Your picks, names still hidden

TaskYour pickYour scoreYour line on why
LinkedIn post from raw notesBkimi-k3 , Cfable-5.1 , Eopus-5.54Human-readable, to the point, with only a few immediate AI tells.; Much more usable as is. Tells a story, human-readable, with far fewer AI quirks.; Effective and does the work without fluff.
Pricing pushback + sensitive emailAgpt-6.1-sol , Ggpt-6-astra3While there is some creativity, net 90 is a big problem.; I don't think offering to remove the managers' sessions is the right move here. I do like that it does not commit to a scope for the $65 million, which tests how sensitive they are.
One-shot exercise for a new audience (wish list)Agpt-6.1-sol , Fgrok-4.7 , Ggpt-6-astra4Very comprehensive, perhaps a bit on the long side.; Well-written, easy-to-follow prompts in a dedicated window; A bit on the longer side but seems the best out of the bunch.
Week plan across engagements (wish list)Fgrok-4.74More to the point than the rest
Course participant hub websiteEopus-5.55The best look and feel, the most carefully thought-out, the most complete experience

The verdicts

Each newer option against what I use today, decided by my own blind scores.

opus-5.5 against opus-5
Split

It joins the lineup for the tasks it wins.

Better on: LinkedIn post from raw notes; Course participant hub website. Weaker on: Pricing pushback + sensitive email; One-shot exercise for a new audience (wish list). Level on 1 task.

grok-4.7 against opus-5
Switch

It becomes the new default.

Better on: LinkedIn post from raw notes; One-shot exercise for a new audience (wish list); Week plan across engagements (wish list). Level on 2 tasks.

gpt-6.1-sol against opus-5
Switch

It becomes the new default.

Better on: LinkedIn post from raw notes; Pricing pushback + sensitive email; One-shot exercise for a new audience (wish list); Week plan across engagements (wish list); Course participant hub website.

fable-5.1 against opus-5
Switch

It becomes the new default.

Better on: LinkedIn post from raw notes; One-shot exercise for a new audience (wish list); Week plan across engagements (wish list); Course participant hub website. Level on 1 task.

gpt-6-astra against opus-5
Split

It joins the lineup for the tasks it wins.

Better on: LinkedIn post from raw notes; Pricing pushback + sensitive email; One-shot exercise for a new audience (wish list). Weaker on: Course participant hub website. Level on 1 task.

kimi-k3 against opus-5
Switch

It becomes the new default.

Better on: LinkedIn post from raw notes; Week plan across engagements (wish list); Course participant hub website. Level on 2 tasks.

Every option, side by side

OptionYour averageTasks you picked itJudge averageTasks the judge picked itTotal costSeconds per answerWords out, roughly
fable-5.1 was C3.2013.801$2.548125.57566
gpt-6.1-sol was A3.2022.801$0.267120.63957
grok-4.7 was F3.0023.401$0.596263.814714
kimi-k3 was B2.8011.800$0.29666.22900
gpt-6-astra was G2.6023.000$1.16593.43443
opus-5.5 was E2.6021.800$1.435142.010681
opus-5 (what I use today) was D2.0002.802$1.417128.88425

How long each one took

Time to produce the answer, and what it cost, for every task. The fastest on each task is highlighted. Speed and cost are tie-breakers: read them after the scores.

Taskfable-5.1gpt-6.1-solgrok-4.7kimi-k3gpt-6-astraopus-5.5opus-5
LinkedIn post from raw notes25s
$0.077
10s
$0.004
32s
$0.016
8s
$0.007
10s
$0.020
20s
$0.037
12s
$0.024
Pricing pushback + sensitive email38s
$0.112
44s
$0.013
1m 31s
$0.037
13s
$0.011
36s
$0.070
37s
$0.060
47s
$0.083
One-shot exercise for a new audience (wish list)1m 32s
$0.289
1m 51s
$0.041
2m 18s
$0.057
54s
$0.031
1m 29s
$0.186
1m 14s
$0.122
1m 27s
$0.152
Week plan across engagements (wish list)1m 18s
$0.271
1m 08s
$0.026
1m 28s
$0.036
28s
$0.021
1m 14s
$0.132
1m 09s
$0.122
1m 15s
$0.132
Course participant hub website6m 34s
$1.800
6m 09s
$0.183
16m 09s
$0.449
3m 48s
$0.225
4m 17s
$0.756
8m 28s
$1.093
7m 01s
$1.025

You and the judge

You picked the same winner on 1 of 5 tasks. Where you differ, the usual causes are a scoring guide that misses something you care about, a judge that misread, or polish that swayed you.

