Claude Opus 5.5 beat GLM 5.3 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. Half of that 16-point gap came from one job, sourced research, where the second judge failed GLM. On meeting follow-ups the two tied with a perfect 30. GLM runs at about a seventh of Opus's rate in Dock, 0.7x against 5x, yet its five-job run cost $18.21 at list price against $2.16 for Opus, about 8 times as much, and took three times as long. In our test Opus 5.5 was the better pick for every job here, including the structured internal work where GLM matched it on quality.
A disclosure: Dock runs agents on Claude by default, our first judge was Opus 5.5 itself, and our second was GPT-6.1 Sol from OpenAI. GLM did not judge. This is one of the head-to-heads behind our guide to the best AI model for business agents.
Opus 5.5 vs GLM 5.3 at a glance
| Claude Opus 5.5 | GLM 5.3 (Z.ai) | |
|---|---|---|
| Combined score (two judges) | 139 / 150 | 123 / 150 |
| Clean runs (no flag) | 9 / 10 | 8 / 10 |
| Sales prospecting | 30 / 30 | 27 / 30 |
| Support triage | 24 / 30 | 22 / 30 |
| Bookkeeping | 29 / 30 | 26 / 30 |
| Ops follow-up | 30 / 30 | 30 / 30 |
| Sourced research | 26 / 30 | 18 / 30 |
| Minutes, all five jobs (estimated) | 6.0 | 18.0 |
| Cost of the five-job run (list price) | $2.16 | $18.21 |
| Dock credit rate | 5x | 0.7x |
A flag is an invented fact or a missed planted trap, and it fails that job whatever the points say. All three flags here came from judge 2: one for Opus, two for GLM. The credit rate is the multiplier in Dock's model picker, a price per unit of work. The run cost is what all five jobs came to at each lab's published prices, from Dock's records. They point in opposite directions: GLM did far more work (more steps, more reading, more retries) to finish the same jobs, so its run cost about 8 times Opus's at one-seventh the rate. Opus's run was the cheapest of all six models and GLM's the most expensive.
Ops follow-up: a dead heat on quality
The job: turn messy meeting notes into an action list and a follow-up email under 150 words. The notes hid two cancelled items, a reassigned owner and a date to work out (Thursday, October 29). Both models caught every trap, and both judges gave both a full 15.
GLM put its date math next to the deadline: Thu Oct 29 ("two weeks from Thursday the 15th"). Both cleared the cancelled items in the email. GLM wrote "Returns come straight from Stockwise once Jun's in, so no CSV needed. Board deck parked." Opus wrote "The board deck is parked until January, and no returns CSV is needed."
GLM's email ran 115 words and Opus's 138. On this job GLM gave up nothing in quality. It was slower, at about a minute against half a minute, and across the whole test its lower rate did not make it the cheaper model to run.
Sourced research: where most of the gap came from
Compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. Both models got the headline prices right and marked Canopy's Premium plan as coming soon, one of the planted traps.
Both were open about TaxDome, whose site blocked automated requests. Opus read the vendor's page through a reader service: "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." GLM used a search engine's copy and said so: "They come from a search index of those vendor pages, last verified ~August 2026".
The trouble was two lines in GLM's write-up. One table cell says "2- and 3-year terms discount to $9,500/$9,000 per seat". Those are totals for ten seats; GLM's own billing row had the per-seat figures right at $950 and $900. Its recommendation then called Karbon "the only one of the three that offers true month-to-month billing", while its own table lists Canopy Plus at $136 a month billed monthly. Judge 1 docked GLM for the month-to-month claim and a four-sentence recommendation, and scored it 12. Judge 2 counted both lines as invented facts and failed the job at 6 out of 15.
Opus scored 14 and 12; judge 2 docked it for a four-sentence recommendation (the prompt asked for two or three) and for leaving out TaxDome's multi-year Essentials rates. Opus took about 2 minutes, GLM about 8. GPT-6.1 Sol scored 28 here; see Opus 5.5 vs GPT-6.1 Sol.
Support triage: one refund apart
Ten customer emails against a refund policy. Both handled every planted item: the duplicate, two out-of-policy returns, a legal threat and a bulk order for sales. The split came on Ben, who paid $35 for expedited shipping on an order held in the warehouse. The policy refunds the fee for a warehouse delay.
Opus refunded it: "Your order HO-58190 was held at our warehouse until October 9, so the delay was on our side." GLM counted from the ship date instead, "Expedited shipping covers 2-3 business days after an order ships, not from the payment date," and ended with "so the $35 fee isn't refundable under our policy."
Both made the same miss on Aisha, who got the wrong lamp and should have been sent the right one. Opus said she was "welcome to place a new order for it on our site." GLM offered to "help you get one ordered." Judge 1 also docked GLM for promising Tomasz a label "today", then asking him to "reply here and I'll send the label over."
Judge 1 scored Opus 13 and GLM 11. Judge 2 scored both 11. Kimi K3 led this job at 27, and every model lost points here, so have a person read refunds before they go out.
