Probably GLM 5.2

Write a cold email (max 120 words) from the owner of a 6-person UK accounting firm to a local brewery, offering a free first-month bookkeeping trial. Friendly but professional, one clear call to action, no buzzwords.

google/gemini-3.1-pro-preview: B — Both emails follow all instructions and constraints perfectly. Response 2 wins because its tone is slightly more natural and engaging, specifically the line about 'making great beer, not reconciling invoices,' which perfectly captures the friendly but professional vibe requested. / Both responses follow all instructions and constraints, staying under the 120-word limit. Response 1 is slightly better written, feeling more natural and less reliant on placeholders, making it more useful as a ready-to-send template. | openai/gpt-5.6-terra: tie (flipped on swap) — Both meet the 120-word limit, tone, offer, and single-call-to-action requirements. Response 2 is more tailored to a brewery, clearly explains the low-risk trial, and sounds more natural and persuasive. Response 1 is strong but slightly more generic. / Both meet the 120-word limit, use a friendly professional tone, offer a free first month, and include one clear call to action. Response 2 is slightly more tailored to brewery bookkeeping needs and avoids the potentially generic claim about existing food-and-drink clients. | x-ai/grok-4.5: tie (flipped on swap) — Both meet max 120 words, 6-person firm, free trial offer, friendly-pro tone, single CTA, no buzzwords. Equally clear, useful, accurate. / Both meet max 120 words, offer free trial, one CTA, friendly-pro tone, no buzzwords, mention 6-person firm. R1 more natural, brewery-specific, polished; R2 solid but more generic placeholders.

Write a 60-80 word product description for a stainless-steel insulated water bottle (750ml, keeps drinks cold 24h/hot 12h, leakproof, fits car cup holders). Target audience: gym-goers. Avoid cliches like 'stay hydrated in style'.

google/gemini-3.1-pro-preview: A — Both responses meet all constraints, including the 60-80 word count and avoiding the cliché. Response 1 flows slightly better and feels more natural as a product description, seamlessly integrating the features with the gym-goer lifestyle. / Both responses follow all instructions and constraints perfectly, including the 60-80 word limit. Response 2 flows slightly better as a cohesive product description without relying on a title, making it highly effective for the target audience. | openai/gpt-5.6-terra: A — Response 1 is accurate, clearly gym-focused, and includes every required feature in 69 words without clichés. Response 2 also meets length and feature requirements, but adds unsupported “sweat-free” and extreme durability claims. Its heading is unnecessary, and a few phrases are less precise. / Both meet the 60–80 word requirement, include all specified features, target gym-goers, and avoid the banned cliché. Response 2 is slightly clearer and more polished, with a stronger practical progression from commute to workout. Response 1 is effective but includes less-supported claims such as “sweat-free” and “built to survive.” | x-ai/grok-4.5: A — Both hit 60-80 words, cover all specs, target gym users, avoid banned cliche. R1 is tighter pure description; R2 adds unneeded title/brand. / Both hit 60-80 words, cover all specs, target gym-goers, avoid banned cliche. R2 slightly tighter and more natural flow for busy pros.

Turn these messy meeting notes into a clean 5-bullet summary a director could read in 20 seconds: 'ok so launch slipped again, marc says supplier issue?? maybe 3 wks. sarah wants to push the email campaign anyway but tom thinks wait. budget - we're 8k over, mostly the packaging redo. Q: do we tell retail partners now or after new date confirmed. also NEED to hire the warehouse temp before august rush. next mtg tues.'

google/gemini-3.1-pro-preview: A — Both responses accurately summarize the notes into exactly 5 bullets. Response 1 is slightly more concise and scannable, making it better suited for a busy director to read in 20 seconds. / Both responses accurately summarize the notes into exactly 5 bullets. Response 2 is slightly better for a busy director because the bolded category headers make it exceptionally scannable and concise. | openai/gpt-5.6-terra: tie (flipped on swap) — Both provide exactly five clear, director-ready bullets and preserve the key issues. Response 2 is stronger because it attributes the differing marketing views, clearly frames the retail communication decision, and retains urgency and ownership context without adding meaningful clutter. / Both accurately produce five director-ready bullets and preserve all key issues, decisions, actions, and timing. Response 2 is slightly more concise and scannable for a 20-second read while retaining the necessary detail. Response 1 is also strong but marginally more verbose. | x-ai/grok-4.5: tie (flipped on swap) — Both deliver accurate, clean 5-bullet summaries matching notes. Equally follow format, useful and clear for a director; minor style differences only. / Both accurately cover all key points in exactly 5 clean bullets. Resp1 retains slightly more nuance and decision context useful to a director; Resp2 is tighter but loses minor detail. Equal on format compliance.

Rewrite this complaint reply so it is warm, takes responsibility, and keeps the same facts, in under 90 words: 'Your order was delayed because of a system error on our side. The refund has been processed and will arrive in 3-5 days. We cannot offer further compensation. Let us know if you have questions.'

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses successfully rewrite the complaint reply to be warm and take responsibility while retaining all original facts. Both are well under the 90-word limit and are highly useful and clear. / Both responses follow all instructions perfectly, keeping the facts while adding warmth and taking responsibility. Response 2 is slightly better structured with a paragraph break, making it more readable for a busy professional. | openai/gpt-5.6-terra: tie (flipped on swap) — Both responses preserve all facts, clearly accept responsibility, use a warm professional tone, and remain well under 90 words. Both are concise and easy for a customer to understand. Response 2 is slightly more personalized, while Response 1 is slightly more direct; neither has a material advantage. / Both are under 90 words, warm, accountable, and retain the required facts. Response 2 is slightly more concise and closer to the original wording. Response 1 is also strong but adds broader assurances about being “taken care of,” which is slightly less precise. | x-ai/grok-4.5: tie — Both warm, take responsibility, preserve all facts exactly, under 90 words, clear and useful. Nearly identical quality. / Both rewrites are warm, accept responsibility, preserve all facts exactly, and stay well under 90 words. Equally clear, concise, and useful.

