Tasks / Metrics & Experimentation

Pick a North Star metric

Can the model pick the one metric that measures the value a product creates, and break it into inputs teams can move?

Task type v1.1 · 2 tasksLast changed 5 Oct 2026 · ChangelogDifficulty

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

From 16 graded outputs by 8 models. 75% were usable with at most a quick edit.

Reliably right

  1. Says how it could be gamed100% pass
    Identifies gaming vectors (tapping through, notification overuse, short-term spikes, revenue manipulation, MAU drift) and sets specific guardrails (quality completion, push caps, blocking metrics, cohort retention checks, subscriber churn tracking).
    Sonnet 5.5 · API · Revenue, actives, or something else?
  2. Sees the growth is narrow100% pass
    It shows the split (chains up, independents' filled shifts and fill rate down, more leaving) and tracks independent venues via segment reporting and guardrails.
    GPT-6 Luna · API · Growing, on paper
  3. Checks the North Star can be measured100% pass
    It flags the 64% confirmation gap and proposes instrumenting check-in/out, reconciliation, and auditing to fix it before the board readout.
    GPT-6 Luna · API · Growing, on paper

Where it slips

  1. Identifies material uncertainty55% pass
    It identifies the confirmation gap as an unknown but does not say how resolving it could change the North Star decision.
    GPT-6.1 Sol · API · Growing, on paper
  2. Addresses the actual decision58% pass
    The output commits to a North Star but does not state what result or condition would change that recommendation.
    GPT-6 Luna · API · Growing, on paper
  3. Avoids unsupported claims67% pass
    It presents 'easy to tap through' and 'we don't track today' as facts without support, and does not label them as assumptions.
    Opus 5.5 · Claude · Revenue, actives, or something else?

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. Measures value created, not captured

    Does the chosen North Star measure the value customers get from the product, rather than revenue or a count of people who merely showed up?

    Passes when Picks a metric of value delivered to customers (a core action that predicts retention), and keeps revenue as a separate business measure.

  2. Breaks it into inputs teams can move

    Does the output break the North Star into a few input metrics, each one a team could own and move week to week, and pair it with a separate business measure of value captured?

    Passes when Gives a small tree of input metrics, each with an owner and a link to the North Star, plus the revenue or business measure it's paired with.

  3. Says how it could be gamed

    Does the output pressure-test the North Star for how teams could raise it without delivering more value, and set guardrail metrics against that?

    Passes when Names a realistic way the metric could be gamed in this product and the guardrails that would catch it.

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 1 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 Head of Product at Simmer. Our CEO wants MRR as the company's North Star metric; our growth lead wants monthly active users. Write a recommendation for both of them, in no more than 700 words: the North Star you'd pick and why, three or four input metrics teams could own, and the guardrails that stop it being gamed. What we know is below.

What the model was given5 items: About Simmer, What people do, Retention (12 months of cohorts), Subscriptions, Last year's notification test
About SimmerA recipe and meal-planning app. Free to use; Simmer Plus ($6 a month) unlocks meal plans and Cook Mode's step-by-step timers for every recipe. 2.1M monthly active users, 140,000 paying subscribers, MRR of $840,000.
What people doSave a recipe (savers save 4 a week on average). Cook a recipe in Cook Mode, logged when someone reaches the last step. Plan a week of meals (7% of monthly actives have ever done it). 68% of monthly actives only browse: they arrive from search, read one recipe and leave.
Retention (12 months of cohorts)Users who cook two or more recipes in Cook Mode in a week: 61% still active at month 6. Weekly planners: 54%. Users who save recipes but never cook one in Cook Mode: 9%. Browse-only users: 4%.
Subscriptions82% of new subscribers had cooked at least three recipes in Cook Mode in their first two weeks.
Last year's notification testA daily 'recipe of the day' push notification raised monthly active users by 6%. Uninstalls doubled over the same period, and Cook Mode completions didn't change.
What a strong answer doesThe answer key the graders mark against

Recommends a North Star built on cooking, such as weekly cooks (people who complete two or more recipes in Cook Mode in a week) or recipes cooked per week, because cooking is the value moment and it predicts both retention (61% at month 6) and subscribing (82% of new subscribers cooked first). Rejects MRR as the North Star because it measures value captured, and lags (keep it as the business measure the North Star is paired with). Rejects monthly actives because 68% of them only browse (4% retention) and the push test shows it can be inflated while the product gets worse. Gives input metrics a team could own, such as first cook within a week of signing up, cooks per active cook, and share of savers who cook what they saved, and guardrails such as uninstalls and notification opt-outs, and subscriber retention.

