Claude Opus 5.5 beat Kimi K3 by 139 to 123 out of 150 when we ran both on five real business jobs in Dock and had two judges score every output blind. Most of that gap came from sourced research, which Kimi K3 failed under both judges. Kimi K3 won support triage 27 to 24, the top support score of all six models we tested. Opus was also faster and cheaper: its five-job run cost $2.16 at list price against $3.25 for Kimi K3, even though Kimi runs at under a third of Opus's rate in Dock. Put research, sales, ops and client emails on Opus. The case for Kimi K3 is support quality: it led that job, so a support inbox is where it earns its place.
A disclosure: Dock runs agents on Claude by default, and our first judge was Opus 5.5 itself. Our second judge was GPT-6.1 Sol, from OpenAI. Neither judge comes from Moonshot AI, which makes Kimi K3. This is one of the head-to-heads behind our guide to the best AI model for business agents.
Opus 5.5 vs Kimi K3 at a glance
| Claude Opus 5.5 | Kimi K3 | |
|---|---|---|
| Combined score (two judges) | 139 / 150 | 123 / 150 |
| Clean runs (no flag) | 9 / 10 | 8 / 10 |
| Sales prospecting | 30 / 30 | 24 / 30 |
| Support triage | 24 / 30 | 27 / 30 |
| Bookkeeping | 29 / 30 | 29 / 30 |
| Ops follow-up | 30 / 30 | 27 / 30 |
| Sourced research | 26 / 30 | 16 / 30, failed under both judges |
| Minutes, all five jobs (estimated) | 6.0 | 13.5 |
| Cost of the five-job run (list price) | $2.16 | $3.25 |
| Dock credit rate | 5x | 1.5x |
| Judge 1 total (Opus 5.5) | 72 / 75 | 64 / 75 |
| Judge 2 total (GPT-6.1 Sol) | 67 / 75 | 59 / 75 |
A flag (an invented fact or a missed planted trap) fails that job whatever the points say. Kimi's two flags are one research job, failed by both judges; Opus's one came from judge 2 on bookkeeping. Of the 16-point gap, research accounts for 10, sales for 6 and ops for 3, and Kimi K3 wins 3 back on support.
Support triage: Kimi K3 led all six models
Ten customer emails against a refund policy. Both models merged the duplicate, declined the two out-of-policy returns, escalated a legal threat at P1, sent a 52-unit order to sales and gave the customer with a broken chair the prepaid return the policy allows. Both also got the hardest call right: Ben's $35 expedited fee, refundable because his order sat in the warehouse. Kimi K3:
Your order was held at our warehouse, which is on us, so we're refunding the $35 expedited fee to your original payment method.
Opus showed its working in a note to the owner:
a warehouse delay can only happen before shipping. So I read "late" as measured from when he paid, and refunded the $35.
The split came on Aisha, who was sent the wrong lamp. Kimi: "We'll get the brass Arden shipped to you and email a prepaid label so the black one goes back at no cost." Opus refunded her and left the reorder to her: "If you'd still like the brass one, you're welcome to place a new order for it on our site." Opus told the owner that "shipping the brass lamp now would be the nicer fix," but the customer still got the weaker reply.
Kimi's only deductions were for tone: one decline apologizes twice and ends "Sorry we can't do more on this one." Judge 1 scored Kimi 14 and Opus 13; judge 2, 13 and 11. Kimi K3 was the only model of the six that both judges agreed got Ben right without missing a second remedy. Nobody scored above 27, so have a person read refund replies before they go out. Our AI customer service agent guide covers that setup.
Sourced research: where the gap opened
Compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. Canopy's Premium plan is marked "Coming soon", and TaxDome bills upfront.
Kimi K3 failed under both judges. It wrote that "reviewers also report server slowdowns around tax deadlines [4]", and the cited review page says nothing about slowdowns. It listed "Premium: $149/user/mo" with no sign the plan isn't on sale yet. And it described a TaxDome "white-labeled portal + mobile app included in every plan, even Essentials", when the white-label app is an add-on. In its favor, Kimi said plainly that TaxDome's site blocked it and "TaxDome figures rely on the two third-party sources above".
Opus got every price right, marked Premium as coming soon and wrote "A white-labeled (firm-branded) mobile app is a paid add-on." It also found Canopy's monthly prices ($93, $136 and $186) on the pricing page's billing toggle, where Kimi wrote "exact monthly rates not published". TaxDome blocked Opus too, so it read TaxDome's own page another way and said it was "worth one click to confirm in a browser before this goes to the client." It took about 2 minutes to Kimi's 5.5.
Judge 1 gave Opus 14 and Kimi 8; judge 2 gave Opus 12 and Kimi 8. Opus lost points for a four-sentence recommendation where the prompt asked for two or three. GPT-6.1 Sol scored 28 here, the best of the six; see Opus 5.5 vs GPT-6.1 Sol if research is your main use.
Sales prospecting: Opus 30, Kimi 24
Rank 20 companies against a customer profile with three planted traps and one borderline case, Saltmarsh, then write a first-touch email with one ask. Both kept every trap out of the top five and left fifth place empty. The difference was the runner-up. Opus named the right one: "Closest is Saltmarsh Cloud (see below): re-check it once their Head of Marketing hire lands." Kimi named a planted trap: "closest is Kestrel Forms, which fails headcount (12) and has no marketer in seat."
Opus's 98-word email kept to one ask: "Want me to send over the 3 things we'd change on Quillmark's site? No call needed unless they're useful." Kimi added a second ask, "If it's useful, we can find 20 minutes after that," and judge 2 marked down its line that the prospect deserves "better than a five-year-old theme." Both judges gave Opus 15 and Kimi 12.
