GPT-6 Luna and GLM 5.3 are the two lowest-rate models we tested in Dock, and on five real business jobs they finished a point apart: 124 and 123 out of 150, with 8 clean runs out of 10 each. Low rates did not mean low cost in our test. Luna's whole five-job run cost $5.71 at list price and GLM's cost $18.21, and both cost more than Opus 5.5's $2.16, one of the two highest-rate models we tested and the cheapest run. Between these two, Luna is the pick: it matched or beat GLM on four of the five jobs and its run cost about a third as much. Neither model should send sourced research to a client unless a person checks the numbers first.
On October 14, 2026 we hired six models as fresh Dock agents, gave each the same five jobs, and had two judges score the outputs blind. The full ranking is in our guide to the best AI model for business agents.
Luna vs GLM at a glance
| GPT-6 Luna (OpenAI) | GLM 5.3 (Z.ai) | |
|---|---|---|
| Combined score (both judges) | 124/150 | 123/150 |
| Clean runs (no flag) | 8/10 | 8/10 |
| Judge 1 (Opus 5.5) | 63/75 | 65/75 |
| Judge 2 (GPT-6.1 Sol) | 61/75 | 58/75 |
| Sales prospecting | 28/30 | 27/30 |
| Support triage | 22/30 | 22/30 |
| Bookkeeping | 29/30 | 26/30, failed under judge 2 |
| Ops follow-up | 30/30 | 30/30 |
| Sourced research | 15/30, failed under both judges | 18/30, failed under judge 2 |
| Minutes for all five jobs (estimated) | 19.0 | 18.0 |
| Cost of the five-job run (list price) | $5.71 | $18.21 |
| Dock credit rate | 0.05x | 0.7x |
A flag is an invented fact or a missed planted trap, and it fails that task whatever the points say. Luna's two flags are one research task, failed by both judges. GLM's both came from judge 2, in research and bookkeeping; judge 1 passed it on all five jobs. A swing of 1 to 3 points per task is normal between runs, so 124 against 123 is a tie.
Speed and cost: low rates, costly runs
The credit rate is the multiplier Dock's model picker shows, a price per unit of work. Luna's 0.05x is the lowest of the six models and GLM's 0.7x the second lowest. What a run costs depends on how much work a model does, and both did a lot: more steps, more reading, more retries to finish the same jobs.
Dock's records put the cost of each five-job run, at each lab's published list prices, at $5.71 for Luna and $18.21 for GLM, the most expensive run of the six. Opus 5.5, at a 5x rate, cost $2.16, and GPT-6 Astra about $2.53 (an estimate). So the two highest-rate models were the cheapest runs, and these two low-rate models were not cheaper to run. This is one run per model, so treat the figures as indicative.
Speed tracked cost. Opus 5.5 finished all five jobs in about 6 minutes; Luna and GLM took 19 and 18.
Bookkeeping: Luna's rows matched the bank export
The job was 40 bank transactions to code, with five planted problems such as a personal vet bill on the business card and a duplicate charge. Both models coded all 40 to the answer key, caught all five and left the real recurring payments alone.
GLM copied one date wrong, an insurance premium dated 2026-08-28 when the bank export says 2026-08-27. It was the only copying error in 240 rows across all six models. Judge 1 took off a point and judge 2 failed the task for it.
Luna lost its one point on explanation. Its flag for the Pixelforge increase gives the reason as "September is higher than August." GLM's tells you what to find out: "A 33% jump on a recurring tool is either an intentional plan upgrade or a billing error". Combined, Luna scored 29 and GLM 26. More in our AI bookkeeping agent guide.
Ops follow-up: both perfect
Messy meeting notes had to become an action list and a follow-up email under 150 words, with a reassigned owner, two cancelled items and a date to work out ("two weeks from Thursday the 15th", which is Thursday, October 29). Both caught all of it, and both judges gave both models 15 of 15.
Both kept the caveat that the dashboard date depends on system access. Luna: "That date assumes access by Oct 19; any delay shifts the draft equally." GLM: "if access slips, the draft slips day-for-day". On this job the quality was identical. Across the whole test Luna's run cost less than GLM's, though more than Opus 5.5 or GPT-6 Astra.
Sales prospecting: same shortlist, two different slips
The job was to rank 20 companies against an ideal 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 agreed on the four companies that clear every rule.
They split on the fifth slot, and the answer key accepts both calls. Luna left it open: "Only four of the 20 companies meet every must-have." GLM put Saltmarsh fifth with a caveat: "treat it as a watchlist pick until the Head of Marketing role posted in Sept 2026 is filled."
Both emails needed work. Luna kept to one ask ("Would you be open to a 20-minute call?") but opened with a broken sentence, "Quillmark is still using the site template launched with in 2021", with no greeting. GLM's read better but offered two asks: "I can send them over, or we can take 20 minutes." Combined, Luna 28 and GLM 27.
Support triage: tied at 22, with different misses
Both handled every planted item in a 10-email inbox: the duplicate merged, two out-of-policy returns declined, a legal threat escalated at P1, a 52-unit order sent to sales. Both lost points on smaller remedies.
Both refused Ben's $35 expedited fee, which the policy refunds when a warehouse delay makes delivery late. Luna: "so the policy doesn't provide a $35 fee refund." GLM: "so the $35 fee isn't refundable under our policy."
Tomasz's chair armrest broke during use, and the policy gives defective items a prepaid return label. GLM got the remedy: "Since this is a defect, we'll cover the return: I'm emailing you a prepaid label today." Then the same reply ends "Just reply here and I'll send the label over." Luna told him "our policy doesn't clearly specify a remedy" and passed the ticket to Operations.
