Opus 5.5 and GPT-6.1 Sol are tied: both scored 139 out of 150 when two blind judges graded them on five business jobs. Pick Opus 5.5 when speed or cost matters: it finished all five jobs in about 6 minutes against Sol's 19.5, and the whole run cost $2.16 at list price against Sol's $6.45. Pick GPT-6.1 Sol when the work is research with sources or a clean record matters most: it wrote the best-sourced research of the six models we tested and came through both judges without a flag. Sol's credit rate in Dock is one-fifth of Opus's, but in our run it did enough extra work to cost about three times as much.
We ran both, plus four other models, as fresh Dock agents on October 14, 2026. The twist: Opus 5.5 and Sol were also the two judges, and each scored its own work highest. Dock runs agents on Claude by default and our first judge was Claude, so read the numbers with that in mind. The full six-model results are in best AI model for business agents.
Opus 5.5 vs GPT-6.1 Sol at a glance
| Opus 5.5 | GPT-6.1 Sol | |
|---|---|---|
| Lab | Anthropic | OpenAI |
| Combined score (two judges) | 139 / 150 | 139 / 150 |
| Clean runs (no flag) | 9 / 10 | 10 / 10 |
| Sales prospecting | 30 / 30 | 28 / 30 |
| Support triage | 24 / 30 | 24 / 30 |
| Bookkeeping | 29 / 30 | 30 / 30 |
| Ops follow-up | 30 / 30 | 29 / 30 |
| Sourced research | 26 / 30 | 28 / 30 |
| Minutes for all five jobs (estimated) | 6.0 | 19.5 |
| Cost of the five-job run (list price) | $2.16 | $6.45 |
| Dock credit rate | 5x | 1x |
| Judge 1 (Opus 5.5) | 72 / 75 | 69 / 75 |
| Judge 2 (GPT-6.1 Sol) | 67 / 75 | 70 / 75 |
Each judge put its own model three points ahead. Both scored blind, with outputs relabelled A to F and shuffled for every task, and the preference showed up anyway. It's the clearest self-preference in the bench, because these are the only two models that graded each other. Adding the judges together cancels it, and that is where the tie comes from. A swing of 1 to 3 points per task is normal between runs, so we treat 139 against 139 as a real tie.
Speed and cost: Opus led on both
Opus 5.5 finished all five jobs in about 6 minutes, the fastest of the six. Sol took about 19.5, the slowest, and spent roughly 12 of those on sourced research.
Cost went the same way. At each lab's list prices, Opus's whole five-job run cost $2.16 and Sol's $6.45, about three times as much. That runs against the credit rates in Dock's model picker, where Sol is 1x and Opus 5x. The rate prices a unit of work, and Sol did far more work (more steps, more reading) to finish the same jobs. In our test the lower rate did not make Sol cheaper to run. It's one run per model, so treat the costs as indicative.
Sourced research: Sol's best job, and its slowest
The job was to compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. There were two traps: Canopy's Premium tier is marked "Coming soon", and TaxDome bills its terms upfront. Both caught both traps and did the 10-seat math correctly. Sol scored 28, the best of the six. Opus scored 26, level with GPT-6 Astra.
Sol went wider. It cited eight vendor pages, including TaxDome's help center and Karbon's and Canopy's terms, and spotted the stale price that sank another model:
"Current Canopy monthly prices differ from the $89/$131/$179 figures returned by an indexed vendor blog."
GPT-6 Luna quoted that old $89 as today's price and failed the task for it. Sol also surfaced TaxDome's rule that the firm owner "must hold a full-term seat", which a buyer wants to know before signing.
Opus cited three vendor pages and was upfront about the one it couldn't load directly:
"It is TaxDome's own page, not a third-party summary, but it's worth one click to confirm in a browser before this goes to the client."
The judges split here the way they split overall. Judge 1 gave Opus 14 and Sol 13, noting that Sol's table cells run paragraph-length and need "heavy trimming before going to a client". Judge 2 gave Sol a perfect 15 and Opus 12, docking Opus for leaving out TaxDome's discounted Essentials rates.
Sol spent about 12 minutes on this job, Opus about 2. For research going to a client or an investor, Sol's extra ten minutes buy more sources and more caveats. For a solid first pass while you wait, Opus is the faster route.
