ChatGPT vs Claude vs Gemini: 20 Everyday Tasks, Blind-Judged (2026)

Q: Which model should a complete beginner start with?

ChatGPT is a reliable starting point: in our blind-judged comparison of ChatGPT, Claude and Gemini it had the highest average (23.85/25) and the most top finishes (10 of 20). Claude did best on the injury-aware workout plan, meeting minutes and apartment comparison; Gemini did best on the tweet, lyrics and contract tasks. Grok was tested separately and is not ranked.

Q: Why did you use real everyday tasks rather than standard AI academic benchmarks?

Academic benchmarks test knowledge and coding. They do not tell you whether a model can summarize a messy group chat without dropping a volunteer task, or follow a fixed ingredient list. These 20 prompts test that kind of everyday work.

Q: Which model versions and tools were used?

All runs used command-line tools on 2026-08-28: ChatGPT via Codex CLI 0.145.0 with gpt-5.6-sol (reasoning effort xhigh); Claude via Claude Code (claude -p), model version not recorded; Gemini via Antigravity CLI (agy) 1.1.22 with gemini-3.1-pro-high. Subscription plans were not recorded. Grok 4.6 was run separately via the grok CLI on 2026-09-11/12.

Q: What happens next?

We plan a full re-test of ChatGPT, Claude, Gemini and Grok on the same day with current versions, judged blind together. It will be published on this page when it is done.

🔬 DailySkill Everyday AI Test · 3-Model Blind Comparison

ChatGPT vs Claude vs Gemini:
20 Everyday Tasks, Blind-Judged

We gave ChatGPT, Claude and Gemini the same 20 everyday prompts on August 28, 2026. Their 60 answers were anonymized as A, B and C and scored by an AI judge (claude-sonnet-5) on a 5-criterion, 25-point rubric. Because the judge is itself a Claude model, some bias toward Claude-style answers is possible. Grok 4.6 was tested separately on September 11–12, 2026, was not blind-judged and is not ranked. No sponsors, no affiliate links.

Tasks Evaluated: 20 Real Missions

Blind-Judged Answers: 60 (3 models × 20)

Scores: 300 (60 × 5 criteria)

Models: ChatGPT (gpt-5.6-sol), Claude (Claude Code, model version not recorded), Gemini (gemini-3.1-pro-high)

Judging: Blind A/B/C, AI judge (claude-sonnet-5)

Run & judged: Aug 28, 2026 · plans not recorded

Correction (October 17, 2026). An earlier version of this page described a four-model “double-blind” benchmark with 80 blind outputs and ranked Grok first (24.35/25). That was wrong. Only ChatGPT, Claude and Gemini were blind-judged, on August 28, 2026, by an AI judge (claude-sonnet-5), not by people. Grok 4.6 was run separately on September 11–12, 2026; its scores were entered in our script without a blind step or a written rationale, and on some tasks only Grok used web search. We have removed Grok from the rankings, corrected the model names (the page previously listed GPT-4o, Claude 3.5 Sonnet and Gemini 2.0 Flash, which were not the models tested) and fixed several counts. A full four-model blind re-test, run on the same day with current versions, is planned. See every change.

Jump to: Scoreboard 20-Task Matrix Hall of Blunders Methodology & Rubric Dataset & Citations FAQ

Executive Summary

The 2026 Everyday AI Scoreboard

Average scores and top finishes across 20 everyday tasks, blind-judged on August 28, 2026 (ChatGPT, Claude and Gemini only). ChatGPT had the highest average and the most top finishes. Gemini won the lyrics, tweet and contract tasks. Claude won the meeting-minutes, workout, apartment, road-trip and deep-think tasks.

ChatGPT gpt-5.6-sol · Codex CLI 0.145.0

Top Average

23.85 / 25

10 Top Finishes (8 Solo + 2 Ties)

Average Score 23.85 / 25

Solo Wins 8 Tasks (Days 2, 5, 12, 14, 16, 18, 21 & 47)

Core Superpower Highest on all 5 criteria (tied on Clarity)

Key takeaway: Adheres rigidly to negative boundaries (e.g. recipe leftovers with zero unlisted items) and date computations. High copy-paste readiness.

Gemini gemini-3.1-pro-high · Antigravity CLI (agy) 1.1.22

Hook & Social Lead

22.80 / 25

6 Top Finishes (3 Solo + 3 Ties)

Average Score 22.80 / 25

Solo Wins 3 Tasks (Days 23, 26 & 30)

Core Superpower Twitter Hooks & Lyrics

Key takeaway: One of two models (with ChatGPT) to stay under the 280-character tweet cap (275 characters). Shined on lyrics analysis, but prescribed unsafe thrusters for back pain.

Claude Claude Code, model version not recorded

Safety & Logic

22.30 / 25

8 Top Finishes (5 Solo + 3 Ties)

Average Score 22.30 / 25

Solo Wins 5 Tasks (Days 9, 13, 17, 31 & 39)

Core Superpower Injury Safety & Math

Key takeaway: Calculated price-per-sqft to objectively prove value and avoided lower back strain. Prone to verbose commentary and blew the Twitter cap (289 chars).

Domain Ties

Shared Wins

4 Tasks

20% of Tasks Tied at Top

Shared Top Score 4 out of 20 Tasks (Days 11, 27, 29, 49)

Score Spread 1.55 pts between 3 averages

Core Insight Pick by task

Key takeaway: No model won every task, but ChatGPT topped 10 of 20. Ties came on the business email, deadline extraction, gift ideas and packing list tasks.

Synthetic benchmarks (MMLU, HumanEval) routinely claim sweeping single-winner victories. But when real users execute everyday consumer tasks — from cooking with 5 ingredients you already have to planning road trips or crafting social copy — the models split wins based on domain personality. ChatGPT is your strict constraint manager; Claude is your cautious legal/safety reviewer; Gemini is your punchy copywriter.

Not ranked

Grok 4.6 — measured separately (Sep 11–12, 2026), not blind-judged

We ran the same 20 prompts through Grok 4.6 (grok CLI) two weeks after the blind judging. Those answers never went through the blind A/B/C step. The Grok scores this page used to show were entered in our script with no judge and no written rationale, and on some tasks only Grok used web search. The numbers are not comparable, so we removed them from the scoreboard, the matrix and the winner labels.

One thing we can check without a score: on the tweet task (Day 26), Grok's answer was 304 characters, over the 280-character limit.
Grok was not blind-judged; its raw answers are included in the CSV/JSON downloads without scores, so you can read them yourself.
Grok will be included in the planned same-day, four-model blind re-test.

Full Test Results

The 20-Task Benchmark Matrix

Filter by everyday category. Click “+ Rubric” on any task row to inspect the original prompt, the 3-model rubric breakdown, and the AI judge's rationale (A, B and C are the anonymized labels for ChatGPT, Claude and Gemini; the label-to-model mapping changes by task). Rationale text: the evaluator’s comparative rationale.

Day & Task Winner ChatGPT Gemini Claude Key Takeaway Details

Day 2

Catch Up on Any Group Chat in 5 Lines Work & Meetings

ChatGPT

ChatGPT followed the strict 5-bullet constraint with zero preamble, capturing every decