GPT-6 Astra and GPT-6.1 Sol tied at 139 out of 150 when we ran both on five real business jobs in Dock, and they were the only two of the six models we tested to come through both judges without a single flag. Pick Astra when turnaround or cost matters, or the work is sales: it finished all five jobs in about 8 minutes against Sol's 19.5, its whole run cost an estimated $2.53 at list price against Sol's $6.45, and it scored a perfect 30 on prospecting. Pick Sol when the work is research a client will read: it was the best researcher of the six. Sol's credit rate in Dock is one-fifth of Astra's, but in our run it did enough extra work to cost about two and a half times as much.
Both are OpenAI models. A disclosure: Dock runs agents on Claude by default, our first judge was Opus 5.5 from Anthropic, and our second judge was GPT-6.1 Sol, one of the two models in this matchup. This is one of the head-to-heads behind our guide to the best AI model for business agents.
GPT-6 Astra vs GPT-6.1 Sol at a glance
| GPT-6 Astra | GPT-6.1 Sol | |
|---|---|---|
| Lab | OpenAI | OpenAI |
| Combined score (two judges) | 139 / 150 | 139 / 150 |
| Clean runs (no flag) | 10 / 10 | 10 / 10 |
| Sales prospecting | 30 / 30 | 28 / 30 |
| Support triage | 24 / 30 | 24 / 30 |
| Bookkeeping | 30 / 30 | 30 / 30 |
| Ops follow-up | 29 / 30 | 29 / 30 |
| Sourced research | 26 / 30 | 28 / 30 |
| Minutes, all five jobs (estimated) | 8.0 | 19.5 |
| Cost of the five-job run (list price) | $2.53 (estimate) | $6.45 |
| Dock credit rate | 5x | 1x |
| Judge 1 total (Opus 5.5) | 70 / 75 | 69 / 75 |
| Judge 2 total (GPT-6.1 Sol) | 69 / 75 | 70 / 75 |
Sol, judging its own work, put itself one point ahead. Opus, the judge with no model in this pairing, put Astra one point ahead. A swing of 1 to 3 points per job is normal from run to run, so this is a tie.
Speed and cost
Astra finished all five jobs in about 8 minutes, second only to Opus 5.5. Sol took about 19.5, the slowest of the six, with roughly 12 of those on sourced research.
At each lab's list prices, Astra's whole five-job run cost about $2.53 (Dock's estimate from Astra's usage) and Sol's $6.45. The credit rates in Dock's model picker point the other way: Astra is 5x and Sol 1x. The rate prices a unit of work, and Sol did far more work to finish the same jobs, so in our test the lower-rate model was not cheaper to run. It's one run per model, so treat the costs as indicative. Opus 5.5 also scored 139, at 5x, and its run cost $2.16.
Sourced research: Sol's best job
The job: compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. The traps were Canopy's Premium tier, marked "Coming soon", and TaxDome's term plans, which are billed upfront. Three of the six models failed this job at least once.
Astra and Sol both got every price right, marked Premium as coming soon and did the 10-seat math correctly. Both put TaxDome Pro first. Astra: "I'd look at TaxDome Pro first." Sol added a caution to "avoid a multi-year commitment until the firm has tested the workflow."
Sol went further on the second trap and told the reader when the money is due: "Full-term total is normally upfront, not a monthly installment implied by an annualized number." Astra explained how multi-year prices add up but never said the full term is paid at the start. Sol also caught the stale price that sank GPT-6 Luna: "Current Canopy monthly prices differ from the $89/$131/$179 figures returned by an indexed vendor blog." It cited eight vendor pages to Astra's five. Astra priced a TaxDome setup for seasonal staffing, with "one 1-year owner seat plus nine monthly seats for 12 months" at $11,800 a year.
Judge 1 gave both 13 out of 15, docking Astra for a dense table with broken bold formatting and Sol for paragraph-length cells that need "heavy trimming before going to a client". Judge 2, which is Sol, gave Sol 15 and Astra 13, docking Astra for leaving out TaxDome's upfront billing. So the two-point gap comes entirely from Sol scoring itself, though the reason it gave is a real trap in our answer key.
Speed is the price. Sol spent about 12 minutes on this job and Astra about 4.
Sales prospecting: Astra wrote the better email
Rank 20 companies against an ideal customer profile with three planted traps, then write a first-touch email in a house voice with one ask. Both picked the same top four in the same order, skipped all three traps, and left the fifth slot empty rather than force in Saltmarsh, which has the revenue but no marketer. Sol: "I would leave the fifth slot unfilled rather than relax the ICP."
The email decided it. Astra opened with "Priya," and wrote to her directly:
"Quillmark raised its Series A in August, and its website still uses the template you launched with in 2021. With you joining as VP Marketing in September, this looks like a useful time to revisit it."
Sol skipped the greeting and mixed "you" with her full name in one sentence:
"Quillmark's website still uses the template you launched with in 2021, and Priya Ostrander joined as VP Marketing in September."
Both judges took a point off Sol for that, which is the whole 30 to 28 gap. Both kept to one ask, and Sol's email was tighter, at 61 words against 80, so a quick edit of the first line would have closed it. Our AI sales agent guide covers the rest of that job.
Support triage: mirror-image misses
Ten customer emails against a refund and returns policy. Both merged the duplicate, declined the two out-of-policy returns, escalated the legal threat at P1 and routed the bulk order to sales. Both scored 24 of 30, and their misses were opposites.