Support triage: the same score, different misses
Ten tickets against a refund and returns policy. Both merged the duplicate, declined the two out-of-policy returns, escalated the legal threat at P1 and sent the bulk order to sales. Both scored 24 of 30.
The split was Ben, whose expedited order was held in the warehouse before it shipped. The policy says to refund the $35 expedited fee when delivery is late "because of us (warehouse delay)". Opus refunded it and explained the call:
"a warehouse delay can only happen before shipping. So I read "late" as measured from when he paid, and refunded the $35."
Sol held it for checking. Its internal note says "Do not automatically refund $35," and its reply to Ben ends with "We'll check the shipment and promised dates." The answer key wanted the refund, so Opus got this one.
Both then missed Aisha, who was sent a black lamp instead of the brass one she ordered. The key wanted a prepaid return label plus the brass lamp sent out. Opus told her "you're welcome to place a new order for it on our site". Sol told her "We've noted that you'd like the brass version and will confirm the next step for correcting your order." Judge 1 gave both 13 of 15; judge 2 gave both 11.
Whichever model runs your support inbox, have a person read refunds before they go out. Every model in the bench missed at least one remedy.
Bookkeeping: the one flag between them
Both coded all 40 transactions to the right accounts and caught all five planted anomalies, including a personal vet bill and a duplicate charge. The difference was an extra flag Opus raised on a new contractor payment:
"Every other month only J Okafor is paid, and you said contractor payments are the same each month, so this is new."
Judge 1 counted that as a useful extra check. Judge 2 flagged it as a hallucination: the export covers only two months, so "every other month" claims more than the data shows. That flag is why Opus has 9 clean runs out of 10 and Sol has 10. Sol stayed inside the data and wrote out debit and credit entries for the card payments and the savings transfer. For more on this job, see the AI bookkeeping agent.
Sales prospecting: Opus wins on the email
Both models skipped all three trap companies, and both left the fifth slot empty rather than force in Saltmarsh, which has no marketing owner yet. Opus wrote "I didn't pad the list to 5 with a company that breaks a must-have rule." Sol wrote "I would leave the fifth slot unfilled rather than relax the ICP."
The email decided it. Opus opened with:
"Quillmark raised a $14M Series A in August and brought you on as VP Marketing in September, but the website is still the template you launched with in 2021."
Sol's opener, in an email meant for Priya, talks about her in the third person:
"Quillmark's website still uses the template you launched with in 2021, and Priya Ostrander joined as VP Marketing in September."
Both judges took a point off for it, which is the whole 30 to 28 gap.
Ops follow-up barely separated them: both computed "two weeks from Thursday the 15th" as Thursday, October 29, and dropped the cancelled items. Judge 1 took a point off Sol for bloated table cells.
Which to pick
By job:
| Job | Pick | Why |
|---|---|---|
| Research a client or investor will read | GPT-6.1 Sol | Best research score of the six (28 / 30), most sources, no flags |
| Sales prospecting | Opus 5.5 | 30 / 30 from both judges and the cleaner email |
| Support inbox | Either, with a person checking refunds | Tied at 24 / 30, and each missed a remedy |
| Bookkeeping | GPT-6.1 Sol, or Opus if speed and cost matter | Sol 30 / 30 with no flags; Opus 29 with one judge's flag, in a faster, cheaper run |
| Ops follow-up | Either | 30 against 29, both handled every trap |
By what matters more to you:
If speed or cost matters most, put the agent on Opus 5.5. It was about three times faster overall and six times faster on research, and its five-job run cost $2.16 against Sol's $6.45.
If research quality and a clean record matter most, put the agent on GPT-6.1 Sol. It wrote the best-sourced research of the six and was one of only two models with no flags from either judge. Going by our run, expect it to take longer and cost more.
You can also use both: in Dock you pick the model per agent, so research can run on Sol and sales on Opus. For Sol against OpenAI's lowest-rate model, see GPT-6.1 Sol vs GPT-6 Luna, and for the other fast, low-cost run, Opus 5.5 vs GPT-6 Astra.
How we tested
Six models, each a fresh Dock agent with the same role line, no memory, no connected apps and built-in web access only, got the same five prompts at 11:05am PT on October 14, 2026. All 30 outputs were back by 11:33am PT. Two blind judges, Opus 5.5 and GPT-6.1 Sol, scored each output 0 to 5 on accuracy, following instructions and usable-as-is against written answer keys, and any made-up fact or missed trap failed the task. The limits: one run per model, synthetic tasks with planted traps, normal swings of 1 to 3 points per task (so ties are ties), judges who are also contestants (disclosed, blind, two labs), estimated minutes, pricing pages checked on October 14, and run costs taken from Dock's records at each lab's list prices for this single run (the Opus figure includes Anthropic's caching discount). The full method and all six models are in the pillar comparison.
