Humanity’s Last PRD
Latest drop: Haiku 5.5 · 8 Oct
Tasks / Metrics & Experimentation

Experiment specification

Can the model design a test that could actually change the decision?

Task type v1.0 · 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. 31% were usable with at most a quick edit.

Reliably right

  1. Randomises by user100% pass
    Randomises by logged-in user ID across platforms, explicitly citing cross-platform usage and session non-independence.
    Sonnet 5.5 · API · Show the fee sooner
  2. Works around the calendar94% pass
    Addresses both the payments test and the Black Friday freeze, proposes concurrent independent assignment with an interaction check, and avoids a one-week squeeze.
    Sonnet 5.5 · API · Show the fee sooner
  3. Addresses the actual decision91% pass
    Commits to a clear experiment plan with a decision rule that maps every outcome (ship, don't ship, kill, extend) to an action.
    Sonnet 5.5 · API · Show the fee sooner

Where it slips

  1. Guardrails with thresholds50% pass
    Guardrail metrics are named but lack explicit numeric thresholds; 'materially worse' and 'drops meaningfully' are too vague to block rollout without further definition.
    Sonnet 5.5 · API · Showing the delivery fee up front
  2. Says why a test is worth it here59% pass
    The output does not explain why this decision merits a test given the traffic and stakes, nor does it recommend a cheaper alternative.
    Opus 5.5 · Claude · Show the fee sooner
  3. Uses the supplied evidence correctly64% pass
    Unsupported claims about the traffic data's adequacy and the code freeze's preapproval requirement are presented as facts, not in the brief.
    GPT-6.1 Sol · API · Show the fee sooner

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. An unambiguous primary metric

    Does the spec name one primary metric, explain why an increase is unambiguously good for the user and the business, and say how the result's trustworthiness will be checked?

    Passes when One primary metric with a clear rationale for its direction, plus a planned trust check such as a sample ratio check.

    Ronny KohaviLenny’s Podcast

  2. Decision rule written before the test

    Does the spec say, before the test runs, which result leads to which action (ship, don't ship, iterate), including what happens if the result is inconclusive?

    Passes when Every outcome, including an inconclusive one, maps to a stated action, with the thresholds that separate them.

    Ronny KohaviLenny’s Podcast

  3. Sized from the real traffic

    Is the sample size and duration worked out from the supplied traffic, baseline and smallest effect worth detecting, run in full weekly cycles?

    Passes when Sample size follows from the baseline and the smallest effect worth acting on, the duration follows from the traffic the brief gives (including how units accumulate), and it runs whole weeks.

    Product and growth leadersOn Lenny’s Podcast

  4. Guardrails with thresholds

    Does the spec name the guardrail metrics that would stop a rollout even if the primary metric wins, each with a threshold?

    Passes when The guardrails that matter for this change are named, each with a threshold that would block the rollout.

    Lauryn IsfordLenny’s Podcast

  5. Says why a test is worth it here

    Does the spec weigh whether a test is worth running here: enough traffic to detect an effect that matters, and a decision the result would actually change, and does it say what it would do instead if not?

    Passes when Explains why this decision merits a test given the traffic and stakes, or recommends a cheaper route where they don't.

    Lauryn IsfordLenny’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 · 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 checkout at Basketful. We want to test showing the delivery fee in the basket, instead of only at the last step of checkout. Write the experiment spec for Ravi, the growth engineer who'll build it, Ines, the analyst who'll read it out, and Chloe, our Head of Growth, who approves it. Keep it under 700 words. The team's draft plan is below. Fix what needs fixing.

What the model was given5 items: About Basketful, Why we're testing it, Traffic and baselines, Smallest change worth acting on, The team's draft plan
About BasketfulAn online grocery service. Delivery costs £3.99, or is free on orders over £60. Today the fee first appears on the final checkout step.
Why we're testing it12% of support contacts last quarter were about delivery fees the customer hadn't expected. Design believes showing the fee early will build trust; Chloe worries it will put people off before they start checkout.
Traffic and baselinesAbout 40,000 different customers view their basket each week. About half of each week's basket viewers are new that week; the rest came back from earlier weeks. 8% of basket viewers place an order. Of customers who start checkout, 62% complete it. Average order value: £47.
Smallest change worth acting onChloe: “Anything smaller than half a percentage point on orders isn't worth arguing about.”
The team's draft planPrimary metric: checkout completion rate (orders ÷ checkout starts). Split: 50/50 by session. Duration: we'll check the dashboard every day and stop as soon as it's significant. Chloe's addition: “While we're at it, let's put the free-delivery-over-£60 banner in the variant too.”
What a strong answer doesThe answer key the graders mark against