Critical failures (cap the score at 40)
  • Recommends MRR or monthly active users as the North Star
Task

Mid-level · v1.0

The outputs

8 setups answered this brief.

Usable with a quick edit11 of 11 checks passed

The output is well-structured, uses evidence correctly, and covers all required elements; a PM could use it with only minor wording tweaks, without re-checking evidence or redoing reasoning.

Output

To: CEO, Growth Lead From: Head of Product Re: Simmer's North Star

Recommendation

Make the North Star Weekly Active Cookers (WAC): users who complete Cook Mode on two or more recipes in a week.

Both of your metrics matter, but neither measures the thing that drives the business.

Why not MAU. 68% of our 2.1M monthly actives arrive from search, read one recipe and leave. They retain at 4% at month 6. Last year's push test shows the risk: MAU rose 6%, uninstalls doubled, and Cook Mode completions didn't move. A MAU target rewards exactly that trade.

Why not MRR. MRR ($840K, from 140K subscribers at $6) is the outcome we want, but it lags. It can also be moved by a price change or a paywall tweak without making the product better, and it tells a team nothing about what to do this week.

Why WAC. People who cook twice or more a week retain at 61% at month 6, against 9% for savers who never cook and 4% for browsers. And 82% of new subscribers cooked at least three recipes in Cook Mode in their first two weeks. Cooking sits upstream of both retention and revenue, so growing WAC should grow MRR, and teams can influence it within a sprint.

MRR stays as our top-line business outcome, reported alongside WAC. MAU becomes a funnel-health metric, since it is the pool WAC is converted from.

One caveat: these are correlations, and committed cooks may simply be our best users anyway. In the first quarter we should run a holdout experiment, such as a stronger first-week cooking prompt, to confirm that causing more cooking causes better retention and conversion.

Input metrics teams can own

  1. New-user activation (Onboarding): % of new users who complete a Cook Mode recipe within 7 days, with the 3-recipes-in-14-days rate tracked as the subscription-intent signal.
  2. Save-to-cook rate (Discovery and Cook Mode): % of savers who cook a saved recipe in Cook Mode within 14 days. Savers who never cook retain at 9%, so this is our biggest leak.
  3. Browse-to-return conversion (Growth and SEO): % of search-landing visitors who save a recipe or come back within 7 days. This gives Growth a real lever on the 68%, and on MAU quality.
  4. Planning adoption (Planning): % of monthly actives who build a weekly plan. Only 7% have ever done it, yet weekly planners retain at 54%.

Guardrails against gaming

Against cheap cooking events. Completion is logged when someone reaches the last step, so tapping through counts. - Track a "quality completion" (minimum elapsed time relative to recipe length, or timer use) and report WAC on that basis. - Monitor median time per completion for sudden drops. - Add an occasional "Did you cook it?" prompt to calibrate against self-report.

Against the notification trap. - Set a hard cap on push frequency. - Treat uninstall rate and notification opt-outs as blocking metrics: any experiment that worsens either is not shipped, whatever it does to WAC or MAU.

Against short-term lifts. - Judge any WAC gain by its 30- and 90-day cohort retention, so we don't buy a spike with users who disappear.

Against revenue gaming. - Track subscriber churn, refund rate and cancellation-flow completion alongside MRR. - Don't count MRR gains from price changes or paywall aggressiveness as product wins. - Don't gate Cook Mode so tightly that new users can't reach the 3-recipe habit that predicts subscribing.

Against MAU drift. - Keep reporting MAU, but as a watch metric with no team target. If MAU rises while WAC is flat, we treat it as a warning sign rather than a win.

What each of you gets

  • CEO: a metric that leads MRR and that teams can act on, with MRR still reported at the top. If WAC grows and conversion holds, MRR follows.
  • Growth lead: ownership of the biggest untapped pool, the browse-only majority. Success is measured by how many of them become cookers, not by how many we can get to open the app.

I'd propose we adopt WAC from next quarter, baseline it in the first month, set targets once we see the quality-adjusted number, and review the causal experiment results at the end of the quarter.

