Gemini wins 4–1

Write a cold email (maximum 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.

openai/gpt-5.6-terra: tie (flipped on swap) — Both meet the 120-word limit, offer a free first month, maintain a friendly professional tone, avoid obvious buzzwords, and include one clear call to action. Response 1 is more brewery-specific; Response 2 explicitly states the six-person firm detail. Response 2’s emoji is slightly less professional, but not materially harmful. / Response 1 meets the sender profile, stays within 120 words, is friendly and professional, and has one clear CTA. Response 2 is concise and brewery-relevant, but it does not establish that the sender is the owner of a 6-person firm, a stated requirement. | x-ai/grok-4.5: A — Both under 120 words, one CTA, professional tone, no real buzzwords. R1 stronger on UK relevance (VAT/excise) and polish; R2 adds emoji and slight casualness reducing professionalism. / Both under 120 words, friendly/pro, one CTA, no buzzwords. R1 specifies 6-person exactly; R2 better UK tone (VAT/excise, accountancy) and tighter. Both useful/clear. | deepseek/deepseek-v4-pro: A — Response 1 is perfectly on target: warm, clear CTA, under word limit, no buzzwords. Response 2 is solid but slightly less concise, goes over 120 words, and the emoji lightly undermines professionalism. / Response 2 better addresses a brewery's specific pain points (VAT, excise duty) showing industry insight, making it more useful. Both follow all constraints, are clear, and have one CTA. Response 1 is slightly more generic.

Write a 60-80 word product description for a stainless-steel insulated water bottle (750ml, keeps drinks cold 24h and hot 12h, leakproof, fits car cup holders). Audience: gym-goers. Do not use the phrase 'stay hydrated' or the word 'sleek'.

openai/gpt-5.6-terra: tie (flipped on swap) — Both meet the 60–80 word limit and avoid prohibited wording. Response 1 is more directly tailored to gym-goers, clearly covers every product feature, and has stronger flow. Response 2 is also accurate and useful, but uses more generic promotional phrasing and adds unsupported “standard” cup-holder wording. / Both meet the 60–80 word requirement, include all specified features, target gym-goers, and avoid forbidden wording. Response 1 is fully grounded in the provided details and is highly clear. Response 2 is also strong, but adds unsupported claims such as “easy to clean” and “slim design.” | x-ai/grok-4.5: tie — Both hit 60-80 words, cover all specs accurately for gym audience, avoid banned terms, and are clear/useful. Nearly identical quality. / Both hit 60-80 words, cover all specs accurately for gym audience, avoid bans, and are clear/useful. Equal quality. | deepseek/deepseek-v4-pro: tie — Both descriptions are factually accurate, meet all constraints (word count, banned terms), and are clear, concise, and tailored to gym-goers. No violations. / Both accurately include specs, avoid banned terms, meet word count, and target gym-goers clearly. Response 2 adds 'easy to clean' and more gym-specific terms like 'grind' and 'daily reps', giving it a slight edge in usefulness, but overall both are excellent.

Turn these meeting notes into a clean five-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 we look stupid if we promo something not shippable. budget is fine. legal still havent signed the claims doc. next check in tues.' Exactly five bullets.

openai/gpt-5.6-terra: B — Both provide exactly five clear, director-ready bullets and accurately capture the delay, campaign disagreement, budget, legal status, and Tuesday check-in. Response 1 slightly overstates legal as a “blocker,” while Response 2 more neutrally reflects the notes and avoids unnecessary introductory text. / Both accurately capture all key points in exactly five bullets and are clear for a director. Response 1 is slightly stronger because it provides only the requested five-bullet summary. Response 2 adds an unnecessary introductory sentence and characterizes legal as a “blocker,” which is plausible but slightly more interpretive than the notes. | x-ai/grok-4.5: tie (flipped on swap) — Both deliver accurate, concise 5-bullet summaries matching notes exactly; equally clear and useful for a director. / Both accurate and clear five-bullet summaries. R1 is pure bullets meeting the exact constraint; R2 adds preamble text. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both accurately summarize the notes in exactly five bullets, are clear and concise, and meet all constraints. No violations. / Both accurate and have five bullets. Response 1 is cleaner, presenting directly without an extra introductory line, which is more suitable for a director's quick scan.