A spec that tests one change and fixes the draft's three problems. The primary metric is orders per basket-viewing customer, not checkout completion rate: showing the fee earlier changes who starts checkout, so completion can rise while orders fall. Customers, not sessions, are randomised, so returning customers see one version. The free-delivery banner stays out of the variant (or gets its own test), or the result can't be attributed. The duration is fixed in advance, with no daily stopping: detecting 0.5pp on an 8% base needs roughly 47,000 customers per arm (about 95,000 in total), and because half of each week's viewers are returning customers, new customers accumulate at about 40,000, 60,000, 80,000 and 100,000 after four weeks, so it runs about four full weeks, not two and a half. Guardrails with thresholds: average order value, fee-related support contacts, and the checkout-start rate as a diagnostic. A sample ratio check before reading results. A decision rule written up front: ship if orders are no worse than 0.5pp down and fee contacts fall; don't ship if orders drop by 0.5pp or more; with a stated plan for an inconclusive result.

Critical failures (cap the score at 40)
  • Uses checkout completion rate (orders ÷ checkout starts) as the primary metric
  • Bundles the free-delivery banner into the variant
Task

Mid-level · v1.3

The outputs

8 setups answered this brief.

Sonnet 5.5API93Usable with a quick editHaiku 5.5API93Usable with a quick editGPT-6 AstraChatGPT80Needs more workGPT-6.1 SolAPI79Usable with a quick editGPT-6 LunaAPI77Usable with a quick editOpus 5.5Claude75Needs more workGemini 3.8 FlashAPI48Needs more workGemini 3.5 Flash-LiteGemini18Critical failure
Haiku 5.5API

Usable with a quick edit12 of 13 checks passed

The spec is complete, directly addresses all team members, and a PM could use it as-is, only needing to fill in the two baselines that are clearly flagged.

Output

Experiment Spec: Delivery Fee in Basket

To: Ravi (build), Ines (analysis), Chloe (approval) From: PM, Checkout Status: For approval

What changed from the draft

  • Primary metric moved from checkout completion to basket-to-order conversion. Completion only measures people who reach checkout, so it would miss the risk Chloe raised: people leaving before checkout.
  • Randomise by customer, not session. Half of each week's viewers are returning customers. Session splits would show the same person both versions.
  • Fixed duration, pre-set sample size. Checking daily and stopping at first significance inflates false positives.
  • Free-delivery banner removed. It changes the same thing we're measuring, so we couldn't attribute any effect. We'll propose it as a separate test.

Hypothesis

Showing the delivery fee in the basket changes basket-to-order conversion. We expect an increase of at least 0.5 percentage points (8.0% to 8.5%). The test is two-sided, so a decrease would also be recorded and acted on.

Variants

  • Control: Basket as today. Fee appears only on the final checkout step.
  • Variant: Basket shows a delivery line under the subtotal: "Delivery: £3.99" or "Delivery: free" once the basket is over £60. Checkout is unchanged.

Ravi: The line must update live as the basket crosses the £60 threshold, and must match the fee charged at checkout exactly.

Population and randomisation

  • All customers who view their basket during the test period.
  • Randomise 50/50 by customer ID at first basket view in the test. Assignment stays fixed for the whole test.
  • Run a sample ratio mismatch (SRM) check on day 1. If the split deviates from 50/50 at p < 0.01, pause and investigate before continuing.

Metrics

Primary: Basket-to-order conversion. Customers who place an order within 7 days of their first basket view in the test, divided by customers who viewed their basket. Baseline: 8%.

Guardrails: - Checkout start rate (checkout starts ÷ basket viewers). Ines to pull baseline. - Checkout completion rate. Baseline: 62%. - Average order value. Baseline: £47. - Delivery-fee support contacts per 1,000 orders. Ines to pull baseline; currently 12% of all contacts.

Sample size and duration

  • Baseline 8.0%, minimum effect 0.5 percentage points absolute, two-sided α = 0.05, power 80%.
  • Requires about 47,500 customers per arm (95,000 total).
  • Traffic is about 40,000 basket viewers per week, of whom about 20,000 are new. Unique customers therefore accumulate at roughly 40k in week 1 and 20k per week after that.
  • Plan: 4 weeks of enrolment, then a 7-day conversion window, for a fixed end date of [date]. Ines to confirm the unique-customer count, since the 20k-new-per-week figure is an estimate.
  • No decisions before the end date, with one exception: at week 2, Ines runs a harm-only check on guardrails. We stop early only if a guardrail shows harm at p < 0.001.