Grades and run details

Decision model 95 · LLM judge 100
Decision model checks
  • passUses the supplied evidence correctly57%
  • passAddresses the actual decision93%
  • passRespects explicit constraints46%
  • passIdentifies material uncertainty75%
  • partialAvoids unsupported claims24%
  • passProduces the required deliverable94%
  • passAnswers both proposals with the data100%
  • passMeasures value created, not captured100%
  • passGets the base of every number right78%
  • passBreaks it into inputs teams can move100%
  • passSays how it could be gamed100%
Run
Run
#1
API response time
22 s
Submitted
5 Oct 2026

Needs more than a quick edit7 of 11 checks passed

The output lacks a condition that would change the recommendation, omits unknowns, makes unsupported causal claims, and does not explicitly pair the North Star with a business measure. A PM would need to rework these gaps.

Output

To: CEO & Growth Lead From: Head of Product Subject: Aligning on Our North Star Metric

Choosing between Monthly Active Users (MAU) and Monthly Recurring Revenue (MRR) is a classic tension between reach and monetization. But choosing either will hurt Simmer.

Here is why neither works as our North Star, what we should measure instead, and how we keep the business honest.

---

Why Not MAU or MRR?

  • MAU is a vanity metric that invites bad behavior. 68% of our monthly actives are transient visitors who read one recipe from Google and bounce, retaining at a dismal 4% by month six. Last year’s push notification test proved MAU is dangerous: blast notifications spiked MAU by 6%, but Cook Mode usage didn’t budge and uninstalls doubled. Optimizing for MAU encourages low-intent traffic acquisition and spammy retention hacks.
  • MRR is a lagging output. MRR tells us if we extracted value last month, not if we created value today. Optimizing directly for MRR leads teams to tighten paywalls, increase discounts, or push annual locks—tactics that burn long-term trust without fixing the core product engine.

---

The Proposed North Star: Weekly Active Cooks (WAC)

Definition: The number of unique users who complete at least one recipe in Cook Mode in a rolling 7-day window.

A great North Star sits at the intersection of customer value and enterprise value. For Simmer, that intersection is Cook Mode:

  1. It predicts retention: Users who cook $\ge$2 recipes a week retain at 61% at month 6 (compared to 9% for passive recipe savers and 4% for browsers).
  2. It fuels monetization: 82% of new subscribers completed at least three Cook Mode sessions in their first two weeks. Cook Mode is the direct gateway to our $6/month Simmer Plus subscription.

If Weekly Active Cooks grows sustainably, MAU becomes higher-intent and MRR growth naturally follows.

---

Team Input Metrics

Product, Growth, and Content teams cannot easily move WAC directly; they need operational inputs they can own day-to-day:

1. Browse-to-Cook Activation Rate (New User Experience & Growth): % of first-time visitors who enter Cook Mode within 7 days of their first session. Goal: Move users out of the 68% bounce bucket and into an interactive cooking experience. 2. Multi-Cook Frequency (Core Product): % of active cooks completing $\ge$2 Cook Mode sessions per week. Goal: Deepen habit formation to hit the threshold associated with our 61% 6-month retention rate. 3. Weekly Meal Plan Adoption (Engagement & Feature Hook): % of monthly actives who schedule $\ge$3 recipes into a weekly meal plan. Goal: Only 7% have ever tried meal planning, yet planners show 54% month-6 retention. Meal planning provides the forward-looking intent that drives repeat Cook Mode completions. 4. Qualified Cook-to-Subscriber Conversion (Monetization): Conversion rate to Simmer Plus among users who have completed their 3rd Cook Mode recipe. Goal: Target our paywalls and free-trial prompts precisely when users reach the high-intent threshold (the 82% cohort), rather than hitting casual browsers with hard walls.

---

Guardrail Metrics (To Prevent Gaming)

To ensure teams do not inflate WAC artificially, we must establish three non-negotiables:

  • Cook Mode Session Integrity: Cook Mode completions are only counted if the user spends at least 5 minutes in the mode with step-progression intervals matching real cooking behavior. This prevents growth teams from auto-completing sessions via deep links.
  • Net Churn & Uninstall Rate: Every experiment that touches engagement or notifications must track app uninstalls and notification opt-outs. Any test that increases uninstalls (like the push notification failure) is an automatic rollback, regardless of active-cook uplift.
  • Monthly Cohort Renewal Rate: We must monitor 30- and 90-day retention of new Simmer Plus subscribers to ensure we are not driving low-intent conversions through deceptive trial dark-patterns.

---

By rallying around Weekly Active Cooks, we bridge the gap between audience growth and revenue: we build a sticky, habitual product first, and harvest the revenue second.