Ben's expedited order sat in the warehouse, and the policy refunds the $35 fee for a warehouse delay. Astra refunded it: "the warehouse delay was ours, so we'll refund your $35 expedited fee." Sol's internal note said "Do not automatically refund $35," and its reply ended with "We'll check the shipment and promised dates."
Tomasz's armrest snapped in use, a defect the policy covers with a prepaid return label. Sol sent one: "We'll email you a prepaid label to return it as defective." Astra held it back: "A senior team member needs to review the defect before we can confirm the available options."
Both sent Aisha a return label for the wrong lamp, and neither committed to sending the brass one she ordered. Whichever model runs your inbox, have a person read refund replies before they go out. No model scored above 27 of 30 here; Kimi K3 led.
Bookkeeping and ops follow-up: nearly identical
Both coded all 40 bank transactions correctly, caught all five planted anomalies and flagged none of the normal recurring payments. All four judgments were a perfect 15. Sol wrote out the journal entries a bookkeeper would want, as in "Asset transfer: debit 1010, credit checking 1000. Not a tax expense or owner's draw." All six models did well here, and the cheapest, fastest runs overall were Astra and Opus 5.5. See the AI bookkeeping agent.
On ops follow-up, both caught the reassigned owner, dropped the two cancelled items and computed Thursday, October 29. Both scored 14 from judge 1 and 15 from judge 2. Astra's email, written as Sam, listed Sam's own work in the third person ("Sam: updated SOW with phase 2 pricing by Oct 23") and opened with a bare "Marcus," line. Sol wrote the email in Sam's voice, "I send the updated SOW with phase 2 pricing Friday, Oct 23," but padded its action list with cells like "Do not assume the Friday deadline waits for Monday access."
How the judges scored them
Both judges scored blind, with outputs relabelled A to F and shuffled for every job.
| Job | Judge 1: Astra | Judge 1: Sol | Judge 2: Astra | Judge 2: Sol |
|---|---|---|---|---|
| Sales prospecting | 15 | 14 | 15 | 14 |
| Support triage | 13 | 13 | 11 | 11 |
| Bookkeeping | 15 | 15 | 15 | 15 |
| Ops follow-up | 14 | 14 | 15 | 15 |
| Sourced research | 13 | 13 | 13 | 15 |
| Total | 70 | 69 | 69 | 70 |
The judges matched point for point on four of the five jobs. Opus split them only on the sales email. Sol agreed about the email, then gave its own research two extra points, so its self-preference came to one point overall.
Which to pick
| Business job | Pick | Why |
|---|---|---|
| Sales prospecting | Astra | 30 vs 28; its email spoke to the buyer directly |
| Research a client or investor will read | Sol | 28 vs 26, caught the upfront-billing trap, eight sources; allow 12 minutes or more |
| Support inbox against a policy | Either, with a person checking refunds | 24 each, opposite misses; Kimi K3 led at 27 |
| Bookkeeping | Either; Astra for a faster, cheaper run | Both 30 / 30 |
| Ops follow-up and client emails | Either | 29 each; Sol's email, Astra's action list |
| One model for everything | Astra, unless research is the main job | 139 each; 8 vs 19.5 minutes; about $2.53 vs $6.45 for the five-job run |
In Dock you choose the model per agent, so your sales agent can run on Astra and your research agent on Sol. For Sol against OpenAI's lowest-rate model, see GPT-6.1 Sol vs GPT-6 Luna. For the two fastest and cheapest runs, see Opus 5.5 vs GPT-6 Astra, and for Sol against Claude, Opus 5.5 vs GPT-6.1 Sol. 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. Each got the same five jobs, one message per job, sent at 11:05am PT; all outputs were back by 11:33am. 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, and a 1 to 3 point swing per job is normal, so a tie is a tie. Both judges are contestants, minutes are estimates, and pricing pages were checked on October 14. Run costs come from Dock's records at each lab's list prices for this one run (Astra's is Dock's estimate), and they move with how many steps a model takes. The full method and all six models are in the best AI model for business agents guide.
FAQ
Is GPT-6 Astra better than GPT-6.1 Sol?
Neither, across our five business jobs. They tied at 139 out of 150 with zero flags each. Astra led sales prospecting by two points and Sol led sourced research by two; the other three jobs were level.
Which is faster, GPT-6 Astra or GPT-6.1 Sol?
Astra, by a wide margin. It finished all five jobs in an estimated 8 minutes against about 19.5 for Sol, the slowest of the six models. Most of the gap was research: about 4 minutes for Astra and about 12 for Sol.
Which costs less to run, GPT-6 Astra or GPT-6.1 Sol?
Astra, in our test. At list prices, Astra's whole five-job run cost an estimated $2.53 and Sol's $6.45, even though Sol's Dock credit rate is 1x and Astra's 5x. The rate prices a unit of work, and Sol did far more work on the same jobs. It's one run per model, so treat it as indicative.
Did GPT-6.1 Sol favor itself as a judge?
A little. Scoring blind, Sol gave itself 70 and Astra 69, with the gap coming from two extra points on its own research. Opus 5.5, our other judge, gave Astra 70 and Sol 69. Combined, the two models tie.
Can I use both GPT-6 Astra and GPT-6.1 Sol in Dock?
Yes. Each Dock agent runs on the model its owner picks, so a sales agent can run on Astra and a research agent on Sol. Dock defaults to Claude, and why your agents run on Claude explains why.
Put each agent on the model that fits their job. Start for free.