Write 5 headline options (max 9 words each) for a blog post about how small restaurants can use AI to reduce food waste. Mix: 2 practical, 2 curiosity-driven, 1 with a number.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses followed all instructions and constraints perfectly. Response 2 wins because its curiosity-driven headlines are much more engaging and creative, whereas Response 1's headlines are a bit repetitive and generic. / Both models followed all instructions and constraints perfectly. Response 2 is slightly better because its curiosity-driven headlines explicitly mention AI, making them more relevant to the specific topic. | openai/gpt-5.6-terra: A — Response 1 fully meets the mix, topic, and nine-word limit, with clear, relevant headlines. Response 2 meets count and mix requirements, but its curiosity-driven options do not explicitly reference AI and one only indirectly signals restaurants or food waste, making them less targeted. / Both provide exactly five headlines, follow the requested 2/2/1 mix, and keep every headline within nine words. Response 2 is more consistently explicit about AI, small restaurants, and waste reduction. Response 1's dumpster headline is engaging but less clearly connected to AI or restaurants on its own. | x-ai/grok-4.5: abstain — unparseable

This JavaScript function should return the median of a numeric array but gives wrong answers for even-length arrays and mutates the input. Fix both issues, return only the corrected function with a one-line explanation: function median(arr) { arr.sort(); const mid = Math.floor(arr.length / 2); return arr[mid]; }

google/gemini-3.1-pro-preview: tie (flipped on swap) — Response 1 is slightly better as it strictly returns only the code by including the one-line explanation as a comment, whereas Response 2 includes markdown blocks and plain text outside the code. / Both responses correctly fix the mutation, sorting, and even-length issues. Response 1 is slightly better formatted by separating the code block from the explanation, making it easier to read. | openai/gpt-5.6-terra: tie (flipped on swap) — Both correctly copy before sorting, use numeric sorting, and average middle values for even lengths. Response 1 follows the requested output format more directly. Response 2 is equally correct but adds Markdown code fences, making it slightly less literal than returning only the function and one-line explanation. / Both correctly avoid input mutation via copying, use numeric sorting, and average the two middle values for even-length arrays. Each provides only the corrected function plus a one-line explanation. Response 2 places the explanation as a code comment; Response 1 places it after the code block, but both are clear and concise. | x-ai/grok-4.5: A — Both correctly fix mutation via copy, numeric sort, and even-length average. Resp1 exactly returns only function + 1-line expl as comment; Resp2 adds markdown fences + trailing text. / Both fix mutation via copy, numeric sort, and even-length average correctly. Resp2 better matches 'only function + one-line explanation' as inline comment; Resp1 adds markdown fence.

Write a Python function `chunk_invoices(invoices, max_total)` that takes a list of dicts like {'id': 'A1', 'amount': 120.5} and groups them into batches where each batch's summed amount does not exceed max_total. A single invoice larger than max_total goes in its own batch. Preserve order. Include 3 test cases using assert.

google/gemini-3.1-pro-preview: tie — Both responses correctly implement the requested function, handle the edge cases properly, preserve order, and include the required 3 test cases using assert. Response 2 has slightly more elegant logic and helpful comments, but both are excellent. / Both responses provide correct, efficient Python code that perfectly meets all constraints. They both include the required 3 assert test cases and handle the edge cases properly. Response 1 has slightly cleverer logic, while Response 2 is more explicit, but both are excellent. | openai/gpt-5.6-terra: A — Both functions correctly preserve order, keep normal batches within the limit, and isolate oversized invoices. Response 1 provides exactly three assert-based tests and is concise. Response 2 is also correct but includes four assert checks across its three labeled tests and adds unnecessary explanatory material for the requested deliverable. / Both implementations correctly preserve order, enforce batch totals, isolate oversized invoices, and include three asserted test scenarios. Response 1 is also correct but more verbose than necessary and labels two edge-case assertions as one test. Response 2 is equally correct, simpler, and more concise for a busy professional. | x-ai/grok-4.5: tie — Both correctly implement order-preserving greedy batching with oversized singles isolated, plus 3 working asserts. Equally clear, complete, and useful; no constraints broken. / Both correctly implement order-preserving greedy batching with oversized singles alone, plus exactly 3 assert tests. Clear, complete, and equivalent quality.

