Kimi K3 and GLM 5.3 tied: 123 out of 150 each when two blind judges scored them on five real business jobs. Put Kimi K3 on your support inbox, where it scored higher than every model we tested, and put GLM 5.3 on meeting follow-ups and other structured work, where it scored perfectly at under half Kimi's rate in Dock. Keep both away from client-facing research unless a person checks the sources first.
We ran both as fresh Dock agents on October 14, 2026, next to Opus 5.5 and three GPT-6 models. The full ranking is in our guide to the best AI model for business agents. This is the close-up on the two that finished level by different routes.
Kimi K3 vs GLM 5.3 at a glance
| Kimi K3 (Moonshot AI) | GLM 5.3 (Z.ai) | |
|---|---|---|
| Combined score, both judges | 123/150 | 123/150 |
| Clean runs (no flag) | 8/10 | 8/10 |
| Judge 1 (Opus 5.5) | 64/75 | 65/75 |
| Judge 2 (GPT-6.1 Sol) | 59/75 | 58/75 |
| Sales prospecting | 24/30 | 27/30 |
| Support triage | 27/30 | 22/30 |
| Bookkeeping | 29/30 | 26/30 |
| Ops follow-up | 27/30 | 30/30 |
| Sourced research | 16/30, failed under both judges | 18/30, failed under judge 2 |
| Minutes for all five jobs | 13.5 | 18.0 |
| Dock credit rate | 1.5x | 0.7x |
A flag is a fabricated fact or a missed planted trap, and it fails that task whatever the points say. Both models have 8 clean runs out of 10. Kimi's two flags are one research task, failed by both judges. GLM's two came from judge 2 alone, in research and bookkeeping. The credit rate is the multiplier in Dock's model picker, a price per unit of work. Models use different amounts of work per job, so treat it as a rate, never a per-task cost.
Support triage: Kimi K3 read the policy to the end
Both models handled every planted item in a 10-email inbox: they merged a duplicate, declined two out-of-policy returns, escalated a legal threat at P1 and sent a 52-unit order to sales. The split came on the smaller remedies further down the refund policy.
Ben paid $35 for expedited shipping, and his order sat in the warehouse first. The policy says to refund the fee when expedited arrives late because of a warehouse delay. Kimi K3 got it:
"Your order was held at our warehouse, which is on us, so we're refunding the $35 expedited fee to your original payment method."
GLM 5.3 started the clock at the ship date instead: "Expedited shipping covers 2-3 business days after an order ships, not from the payment date." It ended the reply with "so the $35 fee isn't refundable under our policy." The arithmetic is careful. The conclusion misses the one case the policy names.
Aisha 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." GLM refunded her and offered to help her buy the right one: "If you'd still like the brass Arden, we can help you get one ordered."
Kimi's only deduction from judge 2 was tone: a decline that apologizes twice, ending "Sorry we can't do more on this one." It was the only model of the six that both judges agreed got Ben's refund right without missing a second remedy, and 27 was the top support score in the test. Every model still lost points here, so have a person read refunds before they go out. Our AI customer service agent guide covers that setup.
Ops follow-up: GLM 5.3 was perfect
Messy meeting notes had to become an action list and a follow-up email under 150 words. The notes hid two items cancelled mid-call, a reassigned owner and a date to work out ("two weeks from Thursday the 15th", so Thursday October 29). Both models caught all of it.
GLM's email came in at 115 words and dealt with the cancelled items in one line: "Returns come straight from Stockwise once Jun's in, so no CSV needed. Board deck parked." Both judges gave it 15 out of 15.
Kimi had the same substance, down to "Board deck parked until January." The difference was length: 139 words to a reader, 154 by a raw count that treats its spaced dashes and bullets as words. Judge 1 counted real words and gave full marks. Judge 2 counted raw and docked it for going over 150.
Sourced research: neither should go out unchecked
The job: 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 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 said TaxDome has a "white-labeled portal + mobile app included in every plan, even Essentials", when TaxDome sells the white-label app as 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".
GLM caught the Premium trap, marking it "coming soon", and got Canopy's monthly prices right. Judge 1 passed it at 12 out of 15. Judge 2 failed it for two lines. One cell says "2- and 3-year terms discount to $9,500/$9,000 per seat", which are ten-seat totals; GLM's own billing row had the per-seat figures right at $950 and $900. Its recommendation also called Karbon "the only one of the three that offers true month-to-month billing", while its own table lists Canopy at $136 a month on monthly billing.
Combined, GLM scored 18 and Kimi 16, against 26 to 28 for the top three. For research a client will read, use GPT-6.1 Sol, Opus 5.5 or GPT-6 Astra; see Opus 5.5 vs GPT-6.1 Sol.