Decision rules

Ship if: - The primary metric is significantly positive (p < 0.05), with an estimated uplift of at least 0.5 percentage points, and - No guardrail is significantly worse.

Don't ship if: - The primary result is not significant, or - It is significant but below 0.5 points (not worth acting on, per Chloe), or - A guardrail is significantly worse.

We don't iterate on a null result by re-slicing the data until something looks significant.

Analysis (Ines)

  • Analyse at the customer level, intention-to-treat: all assigned customers, regardless of whether they saw the fee line.
  • Report absolute difference in percentage points and relative difference, each with a 95% confidence interval.
  • Report all guardrails.
  • Pre-registered segment: new vs returning customers. Returning customers may already know the fee, so the effect may differ. Treat any other segment cuts as exploratory.

Timeline and sign-off

  • Build complete: [date]
  • Launch: [date]
  • Fixed end: [date]
  • Readout: [date + 3 working days]

Approved by: _____ Chloe, Head of Growth Date: ___

Grades and run details

Decision model 85 · LLM judge 100
Decision model checks
  • failUses the supplied evidence correctly33%
  • passAddresses the actual decision79%
  • passRespects explicit constraints14%
  • passIdentifies material uncertainty45%
  • passAvoids unsupported claims28%
  • passProduces the required deliverable34%
  • passTests one change at a time95%
  • passFixes the duration up front97%
  • passAn unambiguous primary metric77%
  • passDecision rule written before the test41%
  • partialSized from the real traffic40%
  • partialGuardrails with thresholds91%
  • passSays why a test is worth it here62%
Run
Run
#1
API response time
29 s
Submitted
8 Oct 2026
GPT-6.1 SolAPI

Usable with a quick edit9 of 13 checks passed

The spec is well-structured and fixes the draft's main problems; a PM could add guardrail thresholds and a brief justification with a quick edit.

Output

Experiment spec: delivery-fee visibility in basket

Owner: Checkout PM Build: Ravi · Analysis: Ines · Approval: Chloe

Objective and hypothesis

Test whether showing delivery costs before checkout increases orders by reducing surprises—or decreases orders by discouraging customers earlier.

Last quarter, 12% of support contacts concerned unexpected delivery fees. This motivates the test but is not the percentage of customers affected.

Experience

Control: Current basket; delivery fee first appears at the final checkout step.

Variant: Show a clearly labelled delivery-fee line and fee-inclusive total in the basket. Update both when the basket changes. Use the production pricing logic: £3.99 delivery, free for orders over £60. Ravi must verify boundary behaviour at exactly £60 and match existing checkout pricing.

Keep the final checkout fee display unchanged. Everything else stays identical.

Do not add the free-delivery banner. It could independently encourage larger baskets, making it impossible to attribute results to earlier fee visibility. Test it separately.

Eligibility and assignment

Include customers viewing an orderable basket in supported markets. Apply existing employee, test-account and bot exclusions identically, before assignment.

Randomise 50/50 by customer, not session, at their first eligible basket view. Persist assignment across visits and devices where identity is available. Use a stable anonymous identifier for guests and a documented account-linking rule; Ravi and Ines must validate identity handling before launch.

Each customer enters the experiment once. Analyse everyone assigned, regardless of whether they start checkout or successfully see the treatment.

Metrics

Primary: Seven-day order conversion: customers placing at least one order within seven days of assignment ÷ all assigned eligible basket viewers.

This captures Chloe’s concern about discouraging checkout entry. Orders ÷ checkout starts is unsuitable as primary: the treatment can change who starts checkout.

Secondary/diagnostic: - Checkout-start rate per assigned customer. - Checkout completion among starters, interpreted cautiously because starter populations may differ. - Revenue per assigned customer, including non-buyers. - AOV and proportion of orders qualifying for free delivery. - Unexpected-fee support contacts per assigned customer within seven days, using consistent contact classification.

Operational guardrails: Basket/checkout errors, incorrect fee calculations and page performance. These can trigger a safety rollback, not an early declaration of success.

Sample size and duration

Interpret Chloe’s minimum worthwhile change as 0.5 percentage points absolute: 8.0% to 8.5%, not a 0.5% relative lift.

With a matching 8% baseline, two-sided 5% significance and 80% power, plan for approximately 95,000 unique customers total, evenly split. Ines must verify the historical baseline uses the same seven-day definition and finalise the calculation before launch.