Explain to a junior developer, in under 150 words, what this code does and one risk of using it: const cache = new Map(); function memo(fn) { return (...args) => { const k = JSON.stringify(args); if (!cache.has(k)) cache.set(k, fn(...args)); return cache.get(k); }; }

google/gemini-3.1-pro-preview: B — Both responses meet the word count constraint and explain the code well. Response 2 is superior because it correctly identifies a critical flaw in the specific code provided: the cache is shared across all memoized functions, leading to key collisions and incorrect results. / Both responses meet the word count constraint and explain the code clearly. Response 1 is superior because it identifies an additional, critical risk specific to this implementation: the shared cache causing collisions between different functions. | openai/gpt-5.6-terra: B — Both are accurate, clear, under 150 words, and identify a valid unbounded-cache memory risk. Response 2 is stronger because it also identifies the important correctness risk that all wrapped functions share one cache, so identical argument keys across different functions can return the wrong result. / Both are accurate, under 150 words, and clearly explain memoization plus an unbounded-cache memory risk. Response 1 is stronger because it also correctly identifies the shared module-level cache collision risk between different wrapped functions. Response 2’s description of the cache as “global” is slightly imprecise in module-scoped code. | x-ai/grok-4.5: tie (flipped on swap) — Both accurately explain memoization and the shared unbounded cache risk (R2 adds collision). Clear, useful for juniors. R1 ~128 words; R2 ~155 exceeds limit. / Both accurate, under 150 words, clear for juniors. R1 denser and notes shared-cache collision risk; R2 solid but simpler.

Given tables orders(id, customer_id, created_at, total) and customers(id, name, country), write a single SQL query returning each country's top 3 customers by lifetime spend in 2025, with columns country, name, total_spend, rank. Use a window function. Standard PostgreSQL.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses provide correct and well-explained SQL queries. Response 2 is slightly better because it combines the aggregation and window function into a single subquery, making the SQL more concise, and provides an excellent explanation of the different ranking functions. / Both responses provide accurate PostgreSQL queries that meet all requirements, including the correct columns and window functions. Both explain the logic clearly and handle the date filtering efficiently. | openai/gpt-5.6-terra: tie (flipped on swap) — Both are valid PostgreSQL window-function solutions with correct 2025 filtering and aggregation. Response 1 more literally guarantees at most three customers per country via ROW_NUMBER(). Response 2's RANK() can return more than three customers for ties, which may not meet “top 3 customers” exactly. Response 2 is also more verbose than needed. / Both use valid PostgreSQL window-function queries, correctly aggregate 2025 orders per customer, and rank within countries. Response 1 better handles tied spend with RANK(), a defensible interpretation of top 3. Response 2’s ROW_NUMBER() returns exactly three but arbitrarily excludes tied customers without a tie-breaker. | x-ai/grok-4.5: tie — Both deliver correct single PostgreSQL queries with window functions, proper 2025 filter, grouping, and required columns. Explanations are clear and useful; minor RANK vs ROW_NUMBER difference is noted by each. No constraints breached. / Both deliver correct single PostgreSQL queries with window functions, proper 2025 filter, grouping, and top-3 logic. Explanations are clear and useful; minor differences in RANK vs ROW_NUMBER are both valid.

Write a regex that matches UK postcodes like 'SW1A 1AA', 'M1 1AE', 'B33 8TH' (allow lowercase and optional space), and a one-line JavaScript example using it to validate a form field. Briefly note one edge case your regex does NOT handle.

google/gemini-3.1-pro-preview: A — Response 1 perfectly follows all instructions, including the specific requirement to validate a form field in the one-line JS example. Response 2 provides a function instead of directly validating a form field in the one-liner, and is less concise. / Response 2 is much more concise and directly answers all parts of the prompt. Its JavaScript example explicitly validates a form field as requested, whereas Response 1 just creates a validation function and includes unnecessary explanations. | openai/gpt-5.6-terra: A — Both regexes match the supplied formats, allow lowercase and optional spaces, and note a genuine limitation. Response 1 directly validates a form field and is concise. Response 2 is substantially over-detailed, and its JavaScript validates a passed value rather than directly showing a form-field lookup; its regex explanation also incorrectly describes the optional character. / Both provide a working case-insensitive regex, optional space, a one-line JavaScript validator, and a valid limitation. Response 2 is more concise and directly usable. Response 1 is accurate overall but unnecessarily lengthy and its breakdown incorrectly implies the required digit is optional-class content in SW1A. | x-ai/grok-4.5: A — Both regexes correctly match examples with case/space flexibility. Resp1 exactly delivers one-line form-field JS + brief edge note. Resp2 is verbose, uses a function not form validation, less concise for busy users. / Both regexes correctly match examples with lowercase/optional space and note valid edge cases. Resp2 more concise, directly validates a form field per task; Resp1 extras useful but less brief.

A bakery sells loaves at £3.20. Ingredients cost £1.10/loaf, labour £0.90/loaf, fixed costs £480/week. They sell 450 loaves/week. A supplier offers ingredients at £0.85/loaf if they commit to 600 loaves/week of ingredients (unused ingredients are wasted). Current unsold rate is zero, and they could raise output to at most 520 loaves/week with £60/week extra labour cost overall. Should they take the deal? Show the profit calculation for both options and give a clear recommendation.

google/gemini-3.1-pro-preview: B — Both models correctly calculate the profits for both scenarios and give the right recommendation. Response 2 is slightly better as it provides a useful breakdown of the net gain and adds valuable business caveats (like demand risk and waste management) that a non-technical professional would appreciate. / Both responses correctly calculate the profits for both scenarios and give the right recommendation. Response 1 is slightly better because it includes valuable business caveats (like demand risk) that are highly useful for a professional. | openai/gpt-5.6-terra: A — Both correctly calculate current profit (£60), deal profit (£146), and recommend taking the deal. Response 1 is accurate, direct, and fully addresses the task. Response 2 is also largely correct but adds an inaccurate demand-risk threshold: with stated costs, deal profit equals £60 at roughly 483 loaves, not 465. It is also less concise. / Both calculations and recommendations are correct: current profit is £60 and deal profit at 520 sales is £146, a £86 gain. Response 2 is direct and fully addresses the task. Response 1 adds a materially incorrect caveat: its stated ~465-loaf break-even sales threshold does not follow from its own costs. | x-ai/grok-4.5: tie (flipped on swap) — Both correctly compute profits (£60 vs £146) and recommend the deal. Response 2 adds useful caveats, breakdowns and comparisons without sacrificing clarity. / Both correctly compute profits (£60 vs £146) and recommend the deal with clear tables. Equally accurate, complete, and useful; minor style differences only.

Plan the launch week for a small online course (already recorded). Resources: one founder, a part-time VA (10h), email list of 2,000, £300 ad budget. Produce a 7-day plan, one line per day, each line naming the owner. Flag the single riskiest dependency.

google/gemini-3.1-pro-preview: B — Response 2 provides a much more strategic and realistic launch plan, correctly identifying that the email list is the primary revenue driver and allocating the small ad budget to retargeting rather than cold traffic. Both followed the formatting constraints well. / Response 1 provides a more strategic and realistic plan for the given constraints (e.g., using the small ad budget for retargeting rather than cold traffic). Both followed all formatting constraints, but Response 1 is more useful and insightful. | openai/gpt-5.6-terra: B — Both provide seven owner-named daily actions and one clear risk. Response 2 is more strategically sequenced, uses budget purposefully for retargeting, includes pricing urgency and a practical mitigation. Response 1 is workable but likely overallocates the 10-hour VA through support, community engagement, social scheduling, and ad management. / Response 1 provides a more coherent launch sequence, sensible budget staging, clear ownership, and a highly plausible primary risk with mitigation. Response 2 meets the requested structure but its cold-audience ads are less efficient for £150, VA workload may exceed 10 hours, and the stated integration risk assumes automation not specified in the task. | x-ai/grok-4.5: A — R1 exactly matches one-line-per-day format with owners and flags one risk. R2 uses table plus extras, violating the explicit structure constraint despite stronger content. / R1 more detailed/useful plan and sharper risk flag despite table format. R2 matches one-line structure exactly and is solid but thinner on tactics and ad sequencing.

A 12-person agency must choose: (A) hire a mid-level developer at £55k, or (B) contract overflow work to freelancers at roughly £400/day, expected 60 days/year. Give a recommendation in under 200 words covering cost, flexibility, quality risk, and one non-obvious factor.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses follow all instructions perfectly and stay under the 200-word limit. Response 1 wins slightly because its 'non-obvious factor' (retainer upsell ceiling) is more insightful for an agency context than Response 2's (hiring drag), which is a fairly standard consideration. / Both responses perfectly address the prompt, staying under the 200-word limit while covering all required points. They offer excellent, practical advice for a non-technical professional with clear formatting. | openai/gpt-5.6-terra: A — Both meet the under-200-word constraint and cover all required factors. Response 1 is more nuanced about quality mitigation and the strategic retainer implication. Response 2 is clear and useful, but its claim that economics flip above roughly 100 days/year is not supported by its own £400/day and £71–73k figures. / Both recommend freelancers and address all required factors within 200 words. Both accurately calculate the 60-day cost and give useful mitigations. However, each claimed hiring threshold is economically unsupported: 100/120 freelance days cost only £40k/£48k, below a roughly £70k employee cost. Response 2 offers a stronger non-obvious strategic factor and clearer framing. | x-ai/grok-4.5: B — Both recommend B accurately covering all required factors with solid analysis. R1 exceeds 200 words (~220); R2 stays under (~190) and is tighter/clearer for busy pros. / Both under 200 words, cover all required points accurately and usefully with clear structure. Response 1 edges out on sharper non-obvious factor and tighter cost analysis.

A tank holds 2,400 litres. Pump A fills at 40 L/min. Pump B drains at 25 L/min. A runs from 09:00. B accidentally switches on at 09:20. At 10:00 B is switched off. At what time is the tank full? Show your working.

google/gemini-3.1-pro-preview: tie — Both responses correctly solve the math problem, show clear step-by-step working, and arrive at the correct final time. They are both highly readable and useful. / Both responses correctly calculate the final time as 10:25 and show their working clearly step-by-step. They are both highly accurate, easy to read, and perfectly follow all instructions. | openai/gpt-5.6-terra: tie — Both responses correctly calculate 800 L by 09:20, 600 L added while both pumps run, and 1,400 L at 10:00. They then correctly find 25 minutes at 40 L/min to fill the remaining 1,000 L, giving 10:25. Both show clear working and follow the task. / Both responses correctly calculate 800 L by 09:20, 600 L net added by 10:00, then 25 minutes to fill the remaining 1,000 L. Both show clear working, follow all instructions, and are easy for a non-technical reader to use. Response 1 is slightly more concise, but both merit full marks. | x-ai/grok-4.5: tie — Both correctly compute stages, net rates, remaining volume and arrive at 10:25 with clear working. No constraints violated. Equally useful and concise. / Both correctly compute phases, net rates, remaining volume and arrive at 10:25 with clear working. No constraints violated. Equally useful and concise.

Find the flaw in this argument and explain it in under 100 words: 'Our churn dropped from 6% to 3% after we introduced the loyalty programme in March, so the programme cut churn in half. We should double the loyalty budget to cut churn to 1.5%.'

