Humanity’s Last PRD
Latest drop: Haiku 5.5 · 8 Oct
Tasks / Growth

Activation & onboarding review

Can the model find the friction that matters most and prioritise the fixes?

Task type v1.1 · 2 tasksLast changed 6 Oct 2026 · ChangelogDifficulty

What AI gets right here, and what you’ll still have to catch

From 16 graded outputs by 8 models. 69% were usable with at most a quick edit.

Reliably right

  1. Respects explicit constraints100% pass
    The output is a review addressed to Omar, well under 800 words, and respects the requested form.
    GPT-6 Luna · API · The checklist isn't the moment
  2. Produces the required deliverable100% pass
    The review covers activation definition, evidence, changes, design proposal, and goal, is addressed to Omar, within length, and is actionable.
    GPT-6 Luna · API · The checklist isn't the moment
  3. Activation defined by what predicts retention100% pass
    Activation is defined by the behaviour that predicts retention (bank+cat 10+), and fixes are judged by their effect on that event.
    GPT-6 Luna · API · The checklist isn't the moment

Where it slips

  1. Uses the supplied evidence correctly44% pass
    The output states 'we have no step-level data for screens 2–8', 'a blank builder is hard for non-analysts', and 'people invite colleagues to something useful, not to a setup screen' as facts, none of which appear in the brief or follow from arithmetic.
    Opus 5.5 · Claude · Analytics tool losing users at setup
  2. Avoids unsupported claims55% pass
    It presents 'a blank builder is hard for non-analysts' and 'people invite colleagues to something useful, not to a setup screen' as established facts without labelling them as hypotheses.
    Opus 5.5 · Claude · Analytics tool losing users at setup
  3. Addresses the actual decision56% pass
    The output commits to a ranked order but does not state what result or condition would change that order, as required.
    GPT-6 Astra · ChatGPT · Analytics tool losing users at setup

How it’s graded

Each check is built on how world-class product and growth leaders do this work, as they’ve described it on Lenny’s Podcast, and names the leader it comes from.

  1. Activation defined by what predicts retention

    Does the review define (or question) the activation event by whether it predicts retention, and judge fixes by their effect on that event rather than on step completion?

    Passes when Names the activation behaviour and its link to retention (from the supplied data, or as an explicit assumption to verify), and ranks fixes by their effect on it.

    Bangaly KabaLenny’s PodcastLauryn IsfordLenny’s Podcast

  2. Gets people to value sooner

    Do the recommended fixes shorten the path to the first moment of value, by removing or deferring steps before it or starting people with templates or sample data, rather than adding tours, tooltips or extra steps?

    Passes when The main fixes cut or defer steps before first value, or give a warm start, and say which steps go.

    Hila QuLenny’s Podcast

Plus our standard checks

Uses the supplied evidence correctly · Addresses the actual decision · Respects explicit constraints · Identifies material uncertainty · Avoids unsupported claims · Produces the required deliverable · Gets the base of every number right · and 2 written for each task, which you’ll see in the tasks below.

The tasks

Read the brief, then compare up to three outputs, each with the LLM judge’s verdict on every check. Highlights mark what a PM had to fix.

The brief

You're the PM for onboarding at Tidepool. Omar, our Head of Product, has asked for a review of onboarding (his note is below). Write it in no more than 800 words: what activation should mean for us, the evidence, what to change in onboarding, and what to make of the design team's proposal and next quarter's goal. The files are attached.

What the model was given4 items: About Tidepool, omar_note.md, onboarding_today.md, first_week_behaviour.csv (August sign-ups: how many who did each thing in their first week were still active at day 60)
About TidepoolA personal finance app: connect your bank, see where your money goes, and budget.
omar_note.md1 lines · Download
From Omar (Head of Product): "Activation is flat at 61%. Review our onboarding and tell me what to fix. The design team has proposed a celebration screen and a progress bar for the checklist."
onboarding_today.md8 lines · Download
# Onboarding today (for 4,000 sign-ups in August)

1. Sign up (100%)
2. Setup checklist: name your account, pick a theme, set a savings goal, enable notifications (61% complete it)
3. Connect a bank (50%)
4. Categorise transactions (34% reach 10)

…
first_week_behaviour.csv (August sign-ups: how many who did each thing in their first week were still active at day 60)first_week_behaviour,retained_at_day_60,churned_by_day_60 All sign-ups,1120,2880 Completed the setup checklist,742,1696 Connected a bank account,964,1036 Categorised 10+ transactions,891,489 Connected a bank AND categorised 10+ transactions,842,238 Created a budget,410,820 Opened the app 4+ days,870,1310
What a strong answer doesThe answer key the graders mark against

Works out retention for each first-week behaviour from the file: all sign-ups 28% (1,120 of 4,000); completed the checklist about 30% (barely better than average); connected a bank about 48%; categorised 10+ transactions about 65%; connected a bank and categorised 10+ about 78%, and that group is about three-quarters of everyone retained (842 of 1,120). So the checklist isn't activation and raising its completion to 75% would do little; activation should be connecting a bank and categorising ten transactions in week one. Recommends cutting or deferring checklist steps that sit before that (theme, notifications, savings goal) and getting people to connect a bank and categorise sooner (for example suggested categories to confirm), and resets the goal to the new activation rate (about 27% today, 1,080 of 4,000). Treats the celebration screen and progress bar as polishing the wrong step. Notes that the data shows correlation and proposes a test to confirm that driving the behaviour improves retention.

Critical failures (cap the score at 40)
  • Recommends raising checklist completion as the main goal
Task

Staff-level · v1.0

The outputs

8 setups answered this brief.

GPT-6 LunaAPI100Usable with a quick editGPT-6 AstraChatGPT92Usable with a quick editGPT-6.1 SolAPI92Usable with a quick editSonnet 5.5API92Usable with a quick editHaiku 5.5API88Usable with a quick editOpus 5.5Claude75Needs more workGemini 3.8 FlashAPI60Needs more workGemini 3.5 Flash-LiteGemini40Needs more work
Gemini 3.5 Flash-LiteGemini

Needs more than a quick edit4 of 11 checks passed

The output misstates the baseline activation rate (34% instead of 27%) and lacks a causation test, so a PM would need to re-check the numbers and add a test plan.

Output

To: Omar, Head of Product From: PM, Onboarding Date: September 2024 Subject: Onboarding Review & Activation Strategy

---

1. What Activation Should Mean for Us

Currently, our definition of activation—completing a checklist of cosmetic preferences (name, theme, savings goal, notifications)—is flawed. It measures vanity completion rather than value realization.

As a personal finance app, Tidepool delivers core value when users connect their financial reality and engage with it. Looking at our August cohort (4,000 sign-ups), the data tells a clear story about Day 60 retention:

  • Checklist completers retain at 30.4% (742/2,438).
  • Bank connectors retain at 48.2% (964/2,000).
  • Users who categorise 10+ transactions retain at 64.5% (891/1,380).
  • Users who do BOTH (Connect + Categorise) retain at 78.0% (842/1,080).

Recommendation: We must redefine "Activation" from “completed the setup checklist” to “Connected a bank AND categorised at least 10 transactions within their first week.” This anchors our success metric to actual product utility, shifting our focus from cosmetic setup to establishing long-term financial habits.

---

2. The Evidence: Where We Are Losing Users

Our current funnel drops off steeply: 1. Sign up: 4,000 (100%) 2. Setup checklist: 2,440 (61%) 3. Connect a bank: 2,000 (50%) 4. Categorise 10 transactions: 1,360 (34%)

The biggest leak is bank connection (dropping 11 percentage points right after the checklist) and transaction categorisation (dropping another 16 points).

Interestingly, the checklist itself acts as a barrier. Users are spending cognitive load on naming accounts and picking themes before experiencing the core product. Meanwhile, power features like "Created a budget" (33% retention) and opening the app 4+ days are strong secondary indicators, but bank linking and categorisation are the undisputed gateways to retention.

---

3. What to Change in Onboarding

To drive our new definition of activation, we need to restructure the first-time user experience (FTUX):

  • Kill the upfront setup checklist: Strip out non-essential steps (naming accounts, picking themes, notifications) from the critical path. Move them to profile settings or prompt for them after value has been proven.
  • Accelerate time-to-bank: Move bank connection immediately following sign-up. Users downloaded Tidepool to see their money; let them see it instantly via Plaid/Aggregator integration.
  • Gamify the "aha" moment: Guide users smoothly from bank connection directly into categorising their first 10 transactions. Use smart auto-categorisation with a clear, frictionless bulk-approval or swipe UI to make reaching the 10-transaction milestone effortless.

---

4. Verdict on the Design Proposal and Next Quarter's Goal

The Design Team’s Proposal (Celebration screen + Progress bar): * Verdict: Reject in its current form. Adding a progress bar and celebration screen for the existing checklist optimizes the wrong metric. It puts lipstick on a superficial activation loop. * Pivot: If we redesign onboarding to focus on connecting a bank and categorising transactions, then we can use a celebration screen and progress bar—not to celebrate picking a theme, but to celebrate linking an account and hitting that crucial 10-transaction milestone.

Next Quarter's Goal (Raise checklist completion from 61% to 75%): * Verdict: Abandon this goal. Optimizing checklist completion is a local maximum that misaligns product and user success. Chasing 75% checklist completion will likely result in higher drop-offs further down the funnel and flat-to-declining Day 60 retention. * New Goal: We should replace this with a target to raise our true activation rate (Bank Connected + 10 Transactions Categorised) from the current baseline (~34%) to 50% next quarter. This directly correlates with our actual business health and user retention.

Claims the judge couldn’t find in the brief

  • The checklist itself acts as a barrier.
  • Users are spending cognitive load on naming accounts and picking themes before experiencing the core product.
  • Bank linking and categorisation are the undisputed gateways to retention.

Grades and run details

Decision model 50 · LLM judge 33
Decision model checks
  • failUses the supplied evidence correctly41%
  • passAddresses the actual decision53%
  • passRespects explicit constraints33%
  • failIdentifies material uncertainty87%
  • failAvoids unsupported claims33%
  • passProduces the required deliverable48%
  • partialComputes retention by behaviour25%
  • failTests before trusting the correlation98%
  • passActivation defined by what predicts retention99%
  • failGets the base of every number right55%
  • passGets people to value sooner93%
Run
Run
#1
Time to output
5 s
Submitted
5 Oct 2026
GPT-6 AstraChatGPT

Usable with a quick edit9 of 11 checks passed

The review is clear, evidence-based, and covers all required points; a PM could act on it with only minor wording tweaks.

Output

Onboarding review: activate on financial value, not setup

Recommendation: Replace checklist completion as our activation metric with connecting a bank and categorising at least 10 transactions within seven days of signup. Reorient onboarding around that outcome. The celebration screen and progress bar are secondary; making the current checklist more engaging risks improving the wrong metric.

What activation should mean

For Tidepool, activation should mean a user has taken the first meaningful steps toward understanding their spending—not finished configuring the app.

Bank connection brings their financial data into Tidepool; categorisation makes that data useful. Together, these are our strongest available proxy for experiencing core value. This is a working definition, not proof that exactly 10 categorisations creates a lasting habit. We should validate the threshold and whether users actually understand something useful about their money.

Budget creation can follow that first insight. The evidence does not support making it a mandatory activation requirement today.

What the evidence says

Among August’s 4,000 signups, overall day-60 retention was 28%.

First-week behaviourUsersDay-60 retention
------:---:
Completed setup checklist2,43830.4%
Connected a bank2,00048.2%
Categorised 10+ transactions1,38064.6%
Connected a bank and categorised 10+1,08078.0%
Created a budget1,23033.3%
Opened the app on 4+ days2,18039.9%

The checklist has weak separation: completers retained at 30.4%, versus 24.2% for non-completers. Its 61% completion rate is therefore a poor headline measure of delivered value.

By contrast, only 27% of signups connected a bank and categorised 10+ transactions, but they accounted for 842 of 1,120 retained users—75% of everyone retained. Users who did not achieve both retained at 9.5%.

These are observational associations. More motivated users may both complete these actions and remain active; we cannot promise 78% retention simply by pushing people through them. Behaviour groups also overlap, so the reported steps should not be treated as a strictly sequential funnel. Before testing, reconcile the rounded onboarding figures with event counts and standardise the day-60 activity definition.

What to change

  1. Remove configuration as a gate to value. Defer theme selection and account naming; make savings goals optional. Request notification permission when there is a clear benefit, such as a relevant budget alert—not as checklist work.
  1. Bring bank connection forward. Explain why access is needed, what data is used, and how users retain control. Instrument connection starts, provider errors, unsupported banks, abandonment and successful imports. Preserve a skip/retry path rather than forcing consent.
  1. Guide users from imported transactions to spending insight. Offer a short, low-effort review of suggested categories. Let users confirm or correct them, then show an understandable spending summary. Do not require ten arbitrary manual edits merely to satisfy the metric.
  1. Offer a budget as the next useful action. Use the spending summary to suggest a realistic starting budget, rather than asking users to invent one before seeing their finances.

Test the redesigned flow against today’s onboarding. Measure seven-day activation first, then downstream retention; monitor connection failures, abandonment and user-reported trust and usefulness.

Design proposal and next quarter’s goal

A progress bar is worth testing if it tracks meaningful steps toward understanding spending. It should not encourage completion of low-value setup tasks. A celebration screen belongs after a genuine milestone and should reinforce the insight gained, not interrupt access to it. Neither should take priority over bank-connection and categorisation friction.

Replace the 61% → 75% checklist goal with a goal for the new activation definition. I propose 27% → 35% within seven days, provisionally, subject to instrumentation validation and early experiment results. That is a planning target—not a forecast supported by this dataset.

Keep checklist completion as a diagnostic. Judge success by increased meaningful activation without sacrificing trust or retention, and confirm the retention effect once next quarter’s cohorts mature.

Grades and run details

Decision model 82 · LLM judge 100
Decision model checks
  • failUses the supplied evidence correctly90%
  • passAddresses the actual decision86%
  • passRespects explicit constraints30%
  • passIdentifies material uncertainty83%
  • passAvoids unsupported claims43%
  • passProduces the required deliverable42%
  • passComputes retention by behaviour32%
  • passTests before trusting the correlation97%
  • passActivation defined by what predicts retention99%
  • failGets the base of every number right82%
  • passGets people to value sooner78%
Run
Run
#1
Time to output
30 s
Submitted
5 Oct 2026
Gemini 3.8 FlashAPI

Needs more than a quick edit6 of 11 checks passed

The output lacks a causation caveat and test proposal, and makes unsupported causal claims; a PM would need to add those before using it.

Output

To: Omar, Head of Product From: PM, Onboarding Subject: Onboarding Review & Activation Strategy

---

Executive Summary

Our current activation rate (61%) is flat because we are measuring and optimizing the wrong milestone. Completing our setup checklist barely moves the needle on long-term retention.

To drive retention, we must redefine activation around core product value—connecting a bank and categorising transactions—and reorient our onboarding funnel and quarterly goals to support it.

---

1. What Activation Should Mean (and the Evidence)

Activation should represent the “Aha!” moment: the point where a user experiences Tidepool’s core value proposition and forms a lasting habit.

Currently, we define activation as completing the setup checklist (name account, pick theme, set savings goal, enable notifications). The data shows this is a vanity metric:

  • Baseline Day 60 Retention: 28.0% (1,120 / 4,000 sign-ups)
  • Checklist Completion: 30.4% Day 60 retention (742 / 2,438) — a negligible +2.4 percentage point lift over baseline.

By contrast, actions tied to our core utility (seeing where money goes) dramatically predict retention: * Connected a bank account: 48.2% Day 60 retention (964 / 2,000) * Categorised 10+ transactions: 64.6% Day 60 retention (891 / 1,380) * Connected bank AND categorised 10+ transactions: 78.0% Day 60 retention (842 / 1,080)

Users who complete both core actions retain at nearly triple the baseline rate. In fact, this group accounts for 75.2% of all retained users at Day 60 (842 of 1,120), despite representing only 27% of initial sign-ups.

New Activation Metric: > A user connects a bank account and categorises at least 10 transactions within their first 7 days.

---

2. Design Proposal & Next Quarter’s Goal

#### The Design Proposal (Celebration Screen & Progress Bar) The design team’s proposal is solving the wrong problem. Adding progress bars and celebration screens to cosmetic tasks (picking themes, naming accounts) creates artificial friction before users experience any real value. We are gamifying administrative overhead instead of utility.

Recommendation: Repurpose the design team’s work, but attach it to the real value milestone (e.g., celebrating when their first 10 transactions are imported and categorized).

#### Next Quarter’s Goal (Raise Checklist Completion from 61% to 75%) We should retire this goal immediately.

Even if we hit 75% checklist completion, model projections indicate Day 60 retention would increase by less than 1%. Optimizing an arbitrary checklist delivers zero business impact.

Proposed New Goal for Next Quarter: > Increase the core Activation Rate (Bank Connected + 10 Categorised Transactions) from 27% to 40% of sign-ups.

---

3. What to Change in Onboarding

Our onboarding funnel currently bleeds users before they experience core value: 50% drop off before connecting a bank, and only 34% categorize 10 transactions. We must clear the runway to value:

1. Eliminate the Cosmetic Checklist from First-Run Flow * Move “pick a theme” and “name your account” to Settings/Profile. * Defer “enable notifications” and “set a savings goal” until after the user sees their financial data.

2. Lead with Bank Connection * Prompt bank linking immediately post-registration. * Pre-frame the security and the payoff clearly: “Connect your account to see your last 30 days of spending instantly.”

3. Accelerate Time-to-Value (TTV) on Categorisation * Upon successful bank connection, immediately import transactions and guide the user through a rapid, swipeable 10-transaction categorisation flow (or auto-categorise and prompt a one-tap review). * Conclude with an instant visual breakdown: “Here’s where your money went this month.” This is Tidepool’s true “Aha!” moment.

4. Contextual Budgeting * Only prompt users to create a budget after categorisation. Budgeting retention is low in week one (33.3%) because users cannot budget accurately without understanding their baseline spend.

Next Steps

  1. Align leadership on the new Activation definition.
  2. Re-scope the design sprint from checklist gamification to a streamlined "Connect & Review 10 Transactions" flow.
  3. Update the team OKR dashboard for next quarter.

Grades and run details

Decision model 68 · LLM judge 58
Decision model checks
  • failUses the supplied evidence correctly24%
  • passAddresses the actual decision56%
  • passRespects explicit constraints49%
  • failIdentifies material uncertainty71%
  • partialAvoids unsupported claims31%
  • passProduces the required deliverable68%
  • passComputes retention by behaviour71%
  • failTests before trusting the correlation95%
  • passActivation defined by what predicts retention100%
  • passGets the base of every number right19%
  • passGets people to value sooner84%
Run
Run
#1
API response time
20 s
Submitted
6 Oct 2026

Check by check

Where the outputs differ comes first. Open a check to read the judge’s reasons. “Mixed” means the two graders disagreed.

Where they differ 7

Uses the supplied evidence correctlyWrongMixedMixed
Gemini 3.5 Flash-Lite · Gemini

The output states 'the checklist itself acts as a barrier' and similar causal claims as facts, which are not in the supplied context and do not follow by arithmetic.

GPT-6 Astra · ChatGPT

All factual claims about the current situation are directly supported by the supplied context or correct arithmetic.

Gemini 3.8 Flash · API

All statements about the current situation are directly from the supplied context or correctly derived from it.

Addresses the actual decisionMixedRightMixed
Gemini 3.5 Flash-Lite · Gemini

The output does not state what result or condition would change its recommendation, so it fails to commit with a clear change condition.

GPT-6 Astra · ChatGPT

Commits early to a clear activation definition and changes, and states that validation and experiment results would change the call.

Gemini 3.8 Flash · API

The output commits to a clear answer but does not state what result or condition would change that answer, as required.

Identifies material uncertaintyWrongRightWrong
Gemini 3.5 Flash-Lite · Gemini

The output does not name any unknowns that could change the decision or how they would be resolved.

GPT-6 Astra · ChatGPT

Names the observational nature, threshold uncertainty, and need for validation; says how to resolve them.

Gemini 3.8 Flash · API

The output does not name any unknowns that could change the decision or say how they would be resolved.

Avoids unsupported claimsWrongRightWrong
Gemini 3.5 Flash-Lite · Gemini

Interpretations like 'the checklist itself acts as a barrier' and 'undisputed gateways' are presented as established fact without being labelled as hypotheses.

GPT-6 Astra · ChatGPT

Hypotheses and forecasts are clearly labelled as such, and confident claims are supported by the evidence.

Gemini 3.8 Flash · API

It presents causal claims (e.g., checklist steps create artificial friction, low budgeting retention is because users lack baseline spend) as fact without labelling them as hypotheses.

Computes retention by behaviourWrongRightRight
Gemini 3.5 Flash-Lite · Gemini

The output does not compute the overall average retention (28%) and does not compare the subgroup rates to it.

GPT-6 Astra · ChatGPT

Computes day-60 retention for each first-week behaviour and compares with the 28% average.

Gemini 3.8 Flash · API

Day-60 retention rates are correctly computed: checklist ~30%, bank+categorise ~78%, average 28%.

Tests before trusting the correlationWrongRightWrong
Gemini 3.5 Flash-Lite · Gemini

The output does not note that the retention differences are correlations and does not propose a test to confirm causation.

GPT-6 Astra · ChatGPT

Notes the correlation risk and proposes a specific test (redesigned flow vs. today, measuring retention) to check causation.

Gemini 3.8 Flash · API

The output does not note that the observed relationships are correlations and does not propose a test to drive the behaviour and measure retention.

Gets the base of every number rightWrongMixedRight
Gemini 3.5 Flash-Lite · Gemini

The baseline for the new activation rate is given as ~34%, but the correct rate for 'connected bank AND categorised 10+' is 1,080/4,000 = 27%.

GPT-6 Astra · ChatGPT

All percentages are computed from the correct denominators, and bases are stated or clear from context.

Gemini 3.8 Flash · API

All percentages and rates use the correct denominators from the data, and the bases are clear.

All got right 4

Respects explicit constraintsRightRightRight
Gemini 3.5 Flash-Lite · Gemini

The output is a memo to Omar, under 800 words, and addresses the requested topics.

GPT-6 Astra · ChatGPT

The review is under 800 words, addresses Omar, and covers all requested elements.

Gemini 3.8 Flash · API

The output is a memo to Omar, under 800 words, and addresses all requested elements.

Produces the required deliverableRightRightRight
Gemini 3.5 Flash-Lite · Gemini

The review is complete, in memo form, within the word limit, and a PM could act on it with light edits.

GPT-6 Astra · ChatGPT

The output is a complete review with recommendation, evidence, changes, and view on proposal and goal, usable as is.

Gemini 3.8 Flash · API

The review is complete, in memo form, within the word limit, and a PM could act on it with light edits.

Activation defined by what predicts retentionRightRightRight
Gemini 3.5 Flash-Lite · Gemini

Activation is redefined based on retention prediction (bank + categorise), and fixes are judged by their effect on that event.

GPT-6 Astra · ChatGPT

Defines activation by retention (bank + categorise 10+) and ranks fixes by their effect on that event.

Gemini 3.8 Flash · API

Activation is defined by the behaviour that predicts retention (bank connect + 10 categorised), and fixes are ranked by their effect on reaching that event.

Gets people to value soonerRightRightRight
Gemini 3.5 Flash-Lite · Gemini

The main fixes cut or defer checklist steps before value and move bank connection and categorisation earlier.

GPT-6 Astra · ChatGPT

Main fixes cut or defer configuration steps before value and bring bank connection and categorisation forward.

Gemini 3.8 Flash · API

The main fixes cut or defer steps before first value (cosmetic checklist, notifications, savings goal) and lead directly to bank connection and categorisation.

Results

Every setup we’ve tested on this task type, across all its tasks and repeats, graded on the current checklist. Calibrated. Where our PM checked the graders by hand, they agreed on 79% of 53 checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1GPT-6 AstraChatGPT85.995.52None
2Sonnet 5.5API78.4100.02None
3GPT-6.1 SolAPI85.990.92None
4GPT-6 LunaAPI90.081.82None
5Haiku 5.5API75.986.72None
6Opus 5.5Claude88.478.42None
7Gemini 3.8 FlashAPI66.651.92None
8Gemini 3.5 Flash-LiteGemini52.539.42None

What you get for the cost

GPT-6 Luna · API scores within 5 points of GPT-6 Astra · ChatGPT at about a hundredth of the cost.

Task score against relative cost for each setup. The table below has the same figures.304050607080901001×3×10×30×100×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreGPT-6 Astra · ChatGPT: task score 91, 122× the cheapest (1 of 2 outputs estimated)Sonnet 5.5 · API: task score 90, 32× the cheapestGPT-6.1 Sol · API: task score 89, 15× the cheapestGPT-6 Luna · API: task score 86, the cheapestHaiku 5.5 · API: task score 82, 1.8× the cheapestOpus 5.5 · Claude: task score 82, 101× the cheapest (1 of 2 outputs estimated)Gemini 3.8 Flash · API: task score 56, 13× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 44, 3.4× the cheapest (1 of 2 outputs estimated)GPT-6 Astra · ChatGPTSonnet 5.5 · APIGPT-6.1 Sol · APIGPT-6 Luna · APIHaiku 5.5 · APIOpus 5.5 · ClaudeGemini 3.8 Flash · APIGemini 3.5 Flash-Lite · Gemini
Best valueNobody beats these on both score and costCost partly estimated (app runs with no usage meter)

Cost is each output’s tokens at the model’s list price, shown relative to the cheapest setup. Runs done in an app have no usage meter, so theirs are estimated from the same setup’s measured runs. How we compare cost

The PM job

Reviewing a signup and onboarding flow that is losing users.

Why it matters

Anyone can list fifty UX nits. The job is finding the two that explain the drop-off, backed by the funnel data supplied.

What good looks like

  • Ties each issue to the funnel data
  • Prioritises by likely impact
  • Defines activation as the behaviour that predicts retention, not finishing onboarding

Deliberately not measured

  • Accessibility audit completeness
  • Visual redesign
Capability tested

Consequential critique

The failure we’re looking for

A generic UX checklist

Grading

Decision model and LLM judge, calibrated against a blind PM review