GPT-6 Luna, the cheapest model we tested, kept pace with GPT-6.1 Sol on bookkeeping and ops follow-up at one-twentieth of Sol's Dock credit rate. It failed the sourced research job under both judges, quoting a stale $89 Canopy price under a citation that showed $93. Put routine structured work on Luna, and put research and anything a client will read on Sol.
Both are OpenAI models, at opposite ends of Dock's model picker. On October 14, 2026 we hired six models as fresh Dock agents, gave each the same five business jobs, and had two judges score the outputs blind. The full ranking of all six is in our guide to the best AI model for business agents.
Sol vs Luna at a glance
| GPT-6.1 Sol | GPT-6 Luna | |
|---|---|---|
| Combined score (both judges) | 139/150 | 124/150 |
| Clean runs (no flag) | 10/10 | 8/10 |
| Judge 1 (Opus 5.5) | 69/75 | 63/75 |
| Judge 2 (GPT-6.1 Sol) | 70/75 | 61/75 |
| Sales prospecting | 28/30 | 28/30 |
| Support triage | 24/30 | 22/30 |
| Bookkeeping | 30/30 | 29/30 |
| Ops follow-up | 29/30 | 30/30 |
| Sourced research | 28/30 | 15/30, failed under both judges |
| Minutes for all five jobs (estimated) | 19.5 | 19.0 |
| Dock credit rate | 1x | 0.05x |
Leave research out and the two are close: Sol scored 111 and Luna 109 on the other four jobs. A swing of 1 to 3 points per task is normal from one run to the next, so treat that as a tie. The 13-point gap on research is where they part ways.
The credit rate is the multiplier Dock's model picker shows. It prices a unit of work, and models use different amounts of work per job, so per-task spend will vary. We haven't published it yet.
Neither model is quick: Opus 5.5 finished all five jobs in about 6 minutes, Sol and Luna in about 19 each. Sol spent roughly 12 minutes on research, its best job.
Sourced research: where Luna fell down
The job was a sourced pricing comparison of Karbon, Canopy and TaxDome for a 10-person accounting firm, with "Not verified" wherever the agent couldn't confirm something. The prompt even warned: "Today is October 14, 2026, so pricing from old blog posts may be out of date."
Luna's 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 its source [4] and described as "page dated July 22, 2026". The old number then carried into the budget ("Month-to-month at $89 would total $10,680 over 12 months [3, 4].") and into the recommendation: "I'd examine Canopy Standard first: based on the published baseline price, it has the lowest estimated annual cost for all 10 users".
Sol ran into the same old blog and set it aside: "Current Canopy monthly prices differ from the $89/$131/$179 figures returned by an indexed vendor blog." It used the live page's $93/$136/$186, put Standard at "$8,880 annually or $11,160 monthly annualized", and marked Canopy Premium as Coming Soon, one of the two traps in the task. When TaxDome's site blocked it, Sol said so: "Direct page loading returned HTTP 403, so this is indexed vendor evidence, not a successful live-page inspection."
Sol wasn't perfect either. Judge 1 gave it 13 of 15: "Table cells are paragraph-length, so it needs heavy trimming before going to a client." Judge 2, which was Sol itself, gave it 15.
In fairness to Luna, judge 1 found that "Karbon, TaxDome and the other URLs check out", and noted that a lenient scorer might count the Canopy miss as an accuracy slip instead of a fail. Both judges flagged it anyway, because the wrong price sat under a citation that looked like it backed it up.
Support triage: both missed something, Luna missed more
Both models got the planted basics right, from merging duplicate emails to escalating a legal threat to the ops lead at P1. The difference showed up in the secondary remedies.
Tomasz's chair armrest snapped during use eight days after delivery. The policy gives defective items a prepaid return label. Sol got it: "We'll email you a prepaid label to return it as defective." Luna told him "our policy doesn't clearly specify a remedy" and passed the ticket to Operations.
Neither refunded Ben's $35 expedited fee, which the policy covers when a warehouse delay makes expedited shipping late. Luna said no: "so the policy doesn't provide a $35 fee refund." Sol hedged: "We'll check the shipment and promised dates." Luna also made the bulk buyer contact sales herself, and some of its replies were curt ("We can't refund this order.").
Sol scored 24 and Luna 22. Most models in the bench misread at least one remedy on this job, so whichever model drafts your support replies, have a person read the refunds before they go out.
Bookkeeping: a tie at a twentieth of the rate
This job was 40 bank transactions to code against a chart of accounts, with five planted problems: a personal vet bill, a savings transfer, a supplier refund, a duplicate charge and a 33% price rise. Both models coded all 40 to the answer key, caught all five, and left the real recurring payments alone.