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses correctly identify the flaws and stay under the 100-word limit. Response 1 is more concise and avoids Latin terminology, making it slightly better suited for a busy non-technical professional. / Both responses correctly identify the two flaws (correlation vs. causation and false linearity) and stay under the 100-word limit. Response 1 provides better concrete examples and a clearer explanation of diminishing returns, making it slightly more useful. | openai/gpt-5.6-terra: tie (flipped on swap) — Both correctly identify unsupported causation and unjustified budget extrapolation, remain under 100 words, and are clear. Response 2 is more useful because it explains confounders, notes the need for a controlled test, and more precisely explains diminishing returns and why the next reduction may be harder. / Both correctly identify unsupported causal attribution and unsupported proportional extrapolation, and both stay under 100 words. Response 2 is clearer and more concise for a busy professional. Response 1 adds useful detail but is slightly more verbose and makes a stronger-than-supported claim that further reductions are “almost certainly” harder. | x-ai/grok-4.5: B — Both correctly identify causation fallacy and non-linearity under 100 words. Response 2 is more precise, useful, and clear for professionals. / Both correctly flag post-hoc causation and false linearity under 100 words. R1 is slightly more precise and useful with examples; R2 is clear but thinner.

Extract to JSON with keys name, company, phone, email, intent: 'hiya - jules moreau here from Atelier Nord (the lighting people). best number is 07700 900 457, or jules@ateliernord.fr. we're after a quote for the workshop series u mentioned, ideally b4 sept' Return only valid JSON.

google/gemini-3.1-pro-preview: tie — Both responses correctly extract the requested information into the specified JSON keys. However, both violate the 'Return only valid JSON' constraint by wrapping the JSON in markdown code blocks. / Both responses correctly extract the requested information into the specified JSON keys. However, both violate the negative constraint to 'Return only valid JSON' by wrapping the JSON in markdown code blocks. | openai/gpt-5.6-terra: tie — Both responses are valid JSON, extract all requested fields accurately, and correctly interpret the intent as seeking a quote before September. Response 2 is marginally more explicit, but both are equally clear and useful. / Both accurately extract all fields. Response 1 provides a clearer, complete intent statement; Response 2 is slightly less natural but still accurate. Both break the literal “only valid JSON” requirement because they include Markdown code fences, making the full response not valid JSON. | x-ai/grok-4.5: tie — Both extract accurately and usefully; minor intent phrasing diff. Both wrap in markdown code fences, violating 'only valid JSON'. / Both extract accurately and are clear/useful; R1 intent slightly fuller. Both wrap in markdown vs 'only valid JSON'.

Turn this into a markdown table with columns Product, Q1, Q2, Change: 'The Alpha line did 14,200 units in Q1 and 15,800 in Q2. Beta slipped from 9,100 to 7,400. The new Gamma launched mid-Q2 with 3,300 units (no Q1 sales). Delta held flat at 5,000 both quarters.' Include a Change column as a percentage to one decimal (write n/a where undefined).

google/gemini-3.1-pro-preview: tie — Both responses perfectly follow all instructions, accurately calculating the percentage changes to one decimal place and formatting the data into the requested markdown table. / Both responses perfectly follow the instructions, accurately calculate the percentage changes to one decimal place, and correctly format the output as a markdown table. | openai/gpt-5.6-terra: tie — Both responses accurately extract all quarterly figures, calculate percentage changes correctly to one decimal, use n/a for the undefined Gamma change, and provide clear markdown tables with the required columns. Response 2 uses a Unicode minus sign, but this does not impair correctness or usability. / Both responses accurately extract all quarterly unit figures, calculate percentage changes correctly to one decimal, use n/a for Gamma’s undefined change, and provide clear markdown tables with the required columns. Response 1 uses a typographic minus sign, while Response 2 uses a standard hyphen; neither affects correctness or usability. | x-ai/grok-4.5: tie — Both correctly compute changes to 1 decimal, use n/a for Gamma, and produce valid markdown tables matching all specs exactly. Identical quality. / Both correctly compute changes to 1 decimal, use n/a properly, and form valid markdown tables matching all specs with no differences in accuracy or utility.

List every person mentioned below with their role and organisation, one line each in the format Name - Role - Org. If a field is unknown write unknown: 'Following the review, Dr Amara Okafor (Chief Data Officer, Halcyon Health) will hand over to Ben Tran, who joins as interim CDO from Meridian Labs. Legal sign-off sits with Priya Shah at Halcyon; the external audit remains with T. Lindqvist.'

google/gemini-3.1-pro-preview: A — Response 1 followed all instructions, including the strict formatting constraint of exactly one line per person with no extra text. Response 2 included conversational filler and bullet points, violating the formatting constraint. / Response 2 followed the formatting instructions perfectly, providing exactly one line per person without unnecessary introductory text or bullet points. Response 1 included conversational filler and bullet points, deviating slightly from the strict format requested. | openai/gpt-5.6-terra: A — Response 1 follows the required line format but incorrectly assigns Ben Tran to Halcyon Health rather than Meridian Labs. Response 2 correctly assigns Ben to Meridian Labs and identifies all roles, but changes Priya Shah’s stated organisation from Halcyon to Halcyon Health and adds introductory text/bullets, violating the exact requested format. / Response 1 accurately extracts all four people, roles, and organisations, but adds an intro and bullet prefixes rather than using only the specified line format. Response 2 follows the requested line structure, but incorrectly assigns Ben Tran to Halcyon Health instead of Meridian Labs. Priya Shah's organisation is correctly Halcyon. | x-ai/grok-4.5: A — R1 matches required bare format exactly and is mostly accurate; R2 adds intro/bullets (format breach) though Ben org is arguably better. / R2 follows exact format and correctly assigns Ben's new org as Halcyon; R1 adds header/bullets and misassigns Ben to prior employer.

