Claude Opus 5.5 beat Claude Sonnet 5.5 by three points, 135 to 132 out of 150, when we ran both on five real business jobs in Dock and had two non-Anthropic judges score every output blind. Both came out with 7 clean runs out of 10, and Sonnet's whole five-job run cost $1.66 at list price against $2.16 for Opus, the lowest of the eight models we have tested. Pick Opus for sales outreach, where it led 30 to 26; pick Sonnet for ops follow-up and for any job where cost per run matters more than three points.
A disclosure: Dock runs agents on Claude by default, and both contestants here are Anthropic models. So neither judge was a Claude model: we used GPT-6 Astra and GPT-6.1 Sol, both from OpenAI. This is one of the head-to-heads behind our guide to the best AI model for business agents.
Sonnet 5.5 vs Opus 5.5 at a glance
| Claude Opus 5.5 | Claude Sonnet 5.5 | |
|---|---|---|
| Combined score (two judges) | 135 / 150 | 132 / 150 |
| Clean runs (no flag) | 7 / 10 | 7 / 10 |
| Sales prospecting | 30 / 30 | 26 / 30 |
| Support triage | 23 / 30 | 23 / 30 |
| Bookkeeping | 28 / 30 | 28 / 30 |
| Ops follow-up | 29 / 30 | 30 / 30 |
| Sourced research | 25 / 30 | 25 / 30 |
| Minutes, all five jobs (estimated) | 6.0 | about 6.6 |
| Cost of the five-job run (list price) | $2.16 | $1.66 |
| Judge 1 total (GPT-6 Astra) | 68 / 75 | 67 / 75 |
| Judge 2 total (GPT-6.1 Sol) | 67 / 75 | 65 / 75 |
The cost row is what each run cost at the lab's published prices, for all five jobs together. Sonnet's run came in about a quarter cheaper than Opus's. For comparison, Fable 5.1, the newest Claude model, scored 128 in the same round and its run cost $5.11 (see Fable 5.1 vs Opus 5.5). One run per model, so treat costs as indicative: they move with how many steps a model takes.
Three of the five jobs ended level. The gap comes from sales, where Opus gained four points, and ops, where Sonnet won one back.
Sales prospecting: Opus picked one ask, Sonnet offered two
The job: rank 20 companies against an ideal customer profile with three planted traps, then write a first-touch email in a set voice with one ask. Both skipped every trap and ranked Quillmark first. They differed on the fifth slot. Opus left it empty: "I didn't pad the list to 5 with a company that breaks a must-have rule." Sonnet filled it with Saltmarsh Cloud, marked conditional, and explained that it "fails the in-house marketing owner rule today". Both judges accepted either approach.
The points went on the email. Sonnet's was tight at 66 words and opened with a clean trigger: "Quillmark raised a $14M Series A in August and you started as VP Marketing in September, but the site still runs on the template Quillmark launched with in 2021." Then it closed with both options from the brief: "Would a 20-minute call be useful? Or I can send you the 3 things we'd change on your site first." The voice guide asked for one ask, and both judges took a point off instructions and a point off usability for it. GPT-6.1 Sol's note: "gives two asks instead of choosing one."
Opus wrote 98 words and chose: "Want me to send over the 3 things we'd change on Quillmark's site? No call needed unless they're useful." It also said why: "I used the "send 3 things" ask instead of the call because it's the lower-effort option for someone a month into a new job." That earned a perfect 15 from both judges. Delete Sonnet's first question and its email is ready to send. Our AI sales agent guide covers the rest of that job.
Support triage: Opus got the refund right and the report wrong
Ten customer emails against a refund 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 23, but they missed different things, and only Opus's miss counted as a fail.
The hardest call was Ben, who paid $35 for expedited shipping on an order our warehouse held up. The answer key says refund the fee. Opus did: "We're refunding the $35 expedited shipping fee to your original payment method." Sonnet declined it, telling Ben "so under our policy the $35 fee isn't refundable in this case", though it raised the call with the owner: "If you'd rather refund the $35 as goodwill, that's your call, not something I can promise." The key treats that as an accuracy error, so it cost Sonnet points without a flag.
Opus's flag came from its note to the founder. The job was drafting replies, and Opus wrote: "So I read "late" as measured from when he paid, and refunded the $35." Both judges flagged that as reporting a refund nobody had made. GPT-6 Astra also caught "Renata owns it now" as a handoff stated as done. Neither model sent Aisha the brass lamp she ordered: Opus told her she was "welcome to place a new order for it on our site", and Sonnet promised "our team will follow up to arrange the brass lamp you ordered." Fable 5.1 led this job at 28.
Have someone read refund replies whichever model you use; neither scored above 23 out of 30 here.
Bookkeeping: identical scores, one flag each
Forty bank transactions to code, five planted anomalies to catch. Both coded all 40 correctly, caught all five, and scored 14 from each judge. Opus added a sixth flag on a new contractor. Sonnet put it in a "Not flagged as problems, but worth knowing" list instead.
GPT-6.1 Sol flagged both for the same habit: describing the state of books they could not see. Sonnet wrote that Stripe payouts mean "revenue and fees are both understated by the fee amount", and Sol called that "an unverified ledger defect" because a bank export can't show what is already booked elsewhere. Opus got flagged for "Every other month only J Okafor is paid", which reaches beyond the two months in the file. GPT-6 Astra wanted the same lines softened but flagged neither. Either model can run this job; Sonnet does it for less.
Ops follow-up: Sonnet's email had nothing to fix
Messy meeting notes into an action list and a follow-up email under 150 words. Both caught the reassigned owner and the two cancelled items, both dated the dashboard draft Thursday, October 29, and both opened the email with "Thanks for yesterday."