Sol scored 30 and Luna 29. Luna lost its point on explanation. Its flag for the Pixelforge increase gives the reason as "September is higher than August." Sol's version: "A $20 increase, about 33.3%, may be an authorized plan/seat change or an unexpected billing change."
Your accountant may prefer Sol's notes, which include debit and credit entries for the card payments, but the codes were identical. For a monthly categorization run, Luna does the same work at 0.05x. Our guide to an AI bookkeeping agent covers what that agent does and how to hire one.
Ops follow-up: Luna's was cleaner
The agents turned messy meeting notes into an action list and a follow-up email under 150 words. The notes hid a reassigned owner, two cancelled items, and a date to work out ("two weeks from Thursday the 15th", which is Thursday, October 29). Both models caught all of it.
Luna scored 30 and Sol 29, mostly on readability. Luna's table rows are short: "Send the revised data-mapping document." Owner Jun, due Fri Oct 16. Sol's dependency cells run long: "none explicitly agreed; #1 helps resolve code questions. Do not assume the Friday deadline waits for Monday access." Both emails came in under the limit (127 words for Sol, 126 for Luna) and tied the dashboard date to getting system access. Luna's version: "That date assumes access by Oct 19; any delay shifts the draft equally."
On sales prospecting they tied at 28, and both first-touch emails needed a light rewrite.
Which to pick, by job
| Job | Pick | Why |
|---|---|---|
| Research with sources a client or investor will see | Sol | 28/30 and clean under both judges; Luna failed it under both |
| Customer replies and other client-facing drafts | Sol | Higher support score; Luna withheld the defective-item remedy and wrote curter replies |
| Bookkeeping categorization | Luna | Same 40 codes and same five flags, at 0.05x |
| Meeting notes to action list and follow-up | Luna | 30 vs 29, with a cleaner table |
| Prospect lists | Either | Tied at 28; Opus 5.5 and GPT-6 Astra scored 30 here |
If you only want one of the two, pick Sol. It tied for the top combined score with no flags from either judge, at one-fifth the rate of Opus 5.5 and GPT-6 Astra, though it was the slowest model in the bench. Our Opus 5.5 vs GPT-6.1 Sol comparison covers that choice.
How to run both: a bookkeeping agent on Luna, a research agent on Sol
In Dock you choose the model for each agent, so one team can use both. 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 Sol for vendor comparisons, market scans and anything with citations.
- Support drafts on Sol, with a person approving refunds before they go out.
Routine work then runs at Luna's rate, and you pay Sol's rate where a wrong number would reach a client. If Luna does produce something with prices in it, 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. For how to hire the first agent, 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, and got each job as one message on October 14, 2026. Outputs were scored blind against written answer keys, 0 to 5 each on accuracy, following instructions and usable-as-is. Any invented fact or missed planted trap failed the task regardless of points. Judge 1 was Opus 5.5 and judge 2 was GPT-6.1 Sol. Both are contestants, and each put its own model first, so Sol scored itself in this matchup. Dock defaults to Claude (here's why), and our first judge was Claude too. Both judges flagged Luna's research miss. 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, and minutes estimated from when each output appeared. The full method is in the pillar guide.
FAQ
Is GPT-6.1 Sol better than GPT-6 Luna?
On our five business jobs, Sol scored 139 of 150 to Luna's 124, with no flags from either judge. Almost all of the gap is sourced research: Sol 28 of 30, Luna 15. On bookkeeping, ops follow-up and sales prospecting they were within a point of each other.
Is GPT-6 Luna good enough for business work?
For structured, routine work, yes. Luna coded all 40 bookkeeping transactions correctly, caught every planted anomaly, and scored 30 of 30 on turning meeting notes into an action list and follow-up email. Don't send its research to a client without a person checking the prices.
How much cheaper is GPT-6 Luna than GPT-6.1 Sol?
In Dock's model picker, Luna's credit rate is 0.05x and Sol's is 1x, one-twentieth. That rate prices a unit of work, and models use different amounts of work per job, so per-task spend can differ from the ratio.
Which is faster, Sol or Luna?
About the same. Sol took an estimated 19.5 minutes across the five jobs and Luna 19.0, against about 6 for Opus 5.5 and 8 for GPT-6 Astra.
Can I use GPT-6 Luna for research?
Only for drafts a person checks. Luna cited Canopy's pricing page for an $89 monthly price that came from a July blog post; the live page showed $93, and both judges failed it. For research a client or investor will read, use Sol.
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 and a research agent on GPT-6.1 Sol in the same workspace.
Put Luna on the routine work and Sol on anything a client will read. Start for free.
