The best AI model for business agents right now is a three-way tie between Opus 5.5, GPT-6 Astra and GPT-6.1 Sol, which each scored 139 out of 150 when we ran six models on five real business jobs. The gap between them and the other three (GPT-6 Luna, Kimi K3 and GLM 5.3) came almost entirely from one job: research with live sources. So the useful question is which job your agent does and what rate you want to pay for it.
We ran this test inside Dock on October 14, 2026, with every model hired as a Dock agent and given the same work. One thing to know up front: Dock runs agents on Claude by default, and our first judge was Claude. We say more about that below, along with how we handled it.
How we tested
We hired six fresh agents in Dock, one per model: Opus 5.5 (Anthropic), GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna (all OpenAI), Kimi K3 (Moonshot AI) and GLM 5.3 (Z.ai). Each got the same role line, "Business operations agent for a small B2B company", with no memory, no connected apps and only built-in web access. We confirmed from Dock's own records which model answered every message, so nothing here rests on a model saying what it is.
Each agent got five jobs, one message per job, identical across models. We sent the first at 11:05am PT and all 30 outputs were back within 28 minutes.
- Sales prospecting: rank 20 companies against an ideal customer profile, with three planted trap companies and one borderline case, then write a first-touch email in a house voice with one ask.
- Support triage: work a 10-ticket inbox against a refund and returns policy, with duplicates to merge, out-of-window requests to decline, a P1 escalation, a sales lead hiding in the queue and three tickets where the right remedy takes a careful read.
- Bookkeeping: code 40 bank transactions to a chart of accounts and flag five planted anomalies (a personal vet bill, a savings transfer, a refund offset, a duplicate charge and a 33% price rise) without flagging the real recurring items.
- Ops follow-up: turn messy meeting notes into an action list and a follow-up email under 150 words, with a reassigned owner, cancelled items and a date to compute ("two weeks from Thursday the 15th" is Thursday, October 29).
- Sourced research: compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm with live citations. The traps were that Canopy Premium is marked "Coming soon" and that TaxDome's term plans are billed upfront.
Scoring was blind. We relabelled the outputs A to F and shuffled them for each task, then scored every output against a written answer key, 0 to 5 on accuracy, following instructions and whether you could use it as-is. That makes 15 points per task and 75 per judge.
One rule mattered more than the points. Any made-up fact or missed planted trap is a flag, and a flagged task fails no matter how many points it earned. A support reply that refunds the wrong customer is a failure even if it reads beautifully.
We used two judges from two labs. Judge 1 was Opus 5.5 and judge 2 was GPT-6.1 Sol, both scoring blind. Both are also contestants, which is a real conflict of interest, and the next section shows what it did to the scores.
The results
Combined scores from both judges. Each job is out of 30 (15 per judge), the total is out of 150, and "Clean" counts the judgments with no flag, out of 10.
| Model | Combined /150 | Clean /10 | Sales | Support | Books | Ops | Research |
|---|---|---|---|---|---|---|---|
| GPT-6.1 Sol | 139 | 10 | 28 | 24 | 30 | 29 | 28 |
| GPT-6 Astra | 139 | 10 | 30 | 24 | 30 | 29 | 26 |
| Opus 5.5 | 139 | 9 | 30 | 24 | 29 | 30 | 26 |
| GPT-6 Luna | 124 | 8 | 28 | 22 | 29 | 30 | 15 |
| Kimi K3 | 123 | 8 | 24 | 27 | 29 | 27 | 16 |
| GLM 5.3 | 123 | 8 | 27 | 22 | 26 | 30 | 18 |
Read the research column first. The top three scored 26 to 28 there and the bottom three scored 15 to 18. On bookkeeping and ops follow-up, all six models scored between 26 and 30, and four of the six scored 29 or 30 on both.
Support triage was the hardest job for everyone. Nobody scored above 27 of 30, and most models misread at least one of the secondary remedies in the refund policy.
What each judge said
Each judge put its own model first. Opus scored Opus 72 of 75. Sol scored Sol 70 of 75.
| Model | Judge 1 (Opus 5.5) /75 | Judge 2 (GPT-6.1 Sol) /75 | Combined /150 |
|---|---|---|---|
| GPT-6.1 Sol | 69 | 70 | 139 |
| GPT-6 Astra | 70 | 69 | 139 |
| Opus 5.5 | 72 | 67 | 139 |
| GPT-6 Luna | 63 | 61 | 124 |
| Kimi K3 | 64 | 59 | 123 |
| GLM 5.3 | 65 | 58 | 123 |
Both judges scored blind, so neither knew which output came from which model. A model may still recognise writing that looks like its own, and the scores suggest that happened a little on both sides. Adding the two judges together cancels it out, and the top three land on exactly the same total.