TaskYour pickThe judge's pickSame?
LinkedIn post from raw noteskimi-k3, fable-5.1, opus-5.5grok-4.7no
Pricing pushback + sensitive emailgpt-6.1-sol, gpt-6-astragpt-6.1-solyes
One-shot exercise for a new audience (wish list)gpt-6.1-sol, grok-4.7, gpt-6-astrafable-5.1no
Week plan across engagements (wish list)grok-4.7opus-5no
Course participant hub websiteopus-5.5opus-5no

The raw scores

How strict was the judge? You gave top marks 1 times out of 35 (average 2.77). The judge gave top marks 5 times out of 35 (average 2.77). A judge that hands out many top marks cannot separate the options, which is a reason to sharpen the scoring guide or to trust your own scores on that task.

How close is the judge to you? Its score is within one point of yours on 26 of 35 answers. It can still pick a different winner, so read its reasons below before you trust it on a task.

We judged the judge

The first judge was told only to score against the guide, and it gave top marks to almost everything. The second was told to act as a tough editor: rank first, at most one 5, competent work earns a 3. Scores per task, highest first:

TaskFirst judgeStrict judgeYou
LinkedIn post from raw notes5 5 4 4 4 1 15 4 3 3 3 1 14 4 4 3 3 3 2
Pricing pushback + sensitive email5 5 5 4 1 1 15 4 3 2 2 1 13 3 2 2 2 2 1
One-shot exercise for a new audience (wish list)no scores5 4 3 2 1 1 14 4 4 3 2 2 1
Week plan across engagements (wish list)5 5 5 5 5 5 15 4 3 3 2 2 14 3 3 3 2 2 2
Course participant hub website5 5 5 5 5 5 55 4 3 3 3 2 25 4 3 3 2 2 1

LinkedIn post from raw notes

OptionYour scoreYour line on whyJudgeJudge vs youThe judge's reason
fable-5.1 was C4Much more usable as is. Tells a story, human-readable, with far fewer AI quirks.4sameDelivers an excellent witty tone and great extrapolations, but requires an edit because it bakes one of the alternative endings into the main post.
opus-5.5 was E4Effective and does the work without fluff.3-1Features solid tone and good structure, but includes unrequested meta-commentary and word counts that need to be edited out.
kimi-k3 was B4Human-readable, to the point, with only a few immediate AI tells.1-3Hits a reject condition by inventing specific times (9am and 5pm) not present in the raw notes, though the storytelling is otherwise engaging.
grok-4.7 was F3I liked how it handled the "useless" part, but overall the writing is just okay and not that inspiring.5+2Follows every rule perfectly and captures the requested zero-hype tone, though the transition into the second paragraph is slightly choppy.
gpt-6.1-sol was A3Useful but a bit on the long side and not every claim is accurate.3sameCompetent and follows all instructions, but the tone is a bit dry and the transition from the hook is abrupt.
gpt-6-astra was G3Something about the phrasing is a bit off, and it is longer than it needs to be.3sameMeets all requirements, but the phrasing is slightly clunky and it reads a bit too much like standard LinkedIn advice.
opus-5 was D2Readable and clear, but a bit too light, and it uses the "A, not B" pattern at least once.1-1Hits two reject conditions by inventing a timeframe (seven hours) and using a variation of the banned 'not X, Y' construction.

Pricing pushback + sensitive email

OptionYour scoreYour line on whyJudgeJudge vs youThe judge's reason
gpt-6.1-sol was A3While there is some creativity, net 90 is a big problem.5+2Meets all rules, perfectly protects value, and smartly advises the user to calculate their own minimums rather than inventing line-item prices.
gpt-6-astra was G3I don't think offering to remove the managers' sessions is the right move here. I do like that it does not commit to a scope for the $65 million, which tests how sensitive they are.4+1Meets all rules and offers excellent strategic advice, though the email is slightly less structured than the top output.
fable-5.1 was C2Better, more human-friendly writing, but it gives away too much in discounts.3+1Competent and meets all rules, but arbitrarily discounts the full package to 92 million, giving up too much value unnecessarily.
kimi-k3 was B2Gives too much ground on price across the board, and suggests moving to a call, which I would avoid here.2sameFails to protect the value of the recordings by offering them for free in Option B, and invents an unprompted pilot workshop.
opus-5 was D2Too many discounts.2sameConfident but sloppy in execution ("I can work to 65... Option 1 - 66 million") and requires the client to do math to understand the 84 million option.
grok-4.7 was F2I would not use the time-bound offer, and the tone of the writing is a bit harsh.1-1Hits a reject condition by inventing a 10 million expiring discount, and the email tone is aggressive rather than warm.
opus-5.5 was E1Offers a large discount, and option 2 moves to a call, which I would avoid here.1sameHits a reject condition by inventing a new cohort size of 24, which compromises the stated hands-on quality of the proposal.