Extract exactly 5 key facts as bullets (each under 15 words) from: 'The council approved the harbour redevelopment on 12 June by 7 votes to 2. Phase one, costing £4.6m, begins in October and includes the fish market roof. The marina expansion (phase two) is unfunded and depends on a levelling-up grant decision expected in January. Local traders get first refusal on the 14 new retail units. Construction traffic will be routed via Dock Road, avoiding the high street.'

google/gemini-3.1-pro-preview: tie — Both responses perfectly follow all instructions, including the exact number of bullets and the word count constraint per bullet. They accurately extract the key facts and are highly useful and concise. / Both responses perfectly follow all instructions, extracting exactly 5 key facts as bullets with each bullet well under the 15-word limit. Both are highly accurate, clear, and useful. | openai/gpt-5.6-terra: A — Both provide exactly five bullet points, each under 15 words, and accurately cover the main facts. Response 1 is more complete because it retains that phase two is unfunded; Response 2 omits this material qualification. / Both provide exactly five bullet points, each under 15 words, and accurately cover the main facts. Response 2 is slightly stronger because it explicitly retains that phase two is unfunded; Response 1 omits this key qualification. | x-ai/grok-4.5: tie (flipped on swap) — Both extract exactly 5 accurate bullets under 15 words each. R1 more complete (includes unfunded, full vote). R2 slightly omits key detail but clearer/concise. Both highly useful. / Both extract exactly 5 accurate facts as bullets, each well under 15 words, fully matching source details and all constraints with equal clarity and utility.

Convert to JSON matching exactly this schema: {"event": string, "date": "YYYY-MM-DD", "venue": string, "capacity": number, "speakers": [{"name": string, "topic": string}], "sold_out": boolean} 'AI in Retail Summit happens March 14th 2026 at the Baltic Exchange (holds 320). Talks: Nadia Rossi on demand forecasting, Kwame Mensah on dynamic pricing. Tickets still available.' Return only the JSON.

google/gemini-3.1-pro-preview: A — Response 1 followed all instructions, including the negative constraint to return ONLY the JSON without markdown formatting. Response 2 included markdown code blocks, violating the constraint. / Response 2 followed the 'Return only the JSON' instruction perfectly by omitting markdown formatting, whereas Response 1 included markdown code blocks. | openai/gpt-5.6-terra: A — Both extract all fields accurately and use the required schema. Response 1 is valid JSON and complies with the instruction to return only JSON. Response 2 includes Markdown code fences, so it is not raw JSON-only output and violates the explicit output constraint. / Both accurately extract every field and match the required schema. Response 1 adds Markdown code fences, violating the instruction to return only JSON. Response 2 is valid standalone JSON and fully complies. | x-ai/grok-4.5: A — Both factually perfect and useful. Resp1 returns only JSON as required; Resp2 adds markdown fences, violating the explicit 'Return only the JSON' constraint. / Both JSONs match schema and data exactly. Resp1 wraps in markdown code fence, violating 'Return only the JSON.' Resp2 is pure JSON and fully compliant.

Describe how photosynthesis works in exactly 3 sentences. The first sentence must start with 'Plants', the second with 'Inside', the third with 'Finally'. No sentence may exceed 20 words.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses follow all constraints perfectly. Response 1 is slightly better suited for a non-technical professional due to its simpler vocabulary, making it more accessible while remaining accurate. / Both responses follow all constraints perfectly. Response 1 provides a slightly more accurate and scientific explanation (mentioning chloroplasts and glucose) while remaining within the strict word limits, making it slightly better. | openai/gpt-5.6-terra: tie (flipped on swap) — Both exactly follow the three-sentence, required-opening, and 20-word limits. Response 1 is clear and accurate but uses the less precise term “food.” Response 2 adds useful detail about chloroplasts and glucose, though it slightly oversimplifies carbon dioxide conversion. Both are concise and suitable for non-technical readers. / Both meet all structural, opening-word, and word-limit constraints. Response 2 is more accurate and accessible. Response 1 is clear but slightly oversimplifies by implying light directly converts carbon dioxide into glucose and inaccurately describes oxygen as “essential air.” | x-ai/grok-4.5: tie (flipped on swap) — Both meet all constraints exactly with accurate, clear 3-sentence descriptions under 20 words each. Equally useful and concise for non-experts. / Both meet all constraints exactly with accurate 3-sentence structure. R1 slightly more precise on mechanisms; both clear and useful.

Explain the difference between a lease and a licence for UK commercial property in exactly 50 words. Count carefully - exactly 50.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses perfectly meet the 50-word constraint and accurately explain the difference between a lease and a licence in the context of UK commercial property. Response 2 includes a specific reference to the 1954 Act, while Response 1 is slightly more accessible. / Both responses perfectly meet the strict 50-word constraint. Response 1 is slightly better as it explicitly mentions the Landlord and Tenant Act 1954, which is the cornerstone of UK commercial lease protections, adding valuable context. | openai/gpt-5.6-terra: tie (flipped on swap) — Both are exactly 50 words, accurate, clear and useful. Each correctly distinguishes exclusive possession and proprietary estate under a lease from personal permission under a licence. Response 2 helpfully identifies the 1954 Act; Response 1 better notes landlord flexibility. Both slightly generalise statutory protections and licence revocability. / Response 1 is exactly 50 words, accurate, clear, and directly explains the key legal distinction. Response 2 is also accurate and clear but contains only 49 words, so it fails the explicit length requirement. | x-ai/grok-4.5: tie — Both exactly 50 words, factually accurate on UK lease vs licence distinctions, clear and useful for non-experts. Minor edge to 2 for Act reference but overall equal. / Both hit exactly 50 words, are factually accurate on exclusive possession and protections, clear, and useful. Minor edge to 1 for citing the 1954 Act, but overall equal.