The judges agreed more than they disagreed. Both put the same three models on top and the same three below. Both failed Kimi K3 and GPT-6 Luna on the research job. Judge 2 was stricter: it also flagged GLM 5.3 twice (research and bookkeeping) and Opus 5.5 once (bookkeeping), where judge 1 flagged neither. Sol and Astra came through both judges with no flags.
Sales prospecting: Opus 5.5 and GPT-6 Astra
Opus 5.5 and GPT-6 Astra both scored 30 of 30. Sol and Luna scored 28, GLM 27 and Kimi K3 24.
Every model ranked the same company first, and none put a trap company in its top five. The split came from the borderline case, Saltmarsh, which had the revenue to qualify but no marketer in the seat, and from the house voice's one-ask rule. Astra applied the right rule to Saltmarsh:
Saltmarsh Cloud: No current in-house marketing owner; an open Head of Marketing role does not count. Its ~$4M revenue meets the financial rule despite its older raise.
Kimi K3 named one of the planted traps as its runner-up instead:
Nothing else passes every must-have; closest is Kestrel Forms, which fails headcount (12) and has no marketer in seat.
Its email also made two asks where the voice guide allows one: "If it's useful, we can find 20 minutes after that." That is small, and it is exactly what a founder would send back for a rewrite. For more on what a prospecting agent does day to day, see the AI sales agent.
Support triage: Kimi K3
Kimi K3 led support with 27 of 30. Opus 5.5, Astra and Sol scored 24, and Luna and GLM 22.
All six handled the obvious parts. They merged the duplicate tickets, declined the two requests outside the return window, escalated the P1 to the right person and routed the sales lead to sales. The split came from a customer named Ben, who paid $35 for expedited shipping on October 5 and got an email saying his order was delayed at the warehouse until October 9. The policy refunds the fee when expedited delivery is late "because of us (warehouse delay)". Kimi K3 read it correctly:
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 5.5 reached the same answer and showed its reasoning:
a warehouse delay can only happen before shipping. So I read "late" as measured from when he paid, and refunded the $35.
GLM 5.3 counted from the ship date and declined:
The warehouse delay was pre-ship, so the $35 fee stands. Decline.
Luna also declined, and Sol hedged with "We'll check the shipment and promised dates." Luna sent a customer with a broken chair to Operations for review, noting that "policy does not clearly specify a remedy", though the policy says to send a prepaid return label.
Kimi was the only model both judges agreed got Ben right without missing a second remedy. Judge 2 still took points off for apologising twice in one reply. Whichever model you pick, have a person read refund decisions before they go out.
Bookkeeping: GPT-6.1 Sol and GPT-6 Astra
Sol and Astra scored 30 of 30. Opus 5.5, Luna and Kimi K3 scored 29, and GLM 5.3 26.
This job barely separated the models. Judge 1 found that all six coded all 40 transactions to the answer key and caught all five planted anomalies, and all six coded both credit card autopays as liability payments. Sol's note on the duplicate charge is what you want from a bookkeeper: it flags the problem and doesn't touch the ledger until someone checks.
Identical date, merchant and amount suggest a duplicate, but separate purchases or licenses are also possible. Removing a real bank charge would break reconciliation.
The two flags here came from judge 2. Opus 5.5 flagged a new contractor payment as worth confirming, but it backed that up with a claim the data couldn't support:
Every other month only J Okafor is paid, and you said contractor payments are the same each month, so this is new.
The export only covered two months, so "every other month" claims more than the data shows. Judge 1 counted it as a useful extra check and didn't deduct. Judge 2 flagged it. GLM 5.3 copied one transaction date as 2026-08-28 when the source said 2026-08-27. It was the only copying error in 240 transcribed rows, and it's the kind of slip an accountant would catch and then stop trusting. More on handing the books to an agent in the AI bookkeeping agent.
Ops follow-up: Opus 5.5, GPT-6 Luna and GLM 5.3
Opus 5.5, Luna and GLM 5.3 scored 30 of 30. Sol and Astra scored 29, and Kimi K3 27.