The one point came from a date nobody agreed on. No send date was set in the meeting for the review invite. Sonnet proposed one in the table ("No send date was set; I suggest Fri Oct 23.") and left it out of the client email. Opus also marked it as an assumption in its table, then put "Fri Oct 23" in the email to Marcus as a plain commitment. GPT-6 Astra took a usability point off for that; Sol gave both a full 15. Sonnet's 30 tied Fable 5.1 for the best score in the round on this job.
Sourced research: same score, one missed trap
Compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. Both used only the three vendors' own pricing pages and got the main prices and the cost maths right.
Sonnet missed the planted trap. Canopy's Premium plan is marked as coming soon on the vendor's page, and Sonnet listed it as a current plan: "Standard $74/user/month billed annually. Plus $109. Premium $149." Both judges failed it for that. It was honest elsewhere: it couldn't read Canopy's monthly prices and said so plainly, "I could not operate the billing toggle (page overlays blocked the click), and I did not estimate it." Its recommendation went the other way from Opus's: "I'd look at Karbon first".
Opus caught the trap, marking Premium "Coming soon", and recommended "I'd look at TaxDome Pro first." It lost points for a four-sentence recommendation when the prompt asked for two or three, and for broken bold formatting in its table. Sol docked both for leaving out TaxDome's cheaper multi-year Essentials rates. Combined, 25 each, but only Sonnet failed. If research goes to a client, have someone check plan availability before it leaves.
How the two judges scored it
Both judges are from OpenAI and neither model under test is theirs. Both scored blind, with outputs relabelled A to C and shuffled per job. Opus's outputs came from our first round, run earlier the same day with the same briefs, and were scored again alongside Sonnet's.
| Job | Astra: Opus | Astra: Sonnet | Sol: Opus | Sol: Sonnet |
|---|---|---|---|---|
| Sales prospecting | 15 | 13 | 15 | 13 |
| Support triage | 12 (flag) | 12 | 11 (flag) | 11 |
| Bookkeeping | 14 | 14 | 14 (flag) | 14 (flag) |
| Ops follow-up | 14 | 15 | 15 | 15 |
| Sourced research | 13 | 13 (flag) | 12 | 12 (flag) |
| Total | 68 | 67 | 67 | 65 |
The judges agreed on direction: both put Opus ahead, by one point and by two. Sol gave Opus the same 67 it gave in our first round, scoring the same outputs against a different field, which is a fair sign the judge is consistent. Round-one scores used a different judge pair, so compare models within a round: Opus scored 139 in round one, which is why our Opus 5.5 vs GPT-6 Astra matchup shows a higher number.
Which to pick
| Business job | Pick | Why |
|---|---|---|
| Sales prospecting | Opus | 30 vs 26. Sonnet's email offered two asks; delete one and it works. |
| Support inbox against a policy | Either, with a human check on refunds | 23 each. Opus got Ben's refund right but reported it as done; Sonnet declined it. |
| Bookkeeping | Sonnet | 28 each, one judge's flag each, and Sonnet's run cost less. |
| Ops follow-up and client emails | Sonnet | 30 vs 29; nothing to edit. |
| Research a client will see | Opus, with a check | 25 each, but Opus caught the coming-soon plan and Sonnet didn't. |
| One model for everything | Opus for outbound email, Sonnet on cost | 135 vs 132, $2.16 vs $1.66, 6.0 vs about 6.6 minutes. |
In Dock you choose the model per agent, so your sales agent can run on Opus while your ops and bookkeeping agents run on Sonnet. Our business agents guide explains how to hire the first one.
Method and limitations
We hired fresh Dock agents on October 14, 2026, each on one model (confirmed from Dock's records for every turn), with the same role, no memory and no connected apps. Each got the same five jobs, one message per job. Opus ran at 11:05am PT and Sonnet at 2:10pm PT, with the same briefs and answer keys. Two blind judges, GPT-6 Astra and GPT-6.1 Sol, scored each output 0 to 5 on accuracy, following instructions and usable-as-is, and any invented fact or missed trap failed that job. Costs are the list price the agent runtime reported for each whole run; Sonnet's figure excludes a short warm-up turn. Minutes are estimates. This is one run per model on synthetic tasks with planted traps, and a 1 to 3 point swing per job is normal from run to run, so three points is close to a tie. Across our two rounds, two agents reported that a tool summarising a web page returned invented content; both caught it and read the raw page, which is one more reason to keep checks on research. The full method is in the best AI model for business agents guide.
FAQ
Is Claude Opus 5.5 better than Sonnet 5.5?
By a small margin in our test. Opus scored 135 out of 150 and Sonnet 132, with two blind judges. Opus led sales prospecting by four points and Sonnet led ops follow-up by one; the other three jobs were tied. Both had 7 clean runs out of 10.
Which costs less to run, Sonnet 5.5 or Opus 5.5?
Sonnet. Its five-job run cost $1.66 at list price against $2.16 for Opus, about a quarter less, and the lowest of the eight models we have tested. That's one run each, so treat it as indicative.
Which is faster, Sonnet 5.5 or Opus 5.5?
Opus, slightly. It took an estimated 6 minutes for all five jobs against about 6.6 for Sonnet. Both were among the fastest models we tested.
Why weren't the judges Claude models?
Both contestants are Anthropic models and Dock runs agents on Claude by default, so we used two judges from OpenAI: GPT-6 Astra and GPT-6.1 Sol. Both scored blind, and both put Opus slightly ahead of Sonnet.
Can I use both Sonnet 5.5 and Opus 5.5 in Dock?
Yes. Each Dock agent runs on the model its owner picks, so a sales agent can run on Opus and an ops agent on Sonnet. Why your agents run on Claude explains why Claude is the default.
Pick the model for the job, then hire the agent. Start for free.