Grades and run details

Decision model 86 · LLM judge 58
Decision model checks
  • passUses the supplied evidence correctly32%
  • passAddresses the actual decision51%
  • passRespects explicit constraints38%
  • failIdentifies material uncertainty95%
  • partialAvoids unsupported claims60%
  • passProduces the required deliverable85%
  • passAnswers both proposals with the data94%
  • passMeasures value created, not captured95%
  • passGets the base of every number right68%
  • passBreaks it into inputs teams can move48%
  • passSays how it could be gamed99%
Run
Run
#1
API response time
22 s
Submitted
5 Oct 2026
Critical failure: Recommends MRR or monthly active users as the North Star. Score capped at 40.

Needs more than a quick edit2 of 11 checks passed

The output recommends MRR as a North Star, violating the core constraint, and contains factual errors (4-49% retention, 7% weekly actives) that would require re-checking evidence and redoing reasoning.

Output

To: CEO & Growth Lead From: Head of Product, Simmer Date: October 24, 2023 Subject: Recommendation: Our North Star Metric, Inputs, and Guardrails

Team, a great debate highlights our current tension: the CEO wants revenue (MRR), while the Growth lead wants reach (MAU).

If we choose MRR, we risk ignoring top-of-funnel health. If we choose MAU, we risk celebrating vanity metrics—as proved by last year’s notification test, where a 6% bump in MAU doubled uninstalls because those users didn't actually find value.

My recommendation is to pick a hybrid North Star: Weekly Active Subscribers (WAS), or more specifically, Paid Users Who Cook Weekly.

However, since a single metric rarely satisfies executive reporting, MRR should remain our ultimate business North Star, but Product Value Realization (Weekly Actives who Cook) must be the operational North Star that guides day-to-day product and growth decisions. MRR is the lagging reward; cooking engagement is the leading engine. Here is why, along with our inputs and guardrails.

Why This North Star Works

Our data reveals a stark truth: 68% of our 2.1M MAUs are "browse-only" users with a catastrophic 4-49% retention rate. MAU rewards us for acquiring transient browsers. Meanwhile, users who experience core value—using Cook Mode—exhibit a stellar 61% six-month retention rate, and 82% of our subscribers cook at least three recipes in their first two weeks before paying.

If we optimize for Weekly Active Cookers, we bridge both worlds: we drive the habit loops that naturally convert free users to Simmer Plus, which directly compounds our $840k MRR, while ignoring empty traffic that drains server costs and drives uninstalls.

Three Input Metrics Teams Can Own

To move our North Star, different pods will own specific, actionable input metrics:

  1. Activation Velocity (Growth & Onboarding Pod): The percentage of new signups who complete at least one recipe in Cook Mode within their first 7 days. (Tied to the 82% subscriber conversion rule).
  2. Core Habit Frequency (Engagement Pod): The average number of Cook Mode completions per active user per week. (Tied to the 61% retention milestone).
  3. Feature Expansion (Core Product Pod): The percentage of weekly actives utilizing meal-planning features. (Currently a low 7%, representing massive upside for long-term retention).

Guardrails to Prevent Gaming

Metrics are easily corrupted if incentives misalign. To ensure we don’t repeat the ghost of last year’s push-notification debacle, we will institute three hard guardrails:

  • The Uninstall Rate Cap: If marketing or growth initiatives drive MAU/signups up, but the monthly uninstall rate exceeds 2% of our active base, campaigns are automatically paused.
  • The "No Empty Calories" Rule for Growth: Growth experiments cannot be measured solely by acquisition or click-through rates. An experiment is deemed a failure if it increases top-of-funnel signups without a corresponding lift in Day-7 Cook Mode completions.
  • Subscription Quality (LTV/Churn Check): To prevent the sales team or pricing experiments from pushing low-intent, high-churn subscriptions just to spike short-term MRR, we will track 90-day subscriber retention. If a cohort churns above baseline, the acquisition channel or incentive driving it is depreciated.

By anchoring on users who actually cook, we align the CEO’s revenue goals with the Growth lead's engagement goals, turning casual browsers into loyal, paying home cooks.