No model missed a trap. All six dropped the cancelled CSV export and the parked board deck, gave the mapping doc to its new owner and worked out the date. Opus 5.5 handled the reassignment in a single line item:
Send the revised data mapping doc (Jun owns it now, not Sam)
GLM 5.3 showed its date math next to the deadline: Thu Oct 29 ("two weeks from Thursday the 15th").
The nearest thing to a miss was length. Kimi K3 signed off its email with "(word count: 136)". Judge 1 counted 139 real words. A raw word count that treats its dashes and bullets as words gives 154, over the 150 limit, and judge 2 took points off for it. If your team has a hard format rule, check it the first few times.
Research with sources: GPT-6.1 Sol
Sol scored 28 of 30. Opus 5.5 and Astra scored 26, GLM 5.3 18, Kimi K3 16 and Luna 15.
On this job a wrong answer looks right. A pricing comparison with citations reads as checked, and a founder may forward it to a client without opening every link. Five of the seven failed judgments in the whole test came from this one job.
Sol caught both traps. On Canopy Premium:
Premium's published rates imply $17,880 or $22,320, respectively, but the tier is Coming Soon.
On TaxDome's billing: "Full-term total is normally upfront, not a monthly installment implied by an annualized number." Sol also spotted the stale price that tripped up another model: "Current Canopy monthly prices differ from the $89/$131/$179 figures returned by an indexed vendor blog."
GPT-6 Luna used that stale blog price as if it were current, citing a page that showed something else:
Standard: $74/user/month billed annually, or $89/user/month month-to-month [3]
Source [3] showed $93. Both judges failed Luna on this job. Kimi K3 cited a review for "server slowdowns around tax deadlines" that the page doesn't mention, listed Premium without the "Coming soon" label, and said TaxDome's white-label app comes with every plan when it's sold as an add-on. Judge 2 failed GLM 5.3 for reading TaxDome's ten-seat totals as per-seat prices.
Opus 5.5 was the most open about its own sourcing. It noted that TaxDome's site blocked its direct request, explained how it read the page instead, and suggested a quick check in a browser before the comparison went to the client. For a standing research role, see the AI research agent.
Speed and rate
| Model | Minutes, all five jobs | Longest job | Dock credit rate |
|---|---|---|---|
| Opus 5.5 | 6.0 | Research, 2.0 min | 5x |
| GPT-6 Astra | 8.0 | Research, 4.0 min | 5x |
| Kimi K3 | 13.5 | Research, 5.5 min | 1.5x |
| GLM 5.3 | 18.0 | Research, 8.0 min | 0.7x |
| GPT-6 Luna | 19.0 | Research, 7.0 min | 0.05x |
| GPT-6.1 Sol | 19.5 | Research, 12.0 min | 1x |
Minutes are estimates, taken from when each output appeared in Dock. Opus 5.5 and Astra were the fastest by a wide margin. Sol was the slowest, mostly because it spent about 12 minutes on research, which was also its best job.
The credit rate is the multiplier Dock's model picker shows next to each model. It's a price per unit of work, and different models use different amounts of work on the same task, so a lower rate doesn't translate directly into a smaller bill for each task. We haven't published per-task spend yet. As a guide, Sol tied for the top score at one-fifth the rate of Opus 5.5 and Astra and took the longest. Luna's rate is one-hundredth of Opus 5.5's, and Luna was fine on structured work and weak on research.
Which model should your agents run on
| If your agent does this | Run it on | Also good | Before it goes out |
|---|---|---|---|
| Research a client or investor will see | GPT-6.1 Sol | Opus 5.5, GPT-6 Astra | Keep Luna, Kimi K3 and GLM 5.3 off this job unless a person checks every source |
| Support inbox against a policy | Kimi K3 | Opus 5.5, Astra, Sol | A person reads every refund decision |
| Bookkeeping | GPT-6 Luna or GLM 5.3, on price | Any of the six | Spot-check dates and amounts against the bank export |
| Ops follow-up from meeting notes | GPT-6 Luna or GLM 5.3, on price | Any of the six | Check hard format limits like word counts |
| Sales prospecting | Opus 5.5 or GPT-6 Astra | Sol, Luna | Read borderline accounts and the ask in the email |
| One model for everything, speed first | Opus 5.5 or GPT-6 Astra | ||
| One model for everything, rate first | GPT-6.1 Sol | Expect slower research |
Mix models across your team
In Dock, each agent runs on the model you pick for them, and different agents can run on different models. The model menu covers the major labs: Claude, which is the default, plus OpenAI, Kimi and GLM. You pick it agent by agent.