Bookkeeping and ops follow-up: close
On bookkeeping, both coded all 40 transactions to the key, caught all five planted anomalies and scored 29. Kimi listed what it left alone, as asked: "Deliberately not flagged (per your note that the recurring items are real)". Opus added a sixth flag on a new contractor: "Every other month only J Okafor is paid, and you said contractor payments are the same each month, so this is new." The export covered two months, so judge 2 flagged that as unsupported; judge 1 saw a useful extra check.
The ops job turned meeting notes into an action list and a follow-up email under 150 words. Both caught the reassigned owner and the two cancelled items and worked out Thursday, October 29. The gap was length. Kimi signed off with "(word count: 136)"; a raw count that treats its dashes and bullets as words gives 154, and judge 2 docked it. Opus reported both counts, 138 and 145, and stayed under. Opus 30, Kimi 27.
Speed and cost
Opus finished all five jobs in about 6 minutes, the fastest of the six models; Kimi K3 took about 13.5. Minutes are estimates, taken from when each output appeared in Dock.
Opus was also the cheaper run. Dock's records put the whole five-job run at $2.16 for Opus and $3.25 for Kimi K3, at each lab's published list prices. That runs against the credit rate in Dock's model picker, 5x for Opus and 1.5x for Kimi. The rate is a price per unit of work, and Kimi used more work to finish the same jobs: more steps, more reading, more retries. Opus's run was the cheapest of all six models. This is one run, so treat the figures as indicative; costs move with how many steps a model takes.
How the judges scored it
| Job | Judge 1: Opus | Judge 1: Kimi | Judge 2: Opus | Judge 2: Kimi |
|---|---|---|---|---|
| Sales prospecting | 15 | 12 | 15 | 12 |
| Support triage | 13 | 14 | 11 | 13 |
| Bookkeeping | 15 | 15 | 14 (flag) | 14 |
| Ops follow-up | 15 | 15 | 15 | 12 |
| Sourced research | 14 | 8 (flag) | 12 | 8 (flag) |
| Total | 72 | 64 | 67 | 59 |
Both judges scored blind, with outputs relabelled and shuffled per job. Each put its own model first overall, which is why we add them. On this matchup they agreed: both put Opus 8 points ahead of Kimi K3, and both put Kimi ahead on support.
Which to pick
| Business job | Pick | Why |
|---|---|---|
| Support inbox against a policy | Kimi K3, with a human check on refunds | 27 vs 24, the top support score of six models |
| Research a client will see | Opus 5.5 | 26 vs 16; Kimi failed under both judges. GPT-6.1 Sol led at 28. |
| Sales prospecting | Opus 5.5 | 30 vs 24; right runner-up, one ask |
| Ops follow-up and client emails | Opus 5.5 | 30 vs 27; stayed under the word limit |
| Bookkeeping | Either | Both 29; all six models did this well. Opus's run was cheaper and faster. |
| One model for everything | Opus 5.5 | 139 vs 123, faster, and a cheaper run ($2.16 vs $3.25) despite the 5x rate |
In Dock you pick the model for each agent, so a support agent on Kimi K3 and a research agent on Opus 5.5 can work for the same company. Don't move bookkeeping to a lower-rate model to save money: in our test GPT-6 Luna and GLM 5.3 had lower rates and costlier runs; see Kimi K3 vs GLM 5.3. Our business agents guide explains how to hire the first one.
Method and limitations
We hired six fresh Dock agents on October 14, 2026, each on a different model (confirmed from Dock's records for every turn), with the same role, no memory and no connected apps, and sent each the same five jobs at 11:05am PT. Two blind judges scored every output 0 to 5 on accuracy, following instructions and usable-as-is against a written answer key; any invented fact or missed trap failed the job. This is one run per model on synthetic tasks, and a 1 to 3 point swing per job is normal, so Kimi's 3-point support lead could narrow on another run. Both judges are contestants, minutes are estimates, run costs are list-price figures from that single run, and pricing pages were checked on October 14. Full method in the pillar guide; for Opus against the model it tied, see Opus 5.5 vs GPT-6 Astra.
FAQ
Is Claude Opus 5.5 better than Kimi K3?
Overall, yes: 139 to 123 out of 150 across five business jobs and two blind judges. Opus led on research, sales and ops follow-up, and they tied on bookkeeping. Kimi K3 led on support triage, 27 to 24.
Is Kimi K3 good for customer support?
Yes. It scored 27 out of 30 on our support inbox, the best of six models, and shipped a customer the lamp she ordered where Opus 5.5 asked her to reorder. Have a person read refunds before they go out, whichever model you use.
Can Kimi K3 do research with sources?
Not for a client or investor without a check. Kimi K3 failed our research task under both judges: it cited a review for "server slowdowns around tax deadlines" that the page doesn't mention and listed Canopy's Premium plan without its "Coming soon" label. Opus 5.5 scored 26 out of 30 on the same job.
Which costs less to run in Dock, Opus 5.5 or Kimi K3?
Opus 5.5, in our test. Its run of all five jobs cost $2.16 at list price against $3.25 for Kimi K3, and it finished in about 6 minutes against 13.5. Kimi has the lower credit rate, 1.5x against 5x, but the rate is a price per unit of work and Kimi used more work for the same jobs. One run, so treat it as indicative.
Is Kimi K3 an open model?
Moonshot AI's own Kimi K3 post calls it "the world's first open 3T-class model". We tested it as a model you pick inside Dock, so openness played no part in the scores.
Was the test biased toward Claude?
Our first judge was Opus 5.5 and Dock defaults to Claude, so we added a second blind judge from OpenAI, GPT-6.1 Sol. Both judges put Opus 8 points ahead of Kimi K3, and both put Kimi ahead on support.
Give support to Kimi K3 and research to Opus 5.5, and keep a person on refunds. Start for free.