Claude Opus 5.5 and GPT-6 Astra tied at 139 out of 150 when we ran both on five real business jobs in Dock and had two judges score every output blind. Astra came out with a clean record from both judges, while Opus was faster (about 6 minutes for all five jobs against 8) and picked up one flag from the second judge on bookkeeping. If you want the safest record, pick Astra; if you want the quickest turnaround, pick Opus; for sales prospecting, pick either, because both scored a perfect 30.
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 Astra's lab. This is one of the head-to-heads behind our guide to the best AI model for business agents.
Opus 5.5 vs GPT-6 Astra at a glance
| Claude Opus 5.5 | GPT-6 Astra | |
|---|---|---|
| Combined score (two judges) | 139 / 150 | 139 / 150 |
| Clean runs (no flag) | 9 / 10 | 10 / 10 |
| Sales prospecting | 30 / 30 | 30 / 30 |
| Support triage | 24 / 30 | 24 / 30 |
| Bookkeeping | 29 / 30 | 30 / 30 |
| Ops follow-up | 30 / 30 | 29 / 30 |
| Sourced research | 26 / 30 | 26 / 30 |
| Minutes, all five jobs (estimated) | 6.0 | 8.0 |
| Dock credit rate | 5x | 5x |
| Judge 1 total (Opus 5.5) | 72 / 75 | 70 / 75 |
| Judge 2 total (GPT-6.1 Sol) | 67 / 75 | 69 / 75 |
The credit rate is what Dock's model picker shows: the price per unit of work. Models use different amounts of work on the same job, so treat it as a rate. The only other model to reach 139 was GPT-6.1 Sol, at 1x.
Sales prospecting: both perfect, different emails
The job: rank 20 companies against an ideal customer profile with three planted traps, then write a first-touch email in a set voice with one ask. Both skipped all three traps, ranked Quillmark first, and left the fifth slot empty. In Opus's words: "I didn't pad the list to 5 with a company that breaks a must-have rule."
The emails differ in style. Opus wrote 98 words and opened with the whole trigger in one sentence: "Quillmark raised a $14M Series A in August and brought you on as VP Marketing in September, but the website is still the template you launched with in 2021." Astra wrote 80 and split it in two: "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."
Both kept to one ask. Opus: "Want me to send over the 3 things we'd change on Quillmark's site? No call needed unless they're useful." Astra: "Would it help if I sent the 3 things we'd change on Quillmark's site?" Both judges gave both models 15 out of 15, so pick on voice. Our AI sales agent guide covers the rest of that job.
Support triage: same score, different misses
Ten customer emails against a refund 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 also got the hardest call right: refunding Ben's $35 expedited fee because the delay happened in the warehouse. Astra's reply: "Your October 9 shipment arrived within that transit window on October 13, but the warehouse delay was ours, so we'll refund your $35 expedited fee." Opus explained its reasoning in the notes: "a warehouse delay can only happen before shipping."
Each then missed one secondary remedy, and they missed different ones. Astra held back the defective-chair return the policy allows, telling Tomasz: "A senior team member needs to review the defect before we can confirm the available options." Opus told Aisha, who got the wrong lamp, to buy the right one herself: "If you'd still like the brass one, you're welcome to place a new order for it on our site." To its credit, Opus raised that call with the owner: "For Aisha especially, shipping the brass lamp now would be the nicer fix."
Judge 1 scored both 13 out of 15; judge 2 scored both 11. Whichever model runs your inbox, have a person read refund replies before they go out; no model scored above 27 out of 30 here.
Bookkeeping: the only flag in this matchup
Forty bank transactions to code, five planted anomalies to catch. Both coded all 40 correctly and caught all five. Astra stayed restrained, describing the software price rise as "A $20 monthly increase, approximately 33.3%, could be an unapproved plan/seat change or billing error," and noting that "Repeated retainers, rent and contractor payments are expected and are not flagged."
Opus added a sixth flag nobody asked for, on a new contractor payment: "Every other month only J Okafor is paid, and you said contractor payments are the same each month, so this is new." Judge 1 read that as a useful extra check and took nothing off. Judge 2 read "every other month" as a claim about months the two-month export doesn't cover and flagged it. That one flag is the whole gap between them on clean runs. If you want an agent that sticks strictly to the file in front of it, Astra did that here.
Ops follow-up: Opus wrote the email you could send
Messy meeting notes into an action list and a follow-up email under 150 words. Both caught the reassigned owner and the two cancelled items, and both computed Thursday, October 29.
The difference was the email. Astra wrote it as Sam but listed Sam's own work in the third person, as in "Sam: updated SOW with phase 2 pricing by Oct 23," and opened with a bare "Marcus," line. Judge 1 took a usability point off for that. Opus wrote "Updated SOW with phase 2 pricing (me): Fri Oct 23" and opened with "Thanks for yesterday. Here's where everything landed." Judge 2 gave both a full 15, so Opus leads this job by one point.