Do not count 40,000 weekly viewers as entirely new experiment entrants. Estimated cumulative reach is 40,000 in week one, then roughly 20,000 additional customers weekly: about four weeks to enrol the target, plus seven days for outcomes to mature.

Lock the enrolment target before launch; stop enrolment when reached, then wait seven days. Extend only if unique-customer accrual is slower—not based on results.

Analysis and decision

Ines reports the intention-to-treat conversion difference, 95% confidence interval and absolute rates. Check allocation imbalance, logging, identity consistency and treatment exposure before interpreting results.

No daily significance-based stopping. Daily dashboards are for operational safety only.

Recommend rollout if the primary improvement is statistically significant, the point estimate is at least +0.5 percentage points, and there are no unresolved operational or economic concerns.

Otherwise, retain control pending review. Distinguish evidence of harm, evidence ruling out a worthwhile gain, and an inconclusive result; “not significant” does not mean “no effect.”

Launch approval

Ravi signs off implementation and QA; Ines signs off measurement and the locked analysis plan; Chloe approves the isolated treatment, decision rule and any economic guardrail thresholds before launch.

Grades and run details

Decision model 77 · LLM judge 79
Decision model checks
  • failUses the supplied evidence correctly6%
  • passAddresses the actual decision66%
  • partialRespects explicit constraints19%
  • passIdentifies material uncertainty38%
  • passAvoids unsupported claims68%
  • passProduces the required deliverable57%
  • passTests one change at a time100%
  • passFixes the duration up front74%
  • passAn unambiguous primary metric80%
  • partialDecision rule written before the test42%
  • partialSized from the real traffic41%
  • partialGuardrails with thresholds98%
  • passSays why a test is worth it here88%
Run
Run
#1
API response time
36 s
Submitted
30 Sept 2026
GPT-6 LunaAPI

Usable with a quick edit7 of 13 checks passed

The core spec is solid and fixes the draft's main problems; a PM could add the missing explanation for no peeking, a sample ratio check, guardrail thresholds, and an approval recommendation with quick edits.

Output

Basket delivery-fee experiment

For: Ravi (build), Ines (analysis), Chloe (approval) Decision requested: Approve a test of showing the existing delivery fee in the basket.

Hypothesis and scope

Showing the delivery cost before checkout will reduce surprise and fee-related support contacts, without materially reducing orders.

  • Control: Current experience; delivery fee first appears on the final checkout step.
  • Variant: Show the applicable fee in the basket: “Delivery £3.99” below the free-delivery threshold, or “Free delivery” when the basket qualifies. Keep the existing “free on orders over £60” rule and wording.
  • Change nothing else. Do not add the free-delivery banner: that would test a second intervention and make the result hard to interpret.

Build and assignment — Ravi

Randomise 50/50 by customer, not session, and keep each customer in the same arm across visits and devices where identity is available. Stratify assignment by new versus returning basket viewers. Exclude staff and test accounts. Log assignment, basket value/fee shown, basket views, checkout starts, orders, order value, and relevant support contacts.

Measurement — Ines

Primary metric: Basket-viewer conversion: percentage of assigned customers who place at least one order within 7 days of their first eligible basket view. Count each customer once, in their assigned arm (intention-to-treat). This captures any deterrence before checkout; checkout completion alone would miss it.

Guardrails / secondary measures: Checkout starts per basket viewer; checkout completion (orders ÷ checkout starts); average order value and revenue per assigned customer; delivery-fee-related support contacts per assigned customer. Support contacts may be too sparse for a definitive test, so treat them as directional, not a substitute for the primary outcome.

Sample, duration, and analysis

Baseline basket-viewer conversion is 8%. Power the test to detect a 0.5 percentage-point absolute difference (Chloe’s smallest worthwhile order effect), with 80% power and a two-sided 5% significance level. This requires approximately 95,000 unique customers total (about 47,500 per arm), assuming independent customer-level outcomes.

There are about 40,000 basket viewers weekly, but many return across weeks. Run for at least four full weeks and until the sample target is reached; then allow 7 days for outcomes to mature before reading results. If the target is not reached, extend in full weeks. Ines should confirm the final sample-size calculation and report actual enrolment.

Do not stop early for significance or check daily results to make a stopping decision. Monitor only for implementation or customer-harm issues. Analyse the primary metric at the customer level, with a 95% confidence interval and the pre-specified two-sided test.