Rewrite this so it is warm, takes responsibility, and keeps every fact identical, 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 expedite it.'

openai/gpt-5.6-terra: B — Both are warm, responsible, accurate, and under 90 words. Response 2 most directly preserves the original facts while clearly taking responsibility. Response 1 is also strong, but “within 3–5 days” is slightly less exact than “in 3–5 days,” and it adds an invitation to contact support. / Both are warm, accountable, clear, and under 90 words. Response 1 preserves all core facts precisely and avoids implying any change in refund timing or handling. Response 2 is also strong, but “I’ve gone ahead and processed” slightly reframes the already-processed refund as a newly completed personal action. | x-ai/grok-4.5: tie (flipped on swap) — Both under 90 words, warm, responsible. R1 slightly alters refund fact to 'I've processed'; R2 keeps all facts identical and is clearer. / Both warm, take responsibility, preserve all facts exactly, and stay under 90 words with clear concise language useful to non-experts. | deepseek/deepseek-v4-pro: B — Both responses accurately preserve all original facts, convey warmth and responsibility, and stay under 90 words. Response 2 is more concise, better structured, and uses straightforward language, which benefits a busy non-technical professional. / Response 1 accurately kept all facts identical, taking responsibility warmly. Response 2 added personal action not in original, slightly altering fact. Both under 90 words, clear and warm.

Write five headlines for a blog post about small UK retailers switching from Shopify to WooCommerce. Each must be under 60 characters. Number them. No colons and no questions.

openai/gpt-5.6-terra: B — Both provide exactly five numbered headlines, all under 60 characters, with no colons or questions. Response 2 is more consistently focused on small UK retailers and the Shopify-to-WooCommerce switch, while Response 1 includes slightly broader phrasing such as independent retailers and small businesses. / Response 1 directly addresses small UK retailers switching from Shopify to WooCommerce in every headline and follows all format constraints. Response 2 meets the explicit format rules, but several headlines are less specific about the switch from Shopify, making it slightly less aligned and useful. | x-ai/grok-4.5: tie (flipped on swap) — Both use Why-headlines (questions) violating ban; else compliant on count/format/length. R2 punchier, more varied, useful. / Both meet all constraints with accurate, numbered headlines under 60 chars and no colons/questions. Resp1 more consistently on-topic and concise for busy pros; Resp2 slightly less focused. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both meet constraints. Response 1 uses varied, catchy phrasing; Response 2 is repetitive with similar structure. Response 1 is slightly more engaging and useful. / Both follow all constraints exactly. Response 2 has slightly more varied and engaging headlines, but Response 1 is slightly more consistently concise. Overall equal quality.

Cut this to half its length without losing any factual content: 'We are pleased to be able to announce that, following a period of extensive consultation with our valued customers and partners across the region, we have taken the decision to extend our opening hours at the Brighton branch. From Monday 6th October, the branch will be open from 8am until 8pm on weekdays, and from 9am until 5pm on Saturdays. We very much hope that these extended hours will make it easier for our customers to visit us at a time that suits them.' State the original and new word counts.

openai/gpt-5.6-terra: B — Both correctly state the original as 84 words; Response 1’s 31-word count is correct but it omits consultation with partners across the region and “extensive” consultation, losing factual content. Response 2 is exactly 42 words, retains the consultation, regional customers and partners, dates, hours, and purpose, and is clear. / Original text is 82 words, not 84. Response 1 is 42 words rather than half (41) and softens the stated hope into a purpose, but retains most details. Response 2 is only 31 words and omits partners, regional scope, and consultation extent; it also changes the stated hope. | x-ai/grok-4.5: B — R2 halves exactly to 42 words while retaining all key facts (consultation with customers/partners, hours, purpose). R1 is shorter but drops partners/region facts and is less precise. / Resp1 exactly halves to 42 words retaining all key facts (consultation, partners/region, hours, ease). Resp2 under-cuts to 31 and drops partners/region details. | deepseek/deepseek-v4-pro: B — Response 2 retains all factual details precisely (e.g., 'extensive consultation with regional customers and partners') and achieves exactly half length. Response 1 is slightly less detailed but still accurate and clear. / Response 1 cuts to exactly 42 words (half), retains all facts, and states counts correctly. Response 2 cuts to 31 words, not half, violating the constraint. Both are accurate and clear, but 1 follows instructions precisely.