The results above make a mixed team the sensible default. A research agent on GPT-6.1 Sol gets the most careful sourcing at a 1x rate. A bookkeeping agent on GPT-6 Luna does near-perfect coding at 0.05x. A support agent on Kimi K3 read the refund policy best. A sales agent on Opus 5.5 is fast and scored top marks on prospecting. They can all work in the same workspace, so you see their work side by side.
If you're starting from one agent, the guide to business agents covers what to hire first. If you want to know why Dock defaults to Claude in the first place, we wrote that up in why your agents run on Claude.
Dock plans are priced per person per month, Starter $20, Standard $100 and Scale $200, and each includes business credits your agents draw on at the rates above.
Limitations
This was one run per model. The same prompt can swing 1 to 3 points per task from one run to the next, so read ties as ties and don't treat a one-point lead as a win.
The tasks were synthetic. The inputs were short and clean with known traps planted in them. Real inboxes and ledgers are messier, and we didn't test long multi-step work or how a model recovers from its own mistakes.
The judges are also contestants. We disclosed it, kept both blind and used judges from two different labs, and the combined score cancels out the self-preference each judge showed. It's still a limit.
Pricing pages for the research job were checked on October 14, 2026, and they change. The minutes are estimates.
The head-to-heads
We wrote up four of the closest matchups in more detail.
- Opus 5.5 vs GPT-6 Astra: tied at 139 with the same 5x rate. Opus 5.5 was a little faster and Astra came through with no flags from either judge.
- Opus 5.5 vs GPT-6.1 Sol: tied at 139. Sol was the better researcher at one-fifth the rate, and Opus 5.5 finished in about a third of the time.
- GPT-6.1 Sol vs GPT-6 Luna: Sol leads 139 to 124 because of research. On bookkeeping and ops, Luna matched it at one-twentieth of Sol's rate.
- Kimi K3 vs GLM 5.3: tied at 123. Kimi K3 is the support pick, GLM 5.3 was stronger on ops and sales at about half Kimi's rate, and neither should do sourced research unchecked.
FAQ
Which AI model is best for business?
In our test of six models on five business jobs, Opus 5.5, GPT-6 Astra and GPT-6.1 Sol tied for first with 139 of 150 points across two blind judges. They pulled ahead on research with live sources. On bookkeeping and ops follow-up, all six models scored 26 or more out of 30, so for that kind of work the cheaper models are a fair choice.
Is Claude better than ChatGPT for business work?
We tested the underlying models inside Dock. Anthropic's Opus 5.5 tied with OpenAI's GPT-6 Astra and GPT-6.1 Sol at 139 of 150. Opus 5.5 was the fastest of the six, at about 6 minutes for all five jobs. Sol was the strongest researcher and runs at one-fifth of Opus 5.5's rate in Dock.
What is the best AI model for research?
GPT-6.1 Sol scored 28 of 30 on our sourced research job and caught both planted traps. Opus 5.5 and GPT-6 Astra scored 26. GPT-6 Luna, Kimi K3 and GLM 5.3 each failed that job with at least one judge for stale prices, misread prices or a claim the cited page didn't make, so give them research only if a person checks the sources.
What is the best AI for bookkeeping?
All six models did well. Judge 1 found that every model coded all 40 transactions correctly and caught all five planted anomalies. GPT-6.1 Sol and GPT-6 Astra scored a perfect 30 of 30. GPT-6 Luna scored 29 at the lowest credit rate in Dock, which makes it a good pick for a bookkeeping agent.
What is the cheapest AI model that still does good work?
GPT-6 Luna, at a 0.05x credit rate in Dock, which is one-hundredth of Opus 5.5's 5x. It scored 29 or 30 on bookkeeping and ops follow-up and failed research. GLM 5.3, at 0.7x, also scored 30 on ops. If you want top-tier quality on every job at a low rate, GPT-6.1 Sol tied for first at 1x.
Can I run different agents on different models?
Yes. In Dock, each agent runs on the model you pick for them, from Claude, OpenAI, Kimi and GLM. A team can run a research agent on Sol, a bookkeeping agent on Luna and a support agent on Kimi K3 in the same workspace.
Is this test biased toward Claude?
Dock defaults to Claude and our first judge was Opus 5.5, so we brought in a second judge from OpenAI, GPT-6.1 Sol, and had both score blind. Each judge ranked its own model first. Combined, Opus 5.5 tied with two OpenAI models, and Opus 5.5 was the only one of the three with a flag.
The best model for your business is the one that fits each job, and in Dock every agent can have their own. Start for free.