Decision

Recommend rollout if the variant improves basket-viewer conversion by at least 0.5 percentage points, the result is statistically significant, and guardrails show no material harm. Do not roll out if there is significant harm. If results are inconclusive, report that clearly; do not interpret “not significant” as proof of no effect.

Grades and run details

Decision model 96 · LLM judge 57
Decision model checks
  • passUses the supplied evidence correctly9%
  • passAddresses the actual decision60%
  • passRespects explicit constraints30%
  • passIdentifies material uncertainty36%
  • passAvoids unsupported claims62%
  • passProduces the required deliverable60%
  • passTests one change at a time94%
  • passFixes the duration up front60%
  • passAn unambiguous primary metric44%
  • passDecision rule written before the test29%
  • passSized from the real traffic24%
  • partialGuardrails with thresholds98%
  • passSays why a test is worth it here72%
Run
Run
#1
API response time
22 s
Submitted
30 Sept 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 correctlyMixedMixedRight
Haiku 5.5 · API

All statements about the current situation are directly from the supplied context or simple arithmetic; no facts are invented.

GPT-6.1 Sol · API

All factual claims about the current situation are taken directly from the brief or derived correctly from it.

GPT-6 Luna · API

All statements about the current situation are taken directly from the brief or supplied context.

Addresses the actual decisionRightRightMixed
Haiku 5.5 · API

The output commits to a clear experimental design and includes decision rules that map results to specific actions (ship/don't ship), which addresses the team's need to decide how to test and what to act on.

GPT-6.1 Sol · API

The spec commits to a clear experiment plan with a decision rule for rollout, addressing the request for an experiment spec.

GPT-6 Luna · API

The output does not commit to a clear answer on whether to approve the test; it only provides a rollout decision rule for after the test.

Fixes the duration up frontRightMixedMixed
Haiku 5.5 · API

The spec sets a fixed 4-week enrolment plus 7-day conversion window, explains that daily peeking inflates false positives, and only allows a harm-based early stop on guardrails.

GPT-6.1 Sol · API

It sets a fixed duration but does not explain why daily peeking inflates false positives; it only states not to do it.

GPT-6 Luna · API

It sets a fixed duration but does not explain why stopping at the first significant day inflates false positives.

An unambiguous primary metricRightRightMixed
Haiku 5.5 · API

A single primary metric (basket-to-order conversion) is named with a clear rationale for its direction, and a sample ratio mismatch check is planned.

GPT-6.1 Sol · API

It names one primary metric (seven-day order conversion) with a clear rationale and includes a trust check (allocation imbalance, logging, etc.).

GPT-6 Luna · API

No sample ratio check or other trust check is planned, only a confirmation of sample size and enrolment.

Decision rule written before the testRightRightMixed
Haiku 5.5 · API

Every possible result is mapped to an action: ship if significant with ≥0.5pp uplift and guardrails safe; don't ship otherwise; no iteration on null results.

GPT-6.1 Sol · API

It states rollout conditions (significant, ≥0.5pp, no concerns) and what to do otherwise, including handling inconclusive results.

GPT-6 Luna · API

The inconclusive outcome maps only to 'report clearly', not to a decision on whether to ship, iterate, or not ship.

Guardrails with thresholdsRightWrongWrong
Haiku 5.5 · API

Checkout start rate, completion rate, average order value, and delivery-fee support contacts are listed as guardrails, with a requirement that they are not significantly worse for rollout to proceed.

GPT-6.1 Sol · API

It lists guardrail metrics but does not set specific thresholds (e.g., for AOV or support contacts) that would block rollout.

GPT-6 Luna · API

Guardrail metrics are listed but no thresholds are given that would block a rollout.

Says why a test is worth it hereRightMixedMixed
Haiku 5.5 · API

The spec demonstrates the test is powered for the smallest change that matters, and the decision rules show the result would change the team's action, which justifies running the test.

GPT-6.1 Sol · API

It does not explain why this test is worth running given the traffic and stakes, nor does it recommend a cheaper alternative.

GPT-6 Luna · API

The output does not explain why a test is worth running here given the traffic and stakes.

All got right 6

Respects explicit constraintsRightRightRight
Haiku 5.5 · API

The spec is within the 700-word limit, is addressed to Ravi, Ines, and Chloe, and proposes an actionable test plan.

GPT-6.1 Sol · API

The output is under 700 words, addresses the named readers, and fixes the draft's problems as required.

GPT-6 Luna · API

The spec is under 700 words, addresses Ravi, Ines, and Chloe, and fixes the draft's problems.