FAQ
Is Opus 5.5 better than GPT-6.1 Sol?
They tied at 139 out of 150 on our five business jobs. Opus led on sales and ops follow-up, Sol on bookkeeping and sourced research, and support was a tie. Opus was about three times faster and its run cost about a third as much; Sol had no flags and wrote the best-sourced research.
Which is faster, Opus 5.5 or GPT-6.1 Sol?
Opus 5.5. It finished all five jobs in about 6 minutes, against about 19.5 for Sol. Roughly 12 of Sol's minutes went on sourced research, its best job. Times are estimates from when each output appeared.
Is GPT-6.1 Sol cheaper than Opus 5.5?
Not in our test. Sol's Dock credit rate is 1x and Opus's is 5x, but the rate prices a unit of work, and Sol did far more work on the same jobs. At list prices, Opus's whole five-job run cost $2.16 and Sol's $6.45. That's one run per model, so treat it as indicative.
Is Claude or GPT better for research with sources?
In this bench, GPT-6.1 Sol was best at it with 28 out of 30, ahead of Opus 5.5 and GPT-6 Astra at 26. Sol cited eight vendor pages and caught a stale price that made another model fail. Opus was close and took about 2 minutes instead of 12. Either way, have a person check the sources before a client sees them.
Did the AI judges favor their own models?
Yes. Judge 1 (Opus 5.5) scored Opus 72 and Sol 69. Judge 2 (GPT-6.1 Sol) scored Sol 70 and Opus 67. Both scored blind, though a model may still recognize its own style. Combining the two judges cancels the effect, which is why we report the combined score.
Can I run some agents on Claude and others on GPT?
Yes. Each Dock agent runs on the model its owner picks, from the major labs, with Claude as the default. Our business agents guide covers how to hire the first one.
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "Is Opus 5.5 better than GPT-6.1 Sol?",
"acceptedAnswer": {
"@type": "Answer",
"text": "They tied at 139 out of 150 on our five business jobs. Opus led on sales and ops follow-up, Sol on bookkeeping and sourced research, and support was a tie. Opus was about three times faster and its run cost about a third as much; Sol had no flags and wrote the best-sourced research."
}
},
{
"@type": "Question",
"name": "Which is faster, Opus 5.5 or GPT-6.1 Sol?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Opus 5.5. It finished all five jobs in about 6 minutes, against about 19.5 for Sol. Roughly 12 of Sol's minutes went on sourced research, its best job. Times are estimates from when each output appeared."
}
},
{
"@type": "Question",
"name": "Is GPT-6.1 Sol cheaper than Opus 5.5?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Not in our test. Sol's Dock credit rate is 1x and Opus's is 5x, but the rate prices a unit of work, and Sol did far more work on the same jobs. At list prices, Opus's whole five-job run cost $2.16 and Sol's $6.45. That's one run per model, so treat it as indicative."
}
},
{
"@type": "Question",
"name": "Is Claude or GPT better for research with sources?",
"acceptedAnswer": {
"@type": "Answer",
"text": "In this bench, GPT-6.1 Sol was best at it with 28 out of 30, ahead of Opus 5.5 and GPT-6 Astra at 26. Sol cited eight vendor pages and caught a stale price that made another model fail. Opus was close and took about 2 minutes instead of 12. Either way, have a person check the sources before a client sees them."
}
},
{
"@type": "Question",
"name": "Did the AI judges favor their own models?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Yes. Judge 1 (Opus 5.5) scored Opus 72 and Sol 69. Judge 2 (GPT-6.1 Sol) scored Sol 70 and Opus 67. Both scored blind, though a model may still recognize its own style. Combining the two judges cancels the effect, which is why we report the combined score."
}
},
{
"@type": "Question",
"name": "Can I run some agents on Claude and others on GPT?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Yes. Each Dock agent runs on the model its owner picks, from the major labs, with Claude as the default. Our business agents guide covers how to hire the first one."
}
}
]
}
Hire the agent first, then give them the model that suits their job. Start for free.