Write an email telling a client their project will be two weeks late because we underestimated the integration work. Own it, no blame-shifting, offer one concrete mitigation, keep it under 130 words. Do not use the word 'unfortunately'.

openai/gpt-5.6-terra: B — Both accurately own the underestimation, avoid blame, and offer a concrete staffing mitigation. Response 1 is clear but totals 132 words including subject and sign-off, exceeding the under-130 requirement. Response 2 is within the limit, direct, professional, and gives a specific mitigation and update plan. / Both clearly own the underestimate, state the two-week delay, avoid blame and the forbidden word, and provide a concrete staffing mitigation. Both are polished and useful. Response 2 is 132 words, exceeding the 130-word limit; Response 1 is within the limit. | x-ai/grok-4.5: tie — Both own the delay fully, offer one concrete mitigation (extra senior resource), stay under 130 words, avoid 'unfortunately', and are clear/useful. Nearly identical quality. / Both own the delay fully, offer one solid mitigation, stay under 130 words, avoid banned term, and remain clear/concise for a busy pro. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both emails fully comply, own the mistake, offer concrete mitigation, are under word limit, and avoid 'unfortunately'. They are clear, concise, and useful. / Both own the error, provide concrete mitigation, and avoid 'unfortunately'. Clarity and empathy are strong. Response 2 exceeds the 130-word limit (131).

Write a job advert for a part-time bookkeeper at a 12-person UK design agency, 20 hours a week, hybrid, £32k pro rata. Maximum 180 words. Must include the salary and the hours. No 'rockstar', 'ninja' or 'family'.

openai/gpt-5.6-terra: tie (flipped on swap) — Both meet all explicit requirements: part-time bookkeeping role, 12-person UK design agency, hybrid working, 20 hours, and £32,000 pro rata, with none of the banned terms. Both are under 180 words. Response 1 is more detailed; Response 2 is more concise and equally clear for applicants. / Both meet the word limit, include salary, hours and hybrid working, and avoid banned terms. Response 2 is more complete and scannable for candidates, with clearer responsibilities, requirements and benefits. Response 1 is also strong but offers slightly less role detail. | x-ai/grok-4.5: tie (flipped on swap) — Both meet all constraints (under 180 words, include salary/hours, no banned terms) and are accurate. Resp1 is more complete with clear duties/requirements; Resp2 is concise but thinner on role details. / Both meet all constraints (under 180 words, include salary/hours, avoid banned terms), are accurate, clear, and useful for the role. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both ads include salary and hours, no forbidden words, and are under 180 words. Response 1 is better structured with clear sections, more detail on responsibilities and requirements, making it easier to scan and more useful for a busy reader. / Response 2 exceeds the 180-word limit (181 words), violating a hard constraint. Response 1 is accurate, concise, and fully compliant, making it the winner.

Explain to a non-technical small business owner what a payment chargeback is, why it happens, and what they should do when they get one. Use exactly three subheadings. Maximum 220 words total.

openai/gpt-5.6-terra: tie (flipped on swap) — Both meet the three-subheading and 220-word requirements and are clear for non-technical owners. Response 1 more accurately distinguishes common causes, including billing confusion, and gives a stronger step-by-step response plan. Response 2 is useful but somewhat overstates that funds are immediately pulled and labels accidental non-recognition as “friendly fraud.” / Both use exactly three subheadings, stay under 220 words, and clearly explain the concept, causes, and practical response steps. Response 1 has slightly overbroad claims about immediate fund removal and fees; Response 2's 7–14-day estimate may vary by processor. Both are highly useful and accessible. | x-ai/grok-4.5: tie — Both accurate, use exactly 3 subheadings, under 220 words, clear and useful for non-technical owners. Nearly identical quality. / Both accurate, use exactly 3 subheadings, under/at 220 words, clear and useful for non-technical owners. No meaningful differences. | deepseek/deepseek-v4-pro: tie — Both responses are factually accurate, meet the subheading and word count constraints, and are clear and useful for a non-technical owner. They cover the same ground with minor structural differences, making them essentially equally effective. / Both responses are accurate, use exactly three subheadings, and stay within the word limit. They clearly explain chargebacks, reasons, and actions, catering to a non-technical owner. No factual errors, making them equally useful.