One-shot exercise for a new audience (wish list)

OptionYour scoreYour line on whyJudgeJudge vs youThe judge's reason
grok-4.7 was F4Well-written, easy-to-follow prompts in a dedicated window4sameMeets all rules and uses clever meta-prompting to avoid brackets, but pushes the testing phase entirely into the second layer instead of the core loop.
gpt-6-astra was G4A bit on the longer side but seems the best out of the bunch.3-1Avoids brackets and provides a polished handout, but forces a fictional workflow instead of having participants pick their own as the source exercise required.
gpt-6.1-sol was A4Very comprehensive, perhaps a bit on the long side.2-2Avoids brackets but forces a fictional workflow, uses fictional test data instead of real inputs, and offers a weak stretch challenge.
fable-5.1 was C3I am not sure why it offers a menu when it interviews people later. The quick references at the end are a bit off.5+2Meets all rules, cleverly avoids brackets by having users paste their playbook below the prompt, and adapts the Jira stretch into native ChatGPT connectors.
kimi-k3 was B2The prompts are hard to follow, and some instructions are written in too much shorthand.1-1Hits a reject condition by including brackets to fill in within the prompts and printing time estimates on the handout.
opus-5 was D2Hard to follow, and it needs more polish.1-1Hits a reject condition by including brackets to fill in within the prompts and printing time estimates on the handout.
opus-5.5 was E1Too basic, and it may have been cut off in the middle.1sameHits a reject condition because the output cuts off mid-sentence and fails to deliver a complete handout.

Week plan across engagements (wish list)

OptionYour scoreYour line on whyJudgeJudge vs youThe judge's reason
grok-4.7 was F4More to the point than the rest4sameDelivers punchy, highly opinionated advice that perfectly balances the constraints, missing only a draft message for the delayed non-prospect emails.
kimi-k3 was B3On the shorter side but overall effective.3sameSolidly addresses the prompt and catches the traps, but misses providing draft messages for the delayed emails.
fable-5.1 was C3A bit on the longer side, but it covers most of what matters well.3sameCompetent and well-organized, catching all traps and providing all requested messages, but lacks the strategic sharpness of the top outputs.
gpt-6.1-sol was A3On the longer side but overall does a decent job.2-1Overly verbose and relies on hypothetical team members despite the prompt stating the user is an independent consultant.
opus-5 was D2Creative in places, but uneven in length and depth.5+3Catches every trap with exceptional strategic insight and realistic pacing, though it misses a draft message for the delayed non-prospect emails.
gpt-6-astra was G22sameBloated with unnecessary tables and invents external dependencies like collaborators and repair providers, missing the independent persona.
opus-5.5 was E2It may have been cut off in the middle. It did not cover everything, and what it did cover ran long.1-1Hits a reject condition by cutting off mid-sentence at the end of the output.

Course participant hub website

OptionYour scoreYour line on whyJudgeJudge vs youThe judge's reason
opus-5.5 was E5The best look and feel, the most carefully thought-out, the most complete experience3-2Brilliant UX innovation with auto-injecting prompts for mobile users, though it slightly altered the requested session titles.
fable-5.1 was C4Effective and the copy is good.4sameBeautifully designed with clever date-based logic and excellent copy, meeting all requirements perfectly.
gpt-6.1-sol was A3Decent look and feel and navigation.2-1Competent design and good copy, but it failed to use the exact session titles requested.
kimi-k3 was B3Decent look and feel. The navigation is average.2-1Clean layout and good copy, but it failed to use the exact session titles requested.
opus-5 was D2The look and feel did not work for me, and the navigation is hard to use.5+3Flawless execution with exceptional, empathetic copy that perfectly captures the requested tone, paired with a highly polished and distinctive design.
grok-4.7 was F2The look and feel did not work for me, the navigation was unclear, and I was not convinced by the content.3+1Elegant typography and a clever participant pass metaphor, with strong copy and perfect adherence to the brief.
gpt-6-astra was G1The weakest look and feel of the set for me, and the copy was not to my taste.3+2Solid design and good copy that meets all requirements, though the illustrations feel a bit abstract compared to the top entries.