Sourced research: a tie built from opposite verdicts
Compare Karbon, Canopy and TaxDome pricing for a 10-person accounting firm, with live citations. This was the job that separated the field: three models failed it at least once. Opus and Astra both got every price right, both marked Canopy's Premium plan as coming soon, and both opened their recommendation with the same words: "I'd look at TaxDome Pro first."
Opus took about 2 minutes, used three vendor pages, and was honest about how it read one of them: "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." Astra took about 4 minutes, cited five vendor pages, and dug up costs a buyer would miss, such as "eSignature requests require separately purchased credits: 100 credits cost $100."
The judges split in opposite directions. Judge 1 gave Opus 14 and Astra 13, docking Astra for a dense table with broken bold formatting. Judge 2 gave Opus 12 and Astra 13, docking Opus for a four-sentence recommendation (the prompt asked for two or three) and for leaving out TaxDome's multi-year Essentials rates, and docking Astra for not mentioning that TaxDome bills upfront. Combined, 26 each. GPT-6.1 Sol scored 28 on this job, the best of the six, so see Opus 5.5 vs GPT-6.1 Sol if research is your main use.
How the judge conflict moved the scores
Judge 1 is Opus 5.5, scoring its own outputs. Judge 2 is GPT-6.1 Sol, from Astra's lab. Both scored blind, with outputs relabelled A to F and shuffled per job.
| Job | Judge 1: Opus | Judge 1: Astra | Judge 2: Opus | Judge 2: Astra |
|---|---|---|---|---|
| Sales prospecting | 15 | 15 | 15 | 15 |
| Support triage | 13 | 13 | 11 | 11 |
| Bookkeeping | 15 | 15 | 14 (flag) | 15 |
| Ops follow-up | 15 | 14 | 15 | 15 |
| Sourced research | 14 | 13 | 12 | 13 |
| Total | 72 | 70 | 67 | 69 |
Each judge put its own side two points ahead. The gaps come from ops (judge 1 docked Astra's email), bookkeeping (judge 2 flagged Opus) and research, where they disagreed in opposite directions. Add both judges together and the lean cancels to a dead heat. On sales and support they agreed point for point.
Which to pick
| Business job | Pick | Why |
|---|---|---|
| Sales prospecting | Either | Both 30/30. Opus writes a fuller email, Astra a shorter one. |
| Support inbox against a policy | Either, with a human check on refunds | Both 24/30, each missed one remedy. Kimi K3 led this job at 27. |
| Bookkeeping | Astra, or a cheaper model | Astra 30, Opus 29 with one judge's flag. All six models did this well. |
| Ops follow-up and client emails | Opus | 30 vs 29; Opus's email was ready to send. |
| Research a client will see | Either, or GPT-6.1 Sol | Tied at 26; Sol led at 28 at one-fifth the rate. |
| One model for everything | Opus if speed matters most, Astra if a clean record does | 139 each, same rate, 6 vs 8 minutes. |
All six models handled bookkeeping and ops follow-up, so on those jobs you can price down to GPT-6 Luna or GLM 5.3 (see Kimi K3 vs GLM 5.3). In Dock you choose the model per agent, so your sales agent can run on Opus and your research agent on 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 run to run, so a tie is a tie. Both judges are contestants, minutes are estimates, and pricing pages were checked on October 14. The full method and all six models are in the best AI model for business agents guide.
FAQ
Is Claude Opus 5.5 better than GPT-6 Astra?
Across our five business jobs, neither. They tied at 139 out of 150 with two blind judges. Opus led ops follow-up by a point and Astra led bookkeeping by a point. Astra had no flag from either judge; Opus had one, from the second judge, on a bookkeeping note.
Which is faster, Opus 5.5 or GPT-6 Astra?
Opus 5.5 was the fastest model in our test, at an estimated 6 minutes for all five jobs against about 8 for Astra. The biggest gap was sourced research: about 2 minutes against 4.
Which costs less to run in Dock?
They have the same credit rate in Dock, 5x, which is a price per unit of work rather than per job. If rate matters, GPT-6.1 Sol tied them at 139 at 1x, though it was the slowest of the six.
Was the test biased toward Claude?
Our first judge, Opus 5.5, scored Opus two points above Astra. Our second, GPT-6.1 Sol from OpenAI, scored Astra two points above Opus. Both judged blind, and combining them cancels the lean. Dock also runs agents on Claude by default, and we say so.
Can I use both Opus 5.5 and GPT-6 Astra in Dock?
Yes. Each Dock agent runs on the model its owner picks, so one agent can run on Opus 5.5 and another on GPT-6 Astra. Why your agents run on Claude explains why Claude is the default.
Pick the model for the job, then hire the agent. Start for free.