Sales and bookkeeping: close calls and one flag each
Sales prospecting meant ranking 20 companies with three planted traps, then writing a first-touch email with one ask. Both ranked the same four real fits and kept every trap out. On the borderline company, Saltmarsh, Opus left fifth place empty ("I didn't pad the list to 5 with a company that breaks a must-have rule"), and GLM put Saltmarsh fifth with the caveat to "treat it as a watchlist pick until the Head of Marketing role posted in Sept 2026 is filled." The key accepts both. The points went on the email, where GLM offered two asks: "I can send them over, or we can take 20 minutes." Opus 30, GLM 27.
On bookkeeping, both coded all 40 transactions to the key and caught all five planted anomalies, and judge 2 flagged each once. GLM copied one insurance date as 2026-08-28 when the bank export says 2026-08-27, the only copying slip in 240 rows across all six models. Judge 1 took a point off; judge 2 failed it. Opus added an extra check on a new contractor, justified with "Every other month only J Okafor is paid", from an export that covered two months. Judge 1 saw a useful extra; judge 2 failed it. Opus 29, GLM 26. More in AI bookkeeping agent.
How the judge conflict moved the scores
Judge 1 is Opus 5.5, scoring its own outputs; judge 2 is GPT-6.1 Sol, with no stake in this matchup. Both scored blind.
| Job | Judge 1: Opus | Judge 1: GLM | Judge 2: Opus | Judge 2: GLM |
|---|---|---|---|---|
| Sales prospecting | 15 | 14 | 15 | 13 |
| Support triage | 13 | 11 | 11 | 11 |
| Bookkeeping | 15 | 13 | 14 (flag) | 13 (flag) |
| Ops follow-up | 15 | 15 | 15 | 15 |
| Sourced research | 14 | 12 | 12 | 6 (flag) |
| Total | 72 | 65 | 67 | 58 |
The Claude judge put Opus 7 points ahead. The OpenAI judge put Opus 9 ahead. So the gap is wider without Claude in the scoring chair. Research drives it: 2 points under judge 1, 6 under judge 2. Judge 2 was stricter on both models, and it flagged Opus too.
Which to pick
| Business job | Pick | Why |
|---|---|---|
| Ops follow-up and meeting notes | Either on quality; Opus 5.5 on cost and speed | 30/30 each from both judges. Opus's whole run cost $2.16 to GLM's $18.21. |
| Research a client will see | Opus 5.5, or GPT-6.1 Sol | Opus 26, GLM 18 with a fail from judge 2; Sol led at 28 |
| Support inbox against a policy | Opus 5.5, or Kimi K3 | Opus refunded a fee GLM denied; Kimi led at 27 |
| Sales prospecting | Opus 5.5 | 30 vs 27; GLM's email made two asks |
| Bookkeeping | Opus 5.5 | 29 vs 26; both coded 40 of 40, GLM copied one date wrong |
| One model for everything | Opus 5.5 | 139 vs 123, 6 vs 18 minutes, $2.16 vs $18.21 for the five-job run |
In Dock you choose the model per agent. GLM 5.3 tying Opus on ops follow-up is a sign it can do that work well; it is no longer a price argument, because in our test the lower rate came with the most expensive run of the six. For GLM against Kimi K3, which tied it at 123, 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 gave each the same five jobs. Two blind judges scored each output 0 to 5 on accuracy, following instructions and usable-as-is against a written answer key, and any invented fact or missed trap failed that job. This is one run per model on synthetic tasks with planted traps; a 1 to 3 point swing per job is normal between runs, so treat the small gaps with care. Both judges are contestants, minutes are estimates, run costs are list-price figures from that single run and move with how many steps a model takes, and pricing pages were checked on October 14. Full method in the pillar guide. Dock's records for GLM 5.3's run can be read two ways; the lower reading is $16.32 instead of $18.21, which doesn't change any comparison here.
FAQ
Is Claude Opus 5.5 better than GLM 5.3?
On our five business jobs, yes: 139 to 123 out of 150 across two blind judges. Opus led on research (26 to 18), sales, bookkeeping and support, and they tied at 30 on ops follow-up. Opus had 9 clean runs of 10, GLM 8.
Is GLM 5.3 cheaper than Claude Opus 5.5?
Not in our test. GLM 5.3 has the lower credit rate in Dock, 0.7x against 5x, but its run of all five jobs cost $18.21 at list price against $2.16 for Opus 5.5, about 8 times as much. The rate is a price per unit of work, and GLM used far more work to finish the same jobs. Opus was also faster: about 6 minutes against 18. One run, so treat it as indicative.
Can GLM 5.3 do research with sources?
Not for a client without a check. Judge 2 failed it for reading ten-seat totals as per-seat prices and for a billing claim its own table contradicted. Judge 1 passed it at 12 of 15.
Was the test biased toward Claude?
Our first judge was Opus 5.5 and Dock defaults to Claude, so we checked. The Claude judge put Opus 7 points ahead of GLM; the OpenAI judge, GPT-6.1 Sol, put Opus 9 ahead. Both scored blind.
Is GLM 5.3 an open model?
Z.ai's own model card calls it "the most capable open-weights model for coding". We tested it as a model you pick inside Dock, so openness played no part in the scores.
Can I use both Opus 5.5 and GLM 5.3 in Dock?
Yes. Each Dock agent runs on the model its owner picks, and the menu includes Claude (the default), OpenAI, Kimi and GLM. Why your agents run on Claude explains the default.
Check what a run costs before trusting the rate, and put Opus 5.5 where the sources have to hold up. Start for free.