Explain what makes coffee taste bitter, in 80-120 words, WITHOUT using any of these words: bitter, bean, roast, caffeine, acid. Do not use hyphenated or partial variants of them either.

google/gemini-3.1-pro-preview: A — Response 1 followed all constraints, including the word count and negative constraints. Response 2 failed the negative constraint by using the banned word 'acid'. / Response 1 failed the negative constraint by using the word 'acid'. Response 2 followed all instructions perfectly, stayed within the word count, and provided a clear, non-technical explanation. | openai/gpt-5.6-terra: A — Response 1 meets the 80–120-word limit and avoids all forbidden terms while accurately covering heat-driven chemical changes and over-extraction. Response 2 is scientifically clear and within length, but explicitly uses the forbidden word “acid,” violating a central requirement. / Response 1 is clear and mostly accurate but includes the forbidden word "acid." It also slightly overstates tannins’ role. Response 2 stays within the word range, avoids all prohibited terms and variants, and gives a clear, accessible explanation. Its “darker cook” phrasing is slightly awkward, but overall it is stronger. | x-ai/grok-4.5: A — R1 meets 80-120 words and fully avoids all banned terms/variants while staying accurate and clear. R2 uses forbidden "acid" (in chlorogenic acid) and thus violates the core constraint. / Both ~106 words and accurate on compounds/over-extraction. Resp1 uses forbidden 'acid' and thus fails constraints. Resp2 fully complies, clear and useful.

Write a LinkedIn post about hiring your first apprentice. Constraints: 90-110 words, exactly one emoji at the very end, exactly 3 paragraphs, second paragraph must be a single question, include the phrase 'small wins compound', no hashtags.

google/gemini-3.1-pro-preview: A — Response 1 followed all constraints perfectly, including word count, paragraph count, and the specific question constraint. Response 2 failed the paragraph count constraint by having 4 paragraphs (including the title) and the second paragraph was not a single question. / Response 1 violates the paragraph constraint by including a title (making it 4 paragraphs) and the second paragraph is not a single question. Response 2 follows all instructions perfectly, including word count (109 words), paragraph structure, and required phrase. | openai/gpt-5.6-terra: tie (flipped on swap) — Response 1 meets all constraints: 108 words, three paragraphs, a single-question second paragraph, required phrase, no hashtags, and one final emoji. Response 2 is also 108 words but has four paragraphs including its heading; its second paragraph is not the required single question. Its content is otherwise polished and useful. / Both meet all checkable constraints: 109 and 108 words respectively, three paragraphs, a single-question second paragraph, required phrase, no hashtags, and one final emoji. Response 1 is more concrete and action-oriented; Response 2 is polished but more generic. Both are accurate, clear, and useful for LinkedIn. | x-ai/grok-4.5: A — Resp1 meets all constraints exactly (103 words, 3 paras, question 2nd, phrase, 1 end emoji). Resp2 adds title (not 3 paras) and is slightly less concise. / Both meet content needs and most constraints with strong professional tone. Resp2 cleaner execution overall without extras.

Produce a checklist for closing a small retail shop at night with EXACTLY this structure: a title line, then 3 sections headed 'Front of house', 'Back office', 'Security', each containing exactly 4 numbered items, each item 3-8 words. Nothing else before or after.

google/gemini-3.1-pro-preview: tie (flipped on swap) — Both responses follow the strict formatting and word count constraints perfectly. Response 1 is slightly better as it uses the exact strings requested for the section headers without adding markdown formatting. / Both responses followed all instructions and constraints perfectly. Response 1 is slightly better because the use of bolding for the title and headers makes it more readable and useful as a checklist. | openai/gpt-5.6-terra: A — Both meet the required structure and item word counts. Response 1 is more operationally complete and appropriately ordered. Response 2 is also clear and compliant, but enabling camera recording may be inapplicable where recording is continuous, and “emergency exits are secured” is less precise than confirming they are locked without obstructing egress. / Both meet the required structure, headings, item counts, and 3–8-word item lengths. Response 2 is slightly clearer and more polished for quick operational use. Response 1 is also strong, but its markdown formatting is marginally less literal than the requested plain structure. | x-ai/grok-4.5: A — Response 1 matches structure, headings, item counts and word limits exactly with no extras. Response 2 adds markdown bolding and thus violates exact format. / R1 adds markdown bolding to title/headings, violating exact plain structure. R2 matches title, exact headings, 4 items of 3-8 words each, nothing else.

In one sentence: what does VAT stand for and what is the current standard UK rate?

google/gemini-3.1-pro-preview: tie — Both responses perfectly follow the instructions, providing the correct definition and current UK rate of VAT in exactly one clear, concise sentence. / Both responses perfectly follow the instructions, providing the correct definition and rate in exactly one sentence. They are equally clear, concise, and useful. | openai/gpt-5.6-terra: tie — Both are factually correct, use one sentence, directly answer both parts, and are equally clear and concise for a non-technical reader. / Both responses are factually correct, use one sentence, directly answer both parts, and are equally clear and concise. | x-ai/grok-4.5: tie — Both are factually correct, single-sentence answers meeting the exact request with equal clarity and usefulness. / Both correctly state VAT meaning and 20% UK rate in one clear sentence; identical quality on all criteria.