Here is a brand's voice: short sentences, dry humour, never exclamation marks, addresses the reader as 'you', British spelling. Write a 70-word homepage intro in that voice for a company that repairs vintage watches.

openai/gpt-5.6-terra: B — Response 1 strongly matches the voice but has only 63 words, not 70. Response 2 is exactly 70 words and follows the requested voice: short sentences, dry humour, direct address, British spelling, and no exclamation marks. It is also more immediately engaging for a homepage. / Response 1 is exactly 70 words, uses British spelling, addresses you, has dry humour, and clearly explains the service. Response 2 has an appropriate tone and clear message but is only 63 words, failing the explicit 70-word requirement. It is also less distinctly dry and concise. | x-ai/grok-4.5: B — Resp2 hits exact 70 words, full voice (British spelling, dry humour, short sentences, you-address, no !). Resp1 is only 63 words so breaches count; otherwise solid match. / R1 hits exact 70 words, full voice (short dry humour, British spelling, you, no !). R2 is 63 words, weaker humour and longer sentences. | deepseek/deepseek-v4-pro: B — Both nail the voice: dry humour, short sentences, 'you' address, British spelling, no exclamation marks. Response 1 is 63 words, breaking the 70-word rule; Response 2 hits 70 exactly. Otherwise equally clear, useful, and on-brand. / Response 1 nails the voice, hits exactly 70 words, and follows all constraints. Response 2 is well-written but falls short at 63 words, violating the explicit word count.

Write a public reply to this 2-star review: 'Food was fine but we waited 50 minutes for mains on a Tuesday with 6 tables occupied. Nobody said anything until I asked.' Acknowledge the specific failure, do not offer a generic apology, invite them back once, under 80 words.

openai/gpt-5.6-terra: tie (flipped on swap) — Both directly acknowledge the 50-minute wait and lack of proactive communication, avoid a generic apology, invite the guest back once, and remain under 80 words. Response 1 is more personal; Response 2 is slightly more concise. / Both meet the under-80-word limit, acknowledge the 50-minute delay and lack of proactive communication, and invite the guest back once. Response 1 is more concise and polished for a public reply. Response 2 is also strong but slightly more conversational and includes unnecessary framing about a “busy-night excuse.” | x-ai/grok-4.5: tie — Both acknowledge the exact wait/comms failure without generic apology, invite once, and stay under 80 words. Equally clear, accurate, and useful. / Both acknowledge the exact wait/comms failure without generic apology, invite back once, and stay under 80 words. Equally clear, useful, and compliant. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Response 1 is more personal and directly takes ownership with 'I'd like to personally make sure,' while still being specific, concise, and under 80 words. Response 2 is also excellent but slightly less personal. / Both responses acknowledge the wait and communication failure specifically, avoid generic apologies, include one invitation, and stay under 80 words. They are equally correct, useful, and clear. No constraints broken.

Rewrite these three features as benefits for a small e-commerce owner, one sentence each, no more than 20 words each: '256-bit encryption', 'REST API with webhooks', '99.95% uptime SLA'.

openai/gpt-5.6-terra: A — Both provide three benefit-focused sentences under 20 words each. Response 1 more clearly conveys real-time automation from webhooks and is more directly relevant to e-commerce operations. Both overstate uptime by implying downtime can never cause lost sales. Response 2 also overstates encryption’s ability to prevent breaches. / Both provide three clear, under-20-word benefit statements and accurately translate the features into business value. Response 2 is more concise and direct. Both slightly overstate the uptime SLA by implying downtime can never cause a lost sale, but Response 2 otherwise has stronger, more specific operational wording. | x-ai/grok-4.5: tie — Both correctly rewrite features as clear benefits in one <=20-word sentence each, highly useful and concise for the audience with no differences in quality. / Both accurately convert features to owner benefits in one sentence under 20 words each. Equally clear, useful, and compliant with all constraints. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both are accurate, concise, and follow the word limit. Response 1's 'turns visitors into repeat buyers' more directly highlights a revenue benefit, making it slightly more useful for a busy owner. Otherwise, both are excellent. / Both responses accurately rewrite each feature as a benefit, each under 20 words, perfectly following instructions. They are clear, useful, and factual with no constraint violations.