Aisha was sent the wrong lamp. GLM refunded her and offered to help her reorder ("we can help you get one ordered"). Luna left the right lamp pending: "confirm the next steps for the brass one." Luna's replies were also curt ("We can't refund this order.").
Kimi K3 led this job at 27, and our Kimi K3 vs GLM 5.3 comparison shows where it read the policy better. Whichever model drafts your support replies, have a person read refunds before they go out.
Sourced research: neither should go out unchecked
The job was a sourced pricing comparison of Karbon, Canopy and TaxDome for a 10-person accounting firm. Canopy's Premium plan is marked "Coming soon", and TaxDome bills upfront.
Luna failed under both judges. Its Canopy row read: "Standard: $74/user/month billed annually, or $89/user/month month-to-month [3]." Source [3] is Canopy's pricing page, which showed $93 that day. The $89 came from a Canopy blog post that Luna listed as source [4], "page dated July 22, 2026".
GLM got Canopy's live prices right, caught the Premium trap ("coming soon"), and said plainly that its TaxDome figures "come from a search engine's index of the vendor's own pages, last verified ~August 2026". Judge 1 passed it at 12 of 15. Judge 2 failed it: one cell says "2- and 3-year terms discount to $9,500/$9,000 per seat", which are ten-seat totals. Its recommendation 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.
Each tripped where the other didn't: Luna had TaxDome's multi-year prices right per seat ("$950/seat/year for 2 years and $900/seat/year for 3 years [5]"). Combined, GLM scored 18 and Luna 15, against 28 for GPT-6.1 Sol. Our GPT-6.1 Sol vs GPT-6 Luna comparison covers that gap.
Which to pick, by job
| Job | Pick | Why |
|---|---|---|
| Bookkeeping categorization | Luna, of these two | 29 vs 26: same codes and flags, no copying slip. Opus 5.5 and GPT-6 Astra had cheaper, faster runs. |
| Meeting notes to action list and follow-up | Either on quality; Luna on cost | Both 30/30; Luna's whole run cost $5.71 to GLM's $18.21 |
| Prospect lists | Luna, narrowly | 28 vs 27; Opus 5.5 and GPT-6 Astra scored 30 |
| Support inbox against a policy | Either, with a person on refunds | Tied at 22; Kimi K3 led at 27 |
| Research a client or investor will read | Neither without a human check | Luna failed under both judges, GLM under one; Sol scored 28 |
If you only want one of the two, pick Luna: the same score, and a run that cost about a third of GLM's. Neither is the budget choice it looks like from the rate, since Opus 5.5's run cost less than both. GLM's case is reasoning a reader can follow. Its bookkeeping flags explained themselves and its research passed one judge.
How to run them in Dock
In Dock you choose the model for each agent. If you have settled on one of these two, a sensible split:
- A bookkeeping agent on Luna to code the month's transactions and flag anything odd for your accountant.
- An ops agent on Luna to turn meeting notes into owners, dates and a follow-up email.
- A research agent on GPT-6.1 Sol for vendor comparisons and anything with citations.
If cost is the reason you are looking at low-rate models, test Opus 5.5 or GPT-6 Astra on the same work first. All six models did bookkeeping and ops follow-up well, and those two had the cheapest and fastest runs.
If Luna or GLM produces prices, check them against the vendor's live page first. Dock plans are per person per month (Starter $20, Standard $100, Scale $200) and each includes a monthly allowance of business credits. To hire the first one, see our business agents guide.
How we tested, and what it can't tell you
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 relabelled, shuffled and scored blind against written answer keys, 0 to 5 each on accuracy, following instructions and usable-as-is. Judge 1 was Opus 5.5 and judge 2 was GPT-6.1 Sol. Both are contestants and each put its own model first, which is why we add the scores. Dock defaults to Claude, and our first judge was Claude too. The limits: one run per model, synthetic tasks with planted traps, normal swings of 1 to 3 points per task (so ties are ties), pricing pages checked on October 14, minutes estimated from when each output appeared, and run costs that come from that single run and move with how many steps a model takes. 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 GPT-6 Luna better than GLM 5.3?
They tied. Luna scored 124 of 150 and GLM 5.3 scored 123, with 8 clean runs of 10 each. Luna led on bookkeeping (29 to 26) and sales (28 to 27), GLM led on research (18 to 15), and they tied on support (22) and ops follow-up (30).
What is the cheapest AI model for business agents?
In our test the cheapest run was Opus 5.5: all five jobs cost $2.16 at list price, then GPT-6 Astra at about $2.53 (an estimate). GPT-6 Luna has the lowest credit rate in Dock at 0.05x, but its run cost $5.71, and GLM 5.3, at 0.7x, cost $18.21, the most of the six. Low rates did not mean low cost. One run per model, so treat it as indicative.
Is GPT-6 Luna cheaper to run than GLM 5.3?
Yes, in our test, though by far less than the rates suggest. Luna's credit rate in Dock is one-fourteenth of GLM 5.3's (0.05x against 0.7x), while its five-job run cost $5.71 at list price against $18.21, about a third. Both cost more than Opus 5.5's $2.16.
Can GPT-6 Luna or GLM 5.3 do research with sources?
Only for drafts a person checks. Luna cited Canopy's pricing page for an $89 price that came from a July blog post, and both judges failed it. 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 on research.
Can I run different agents on different models in Dock?
Yes. Each Dock agent runs on the model its owner picks, from the major labs, with Claude as the default. You can put a bookkeeping agent on GPT-6 Luna, a support agent on Kimi K3 and a research agent on GPT-6.1 Sol in the same workspace.
Check what a run costs before trusting the rate, and keep a person between either model and your clients. Start for free.