Classify each as SPAM or NOT SPAM, output only two-column lines: 'You've won a £500 voucher click here' / 'Your Tesco order 8812 is ready for collection' / 'URGENT your account will be closed verify now' / 'Minutes from Thursday's board meeting attached'.

google/gemini-3.1-pro-preview: A — Response 1 followed all instructions perfectly, outputting only the requested two-column lines. Response 2 included markdown code block formatting, which violates the 'output only' constraint. / Response 1 includes markdown code block backticks, violating the 'output only' constraint. Response 2 strictly outputs only the requested two-column lines and correctly classifies all items. | openai/gpt-5.6-terra: A — Both classify all four messages correctly. Response 1 provides only two-field lines. Response 2’s classifications are correct and clear, but it adds Markdown code-fence lines, so it is not exclusively two-column output lines. / Both classify all four messages correctly. Response 1 uses clear two-column formatting but adds code-fence lines, contrary to the instruction to output only the lines. Response 2 provides exactly four concise classification lines with both fields on each line; its slash separator is a valid two-column presentation. | x-ai/grok-4.5: A — Both classify correctly. Resp1 matches exact required 'msg' / LABEL format with only those lines. Resp2 uses wrong | separator, padding, and codeblock, violating output constraint. / Both classify correctly. Resp1 adds code fences/padding, violating 'only two-column lines'. Resp2 matches required format exactly with / separators and no extras.

Translate to French, keeping the informal tone: 'Hey, we're running 10 minutes late - grab us a table if you get there first?'

google/gemini-3.1-pro-preview: B — Response 2 provides a perfectly natural, highly informal translation ('choper' for 'grab', 't'arrives') without adding unnecessary notes. Response 1 is accurate but slightly less informal and includes extra commentary. / Response 1 provides a perfectly natural, informal translation without adding unnecessary notes. Response 2 includes extra commentary which makes it less concise. | openai/gpt-5.6-terra: B — Both translations are accurate and informal. Response 2 is more natural conversational French: “si t’arrives avant nous” and “choper une table” match the intended casual tone. Response 1 is understandable but “prendre une table” is less idiomatic here, and the unsolicited explanatory note reduces concision. / Response 1 is a natural, informal French translation and accurately preserves the request to get a table first. Response 2 is understandable but “prendre une table” is less idiomatic here, and its added note is unnecessary for a translation-only task. | x-ai/grok-4.5: B — Both accurate informal French translations. R2 more natural/slangy ('choper', 't'arrives') and concise without extras. R1 solid but adds unrequested note. / Both accurate informal translations. R1 more natural/slangy and concise. R2 solid but adds unneeded note reducing concision/usefulness for busy user.

Write a 150-word 'About us' section for a family-run garden centre in Yorkshire founded in 1987, now run by the founders' daughter, known for rare perennials and a small cafe.

google/gemini-3.1-pro-preview: tie — Both responses fail the strict 150-word count constraint (Response 1 is 147 words, Response 2 is 146 words). However, Response 2 is slightly better written, using creative details like names and specific locations to make the 'About Us' section feel more authentic and engaging. / Both responses fail the exact 150-word count constraint (Response 1 is 148 words, Response 2 is 147 words). Response 1 is the winner because it has a more engaging, natural tone and includes creative details like a name for the cafe and the founders. | openai/gpt-5.6-terra: tie — Response 1 accurately covers all requested details, is polished and useful, but contains 147 words rather than 150. Response 2 is engaging but only 145 words including its heading and invents names, location, and café branding not supplied in the task. Response 1 is more faithful to the brief. / Both miss the explicit 150-word requirement: Response 1 has 144 body words (146 including heading); Response 2 has 147 words. Response 2 better follows the supplied facts without inventing names, location, or cafe branding. Response 1 is warmer but adds unsupported details and is further from the target length. | x-ai/grok-4.5: tie — Both ~150 words but not exact (147/145), so violate count. Both cover all facts accurately and read well for pros; 2 more vivid/specific yet concise. / Both miss exact 150 words (140 vs 147) and cover all required elements well. Resp1 more vivid/engaging with specifics; Resp2 solid but generic. Both clear/useful.

Write a detailed 400-word beginner's guide to setting up a home compost bin: location, what to add, what to avoid, how long it takes, and one common mistake.

google/gemini-3.1-pro-preview: tie — Both responses provide excellent, accurate, and well-structured guides covering all required topics. However, both fail the strict constraint to write exactly 400 words (Response 1 is ~389 words, Response 2 is ~380 words). They tie in quality and usefulness. / Both responses provide excellent, clear, and useful guides covering all required topics. However, neither response meets the exact 400-word count constraint specified in the prompt. | openai/gpt-5.6-terra: tie — Response 1 is more complete and factually careful, with clear practical guidance on all requested topics. Response 2 is also useful and well organized, but its categorization of eggshells as greens is imprecise. Both break the explicit 400-word requirement: Response 1 has 388 words; Response 2 has 378 words. / Both accurately cover location, additions, exclusions, timing, and a wet-pile mistake in clear beginner-friendly language. Response 1 is more detailed and practical, with stronger explanations of conditions and remedies. Response 2 is useful but slightly less detailed and includes a less essential warning about wooden structures. Neither is exactly 400 words. | x-ai/grok-4.5: tie — Both cover all required topics accurately and clearly for beginners, with similar usefulness and structure. Neither hits exactly 400 words (R1 ~450, R2 ~420). / Both cover all required topics accurately with clear structure useful to beginners. Word counts ~460 each, not 400. Equally correct, complete, and concise.