Claims the judge couldn’t find in the brief

  • browse-only users have a 4-49% retention rate
  • the percentage of weekly actives utilizing meal-planning features is currently a low 7%

Grades and run details

Decision model 41 · LLM judge 17
Decision model checks
  • failUses the supplied evidence correctly71%
  • partialAddresses the actual decision72%
  • failRespects explicit constraints91%
  • failIdentifies material uncertainty90%
  • failAvoids unsupported claims33%
  • partialProduces the required deliverable32%
  • passAnswers both proposals with the data35%
  • partialMeasures value created, not captured20%
  • failGets the base of every number right40%
  • passBreaks it into inputs teams can move94%
  • passSays how it could be gamed54%
Run
Run
#1
Time to output
5 s
Submitted
5 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 10

Uses the supplied evidence correctlyRightRightWrong
Sonnet 5.5 · API

All statements about the current situation are taken directly from the supplied context, with no invented facts.

Gemini 3.8 Flash · API

All factual numbers and statements about the current situation are taken correctly from the brief.

Gemini 3.5 Flash-Lite · Gemini

States browse-only retention as '4-49%' when the brief gives 4%, and claims 7% of weekly actives use meal planning when the brief says 7% of monthly actives ever.

Addresses the actual decisionRightMixedWrong
Sonnet 5.5 · API

Commits unambiguously to Weekly Active Cookers as the North Star, addresses both the CEO and Growth Lead, and states that a holdout experiment would confirm causality and could change the call.

Gemini 3.8 Flash · API

The output commits to Weekly Active Cooks but never says what result or condition would change that recommendation.

Gemini 3.5 Flash-Lite · Gemini

Does not commit to one North Star; proposes a hybrid (Paid Users Who Cook Weekly), then says MRR is the ultimate business North Star and Weekly Actives who Cook is the operational North Star, leaving no single clear answer.

Respects explicit constraintsRightRightWrong
Sonnet 5.5 · API

The output is a memo to the CEO and Growth Lead, stays under 700 words, and respects the requested form and reader.

Gemini 3.8 Flash · API

The output is a memo to the CEO and Growth Lead, under 700 words, and respects the requested form.

Gemini 3.5 Flash-Lite · Gemini

Recommends MRR as the ultimate business North Star, which violates the explicit prohibition against MRR or MAU as the North Star.

Identifies material uncertaintyRightWrongWrong
Sonnet 5.5 · API

Explicitly names the correlation-vs-causation uncertainty and proposes a holdout experiment to resolve it, with the result potentially changing the decision.

Gemini 3.8 Flash · API

No unknowns that could change the decision are named, nor how they would be resolved.

Gemini 3.5 Flash-Lite · Gemini

Names no unknowns that could change the decision or how they would be resolved.

Avoids unsupported claimsRightWrongWrong
Sonnet 5.5 · API

Forecasts and causal claims are labelled as such (e.g., 'should grow', 'caveat: these are correlations'), and no interpretation is presented as established fact.

Gemini 3.8 Flash · API

It presents 'Cook Mode is the direct gateway' and 'Meal planning provides the forward-looking intent that drives repeat Cook Mode completions' as facts, not hypotheses.

Gemini 3.5 Flash-Lite · Gemini

Presents the '4-49%' retention figure and the 7% weekly-actives claim as fact without support, and does not label them as assumptions.

Produces the required deliverableRightRightWrong
Sonnet 5.5 · API

The memo is complete, within the word limit, and directly usable by the CEO and Growth Lead with at most light edits.

Gemini 3.8 Flash · API

The memo is complete, within the word limit, and usable by the named readers with light edits.

Gemini 3.5 Flash-Lite · Gemini

The memo recommends MRR as a North Star, which is prohibited, and does not provide a single, usable North Star recommendation as requested.

Answers both proposals with the dataRightRightMixed
Sonnet 5.5 · API

Uses the data to reject MAU (68% browse, 4% retention, push test inflated MAU while uninstalls doubled) and MRR (lags, can be moved by price/paywall without product improvement).

Gemini 3.8 Flash · API

It uses the 68% browse/4% retention and the push test data against MAU, and calls MRR a lagging measure of value captured.

Gemini 3.5 Flash-Lite · Gemini

Does not use data to reject MRR (only says it risks ignoring top-of-funnel health, not that it lags and measures value captured); uses data against MAU but not against MRR as required.

Measures value created, not capturedRightRightWrong
Sonnet 5.5 · API

The North Star (Weekly Active Cookers) measures value delivered (cooking), and MRR is kept as a separate business outcome.

Gemini 3.8 Flash · API

Weekly Active Cooks measures value delivered (cooking), and revenue is kept separate.

Gemini 3.5 Flash-Lite · Gemini

The recommended North Star (Paid Users Who Cook Weekly) mixes value captured (payment) with value created, rather than measuring pure value delivered.