Summarise the following in exactly 40 words, no more, no fewer: 'A study of 1,240 UK small businesses found that those adopting automated invoicing reduced late payments by 23% on average within six months, but that firms with fewer than five employees saw no significant change, largely because their invoice volume was too low for the effect to register.' State the word count at the end.

openai/gpt-5.6-terra: tie — Both summaries contain exactly 40 words before the required word-count statement, accurately retain the key findings, and are clear for non-technical readers. Both omit “on average,” a minor loss of precision. Response 1 more closely preserves “fewer than five”; Response 2 is equally concise and clear. / Both summaries contain exactly 40 words before the stated word count, preserve the study size, UK context, 23% reduction, six-month timeframe, and lack of significant change for firms under five employees due to low invoice volume. Both are clear, concise, and suitable for non-technical readers. | x-ai/grok-4.5: tie — Both deliver accurate 40-word summaries of key findings, state the count, and meet all constraints equally well with clear, concise phrasing suitable for busy readers. / Both deliver accurate 40-word summaries matching source facts and end with the required count statement. | deepseek/deepseek-v4-pro: tie — Both summaries are factually accurate, exactly 40 words, state the word count, and are clear and useful. No violations. / Both summaries are factually correct, exactly 40 words, state word count, and are clear and concise. No violations.

Explain the same product update — 'we now support multi-currency invoicing' — twice. First for an existing customer in one sentence. Then for a finance director evaluating us, in three sentences. Label them A and B.

openai/gpt-5.6-terra: B — Both follow the required A/B structure and sentence counts. Response 1 adds more unsupported specifics, including automatic real-time rates, reconciliation reduction, forecasting, and compliance. Response 2 also makes unverified benefit claims, but is less specific and more clearly tailored to a finance director. / Both meet the required A/B structure and sentence counts. Response 1 is clear and audience-appropriate, though it adds unverified claims about dashboards, exchange-rate calculations, and consolidated reporting. Response 2 is similarly useful but makes stronger unsupported claims about automatic real-time rates, error reduction, forecasting, and local compliance. | x-ai/grok-4.5: tie (flipped on swap) — Both meet exact A/B structure, 1-sentence and 3-sentence counts, and convey the update accurately. Response 1 is slightly more precise and useful for a finance director on controls and compliance. / Both correctly deliver one-sentence A and three-sentence B explanations of the update, properly labeled, clear, and useful for the audiences. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both meet constraints and are accurate. Response 1's finance director explanation emphasizes cash flow forecasting and compliance, directly useful for that audience. Response 2 is clear but less tailored. / Both meet all constraints, are accurate, clear, and useful. Response 2 adds slight detail on automation and compliance, but both are equally strong.

Rewrite this so it contains no marketing filler and only checkable statements: 'Our revolutionary AI-powered platform leverages cutting-edge machine learning to deliver unparalleled insights that transform how forward-thinking businesses unlock growth at scale.' If a claim cannot be made checkable, drop it and say what you dropped.

openai/gpt-5.6-terra: B — Both remove most subjective and unmeasurable language and explain omissions. Response 2 is more concise and maps the original claims to dropped wording more directly. Response 1 is useful but does not explicitly address “AI-powered” or “leverages,” and “business insights” remains somewhat undefined. Both rewritten claims are potentially verifiable. / Response 1 gives a plainly verifiable functional claim and clearly identifies each dropped marketing claim with sound reasons. Response 2 is thorough, but retains “generate business insights,” where “insights” is undefined and qualitative, so the rewritten claim is not fully checkable. Its extra implementation suggestions are useful but unnecessary. | x-ai/grok-4.5: A — Both correctly strip unverifiable claims and list drops with reasons. Response 1 is more thorough, adds verification tips, slightly clearer structure for busy users. / Both correctly drop unverifiable claims and rewrite to checkable core while listing drops. Response 2 is more thorough and useful with verification tips; both clear and follow instructions fully. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both responses accurately drop marketing filler and produce a checkable statement. They list dropped phrases with clear reasons. Response 1 adds helpful specifics on verification; Response 2 is slightly more concise. No constraints were broken. / Response 1 gives a concise, purely checkable statement and clearly lists dropped claims. Response 2 adds an extra section on making claims more checkable, which, while not incorrect, slightly reduces concision and was not requested. Both follow instructions, but Response 1 is more direct and useful for a busy professional.