Sales and bookkeeping: close, with one slip each
Sales prospecting meant ranking 20 companies against a customer profile, with three planted traps and one borderline case, Saltmarsh, then writing a one-ask email. Both kept every trap out of the top five. GLM put Saltmarsh fifth with the right caveat: "treat it as a watchlist pick until the Head of Marketing role posted in Sept 2026 is filled." Kimi left fifth place empty and named a trap company as the nearest miss: "closest is Kestrel Forms, which fails headcount (12) and has no marketer in seat."
Both emails broke the one-ask rule. GLM offered "I can send them over, or we can take 20 minutes." Kimi added "If it's useful, we can find 20 minutes after that," and judge 2 also marked down its line that the prospect deserves "better than a five-year-old theme." Combined: GLM 27, Kimi 24.
Bookkeeping meant coding 40 bank transactions and flagging five planted problems, such as a personal vet bill on the business card. Both coded all 40 to the key and caught all five. Kimi also listed what it left alone, as asked: "Deliberately not flagged (per your note that the recurring items are real): the Halverson/Fenwick retainers, rent, J Okafor's contractor payments, and the varying Stripe payout amounts."
GLM copied one date wrong, an insurance premium dated 2026-08-28 when the bank export says 2026-08-27. It was the only transcription error in 240 rows across all six models. Judge 1 took off a point; judge 2 counted it as a fail. Combined: Kimi 29, GLM 26. More in AI bookkeeping agent.
Which to pick
| Job | Pick | Why |
|---|---|---|
| Support inbox against a written policy | Kimi K3 | 27/30, the top support score of all six models. Have a person read refunds before they send. |
| Meeting notes to action list and follow-up | GLM 5.3 | 30/30 from both judges, at a 0.7x rate |
| Bookkeeping first pass | Kimi K3 if nobody rechecks rows, GLM 5.3 if your accountant does | Both coded 40 of 40; GLM had one wrong date |
| Sales prospecting shortlist | GLM 5.3, of these two | Right call on the borderline company. Opus 5.5 and GPT-6 Astra scored 30. |
| Research a client will read | Neither without a human check | Both failed at least once; GPT-6.1 Sol led at 28 |
For one model across everything, the score is a tie. GLM 5.3 has the lower rate, 0.7x against 1.5x. Kimi K3 was faster, 13.5 minutes against 18 for all five jobs. In Dock you choose the model per agent, so a support agent on Kimi K3 and an ops agent on GLM 5.3 can work for the same company. More in business agents.
How we tested, and the limits
Each model ran as a fresh Dock agent with the same role line, no memory, no connected apps and only built-in web access, one message per job. Outputs were relabeled and shuffled, then scored 0-5 against written answer keys on accuracy, following instructions and usable-as-is, 15 points per job per judge. Judge 1 was Opus 5.5 and judge 2 was GPT-6.1 Sol; each is a contestant and put its own model first, which is why we add the scores. Dock defaults to Claude and our first judge was Claude. Neither Kimi K3 nor GLM 5.3 judged. Limits: one run per model, synthetic tasks with planted traps, and a 1-3 point swing per task is normal between runs, so 123 against 123 is a tie. Pricing pages were checked October 14; minutes are estimates. Full method in the pillar post.
FAQ
Is Kimi K3 better than GLM 5.3?
They tied at 123 out of 150. Kimi K3 led on support triage (27 to 22) and bookkeeping (29 to 26). GLM 5.3 led on ops follow-up (30 to 27), sales (27 to 24) and research (18 to 16). Both had 8 clean runs of 10.
Which is cheaper to run, Kimi K3 or GLM 5.3?
In Dock, GLM 5.3 runs at a 0.7x credit rate and Kimi K3 at 1.5x, under half. The rate is a price per unit of work, and models use different amounts of work per job, so it is not a cost per task.
Which is better for customer support, Kimi or GLM?
Kimi K3. It scored 27 out of 30 on our support inbox, the highest of six models, and correctly refunded a warehouse-delay shipping fee that GLM 5.3 denied. GLM scored 22. With either, have a person read refunds before they go out.
Can Kimi K3 or GLM 5.3 do research with sources?
Not for a client or investor without a check. Kimi K3 failed our research task under both judges, citing a review for a claim the page does not make. GLM 5.3 passed with judge 1 and failed with judge 2 for reading ten-seat totals as per-seat prices. GPT-6.1 Sol scored highest.
Are Kimi K3 and GLM 5.3 open models?
Moonshot's own Kimi K3 post calls it "the world's first open 3T-class model". Z.ai's model card for GLM 5.3 calls it "the most capable open-weights model for coding". We tested both as models you pick inside Dock, so openness played no part in the scores.
Can I run my agents on Kimi K3 or GLM 5.3 in Dock?
Yes. Dock's model menu includes Claude (the default), OpenAI, Kimi and GLM, and you choose the model for each agent.
Give each agent the model that won its job, and keep a person on refunds and research. Start for free.
