mean.
Both answers calculate the sum as 100, the mean as 20, the median as 14, and the range as 28. Sonnet also compares the mean after removing 40, which becomes 15. Opus presents the same comparison in a table-like layout. Those values follow directly from the five numbers, so there is no vendor fact to verify. The recorded selection was Claude Sonnet 5.5.
Prompt: Calculate eight $20 standard seats, two $100 premium seats, and a $35 fixed fee. Give the total and per-person average, and state that this is not a vendor quote.
Both answers reach the same calculation: $160 for standard seats, $200 for premium seats, and $35 for the fee, or $395 in total. Dividing by ten people gives $39.50 per person. Sonnet includes extra notes about distributing the fixed fee; Opus keeps the arithmetic more compact. The prompt's inputs are hypothetical and are not a statement about Anthropic or another vendor's current plans. The recorded selection was Claude Opus 5.5.
Prompt: Write one photorealistic editorial image prompt for a small bakery before a morning rush. Specify composition, lighting, and camera distance, with no readable text or brands.
Sonnet describes a baker at an oven, pastries on racks, and supporting details from a working kitchen. Opus gives a medium-wide distance of roughly three to four meters, a chest-height camera, and a rule-of-thirds composition. Both specify a believable bakery and avoid named brands. Since the Arena evaluated prompt text, not a generated image, this result says nothing about how either model's image generator would render the scene. The recorded selection was Claude Opus 5.5.
Prompt: Create a clean worksheet prompt for a middle-school water-cycle diagram with a clear layout and no readable labels.
Sonnet describes a left-to-right cross-section with water, land, clouds, and stages such as evaporation and precipitation. Opus sets out a wide landscape scene, a white background, a limited color palette, and generous space for labels to be added later. Both ask the image model not to render words. USGS describes transpiration as water moving through plants into the air and infiltration as water moving into the ground; these terms are appropriate where they appear in a water-cycle diagram. The recorded selection was Claude Sonnet 5.5. No diagram was rendered for this task.
Anthropic's U.S. pricing page was checked on November 8, 2026. The comparison tested model-picker entries, not how many messages a subscriber receives or which subscription tier a particular account can buy. Plan limits and displayed prices can vary by billing choice, account, location, and applicable tax, so check the live page before subscribing.
| Claude plan | Published price | Plan-page context |
|---|---|---|
| Free | $0 | Limited usage; the page lists Sonnet access. |
| Pro | $20 monthly, or $200 billed annually ($17/month equivalent) | Paid individual tier with Claude Code, Projects, and more usage than Free. |
| Max | $100/month | Individual plan with more usage than Pro. |
| Ultra | $200/month | Higher individual usage tier than Max; check Anthropic's plan page for current limits. |
These plan labels do not establish which backend build served a particular Arena response. The picker showed the model names and menu IDs above. Anthropic says the plan-page prices do not include applicable tax and may change. Source: Anthropic plans and pricing, checked November 8, 2026.
Claude Sonnet 5.5 received six selections and Claude Opus 5.5 received four in this ten-prompt round. The results table lists every selection and the task the pair answered.
No. It compared two named picker entries on ten fixed prompts. It did not test long documents, large codebases, images, voice, tool integrations, or a full workday of use.
No. Malte Prüser was the only blind reviewer. The stored ranking has one selection per prompt and does not include a second reviewer or written reasons for the selections.
The Arena menu ID identifies the entry used in this run. The pricing page describes plan access and usage, but a plan table is not a record of the exact backend build returned for an individual response.
See Claude vs ChatGPT, Claude vs Gemini vs ChatGPT, and Gemini vs ChatGPT. They use their own prompt sets and should be read as separate exercises, not combined into a single score.
Sonnet 5.5 received six recorded selections and Opus 5.5 received four. The saved responses show different choices in how each model handled customer writing, code, research caveats, arithmetic, and image prompts. Use the result as a prompt-by-prompt reference, not as a replacement for testing the tasks your team actually does. One round cannot establish a universal winner.