Identifies material uncertaintyRightRightRight
Haiku 5.5 · API

The output names uncertainty in the new-customer accumulation estimate and asks Ines to confirm, and notes the segment effect may differ; it bounds where possible.

GPT-6.1 Sol · API

It identifies unknowns (baseline verification, identity handling, enrollment speed) and says how to resolve them, and distinguishes inconclusive results.

GPT-6 Luna · API

It acknowledges that support contacts may be too sparse and that inconclusive results are possible, and says how to handle them.

Avoids unsupported claimsRightRightRight
Haiku 5.5 · API

Hypotheses, forecasts, and reasoning are clearly labelled as such; no confident claims go beyond the supplied evidence.

GPT-6.1 Sol · API

Hypotheses and interpretations are clearly labeled, and no confident claims go beyond the supplied evidence.

GPT-6 Luna · API

No interpretations or forecasts are presented as established facts; the sparsity note is hedged as 'may'.

Produces the required deliverableRightRightRight
Haiku 5.5 · API

The output is a complete experiment spec with sections for build, analysis, and sign-off, and is usable by the intended readers with minimal editing.

GPT-6.1 Sol · API

The output is a complete experiment spec in the requested form, under 700 words, and usable by the named readers.

GPT-6 Luna · API

The experiment spec is complete, under 700 words, and usable by the named readers.

Tests one change at a timeRightRightRight
Haiku 5.5 · API

The free-delivery banner is explicitly removed from the variant with a clear explanation that it would confound attribution, and it is suggested as a separate test.

GPT-6.1 Sol · API

It explicitly excludes the free-delivery banner and explains that bundling would make attribution impossible.

GPT-6 Luna · API

The spec explicitly keeps the free-delivery banner out of the variant and explains why bundling it would make the result uninterpretable.

Sized from the real trafficRightRightRight
Haiku 5.5 · API

Sample size is correctly derived from the 8% baseline, 0.5pp effect, 80% power; duration follows from the 40k weekly traffic and half-new accumulation, running whole weeks.

GPT-6.1 Sol · API

Sample size is calculated from the 8% baseline and 0.5pp effect, and duration follows from the traffic accumulation pattern, running whole weeks.

GPT-6 Luna · API

Sample size is correctly derived from the 8% baseline and 0.5pp effect, and duration accounts for weekly traffic and returning customers.

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 85% of 39 checks.

#Model · HarnessTask scoreDecision modelLLM judgeRunsCritical failures
1Sonnet 5.5API92.696.42None
2Haiku 5.5API83.496.72None
3Opus 5.5Claude88.779.02None
4GPT-6 AstraChatGPT88.976.02None
5GPT-6 LunaAPI89.165.22None
6GPT-6.1 SolAPI83.169.32None
7Gemini 3.8 FlashAPI77.347.92None
8Gemini 3.5 Flash-LiteGemini47.813.321 capped

What you get for the cost

No setup scores within 5 points of the leader at half its cost or less here, so the leader is also the best value.

Task score against relative cost for each setup. The table below has the same figures.20304050607080901001×3×10×30×100×Cost per output, relative to the cheapest setup (log scale) →↑ Task scoreSonnet 5.5 · API: task score 95, 33× the cheapestHaiku 5.5 · API: task score 89, 2.7× the cheapestOpus 5.5 · Claude: task score 81, 49× the cheapestGPT-6 Astra · ChatGPT: task score 79, 79× the cheapestGPT-6 Luna · API: task score 76, the cheapestGPT-6.1 Sol · API: task score 75, 13× the cheapestGemini 3.8 Flash · API: task score 60, 7.4× the cheapestGemini 3.5 Flash-Lite · Gemini: task score 29, 2.1× the cheapest, 1 critical failureSonnet 5.5 · APIHaiku 5.5 · APIOpus 5.5 · ClaudeGPT-6 Astra · ChatGPTGPT-6 Luna · APIGPT-6.1 Sol · APIGemini 3.8 Flash · APIGemini 3.5 Flash-Lite · Gemini
Nobody beats these on both score and costAt least one critical failure

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

Specifying an A/B test before running it.

Why it matters

Most tests are underpowered, or measure a metric nobody agrees is good.

What good looks like

  • One primary metric everyone agrees on the direction of
  • Decision rule stated up front
  • Guardrails
  • Power considered

Deliberately not measured

    Capability tested

    Test design

    The failure we’re looking for

    A test with no decision attached

    Grading

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