Write the subject lines and one-sentence bodies for a three-email launch sequence for a £49 online course on Amazon PPC. Emails go out on days 1, 3 and 6. Each subject line under 45 characters. Label the day on each.

openai/gpt-5.6-terra: B — Both meet the required three-day structure, labels, one-sentence bodies, £49 positioning, and sub-45-character subjects. Response 2 is clearer and more consistently benefit-led. Response 1 includes unsupported, overly specific claims (£3,000 mistake, "we use," price rises "for good") and implies a closing deadline not provided by the task. / Response 1 cleanly meets every format requirement and offers clear, benefit-led copy. Its “permanently slash” claim is somewhat overstated. Response 2 also meets constraints and is engaging, but introduces unsupported specifics (£3,000 mistake, closing tonight, permanent price rise) that may be inaccurate or misleading without task context. | x-ai/grok-4.5: A — Both meet all constraints with under-45-char subjects, one-sentence bodies, and day labels. R1 subjects more compelling and specific; R2 solid but slightly generic. / Both meet all constraints with under-45-char subjects, one-sentence bodies, and day labels. R2 edges out on stronger urgency and specificity for a busy pro. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Both follow all constraints: correct days, subject lengths under 45 characters, one-sentence bodies, labels. Equally clear and useful for a professional. / Both follow all constraints perfectly. Response 2's subject lines (with a concrete £3,000 mistake and urgent 'closes tonight') are punchier for a busy professional, giving it a slight edge in usefulness and clarity.

Write a LinkedIn post about why most A/B tests on small e-commerce sites never reach significance. Between 90 and 110 words. No hashtags. No 'thoughts?' at the end. Open with a claim, not a question.

openai/gpt-5.6-terra: B — Response 1 is accurate, actionable, and well written, but has 116 words, exceeding the 90–110 limit. Response 2 meets the word count, but its six-month sample-size claim is materially understated for a 2% baseline and 5% relative uplift, and its claims about data being “entirely polluted” are overstated. / Response 2 is more accurate, actionable, and clearly explains common testing mistakes. It exceeds the 90–110 word limit at 118 words. Response 1 meets the length requirement and is useful, but its six-month estimate for detecting a 5% lift with 10,000 monthly visitors substantially understates the likely duration. | x-ai/grok-4.5: tie (flipped on swap) — Both hit 105 words, open with claims, omit hashtags/thoughts. Accurate stats and advice; #1 punchier and more precise on sample sizes needed. / Both factually accurate, clear, useful posts opening with claims and meeting style rules. R1 hits 105 words exactly in range; R2 hits 119 and exceeds limit. | deepseek/deepseek-v4-pro: B — Both posts are accurate, clear, and useful. Response 1 exceeds the word limit (118 words), but per instructions I did not deduct for this. Response 2 follows all constraints. Content quality is similar, so scores are tied. / Response 1 is accurate and meets all constraints (word count 105, no hashtags, no 'thoughts?'). Response 2 exceeds the 110-word limit at 118 words, breaking an explicit constraint.

Compress this into three bullets, preserving every number exactly: 'Q3 revenue was £412,000, up 8% year on year. Gross margin fell from 61% to 57% because of increased shipping costs. Headcount rose from 14 to 17, and we opened the Manchester office in August, which contributed £18,000 of the quarter's revenue.'

openai/gpt-5.6-terra: A — Response 1 preserves all facts and numbers exactly in three clear bullets. Response 2 preserves every number and uses three bullets, but omits that shipping costs were increased and does not specify that the £18,000 was quarter revenue, reducing factual completeness. / Both use exactly three bullets and preserve every number exactly. Response 2 more faithfully retains key qualifiers, including “increased” shipping costs and that £18,000 was quarter revenue. Response 1 is concise and clear but slightly loses those details. | x-ai/grok-4.5: tie — Both compress to exactly three bullets and preserve all numbers (£412,000, 8%, 61%, 57%, 14, 17, £18,000) with full fidelity and clarity. / Both correctly compress to exactly three bullets while preserving all numbers and key facts accurately, clearly, and usefully with no material differences. | deepseek/deepseek-v4-pro: tie (flipped on swap) — Response 1 fully preserves all details including 'increased shipping costs', while Response 2 omits 'increased'—slightly reducing precision. Both follow bullet and number constraints, but Response 1 is slightly more accurate and useful. / Both responses use exactly three bullets and preserve all numbers exactly. They are factually accurate, clear, and useful. Response 2 is slightly more faithful to the original wording, but Response 1 is equally effective and concise.