Gets the base of every number rightRightRightWrong
Sonnet 5.5 · API

All percentages are quoted directly from the supplied context; no derived figures require a base check.

Gemini 3.8 Flash · API

All percentages and rates are correctly derived from the supplied data, with no base errors.

Gemini 3.5 Flash-Lite · Gemini

The '4-49%' retention claim is not derived from the supplied data, and the 7% figure is misapplied to weekly actives.

Breaks it into inputs teams can moveRightMixedRight
Sonnet 5.5 · API

Provides four input metrics (new-user activation, save-to-cook rate, browse-to-return conversion, planning adoption) each with an owner, and pairs WAC with MRR as the business measure.

Gemini 3.8 Flash · API

It gives input metrics with owners but does not explicitly pair the North Star with a separate business measure of value captured (e.g., MRR as the revenue metric to track alongside).

Gemini 3.5 Flash-Lite · Gemini

Provides three input metrics with owners (Activation Velocity, Core Habit Frequency, Feature Expansion) linked to cooking engagement, and pairs with MRR as a business measure.

All got right 1

Says how it could be gamedRightRightRight
Sonnet 5.5 · API

Identifies gaming vectors (tapping through, notification overuse, short-term spikes, revenue manipulation, MAU drift) and sets specific guardrails (quality completion, push caps, blocking metrics, cohort retention checks, subscriber churn tracking).

Gemini 3.8 Flash · API

It names auto-completing sessions via deep links as a gaming risk and sets a guardrail (session integrity with time and step-progression checks).

Gemini 3.5 Flash-Lite · Gemini

Identifies gaming via push notifications, growth experiments without cooking lift, and low-intent subscriptions, with guardrails (uninstall cap, no-empty-calories rule, 90-day subscriber retention).

Results

Every setup we’ve tested on this task type, across all its tasks and repeats, graded on the current checklist. Calibrated.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1Sonnet 5.5withAPI95.996.72None
2GPT-6.1 SolwithAPI97.793.32None
3GPT-6 LunawithAPI95.588.32None
4GPT-6 AstrawithChatGPT94.288.32None
5Opus 5.5withClaude94.680.82None
6Haiku 5.5withAPI88.881.72None
7Gemini 3.8 FlashwithAPI78.965.82None
8Gemini 3.5 Flash-LitewithGemini54.438.321 capped

What you get for the cost

GPT-6 Luna · API scores within 4 points of Sonnet 5.5 · API at about a twentieth of the cost.

Task score against relative cost for each setup. The table below has the same figures.304050607080901001×3×10×30×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreSonnet 5.5 · API: task score 96, 22× the cheapestGPT-6.1 Sol · API: task score 96, 9.6× the cheapestGPT-6 Luna · API: task score 92, the cheapestGPT-6 Astra · ChatGPT: task score 91, 52× the cheapestOpus 5.5 · Claude: task score 86, 36× the cheapestHaiku 5.5 · API: task score 86, the cheapestGemini 3.8 Flash · API: task score 71, 7.5× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 44, 1.8× the cheapest, 1 critical failureSonnet 5.5 · APIGPT-6.1 Sol · APIGPT-6 Luna · APIGPT-6 Astra · ChatGPTOpus 5.5 · ClaudeHaiku 5.5 · APIGemini 3.8 Flash · APIGemini 3.5 Flash-Lite · Gemini
Best valueNobody beats these on both score and costAt least one critical failure
Task score and relative cost by setup
SetupTask scoreRelative cost (cheapest = 1)
GPT-6 Luna · API921.0×
Haiku 5.5 · API861.1×
Gemini 3.5 Flash-Lite · Gemini441.8×
Gemini 3.8 Flash · API717.5×
GPT-6.1 Sol · API969.6×
Sonnet 5.5 · API9622×
Opus 5.5 · Claude8636×
GPT-6 Astra · ChatGPT9152×

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

About the task

The PM job

Choosing the metric the whole company steers by.

Why it matters

Pick revenue and you steer by the rear-view mirror. Pick a vanity metric and teams will happily grow it while the product gets worse.

What good looks like

  • Measures value delivered to customers, not value captured
  • Grounded in what predicts retention in the data
  • Broken into input metrics a team can own
  • Says how it could be gamed, with guardrails

Deliberately not measured

  • Dashboard design
  • Naming
Capability tested

Metric design

The failure we’re looking for

Revenue, or a vanity metric, as the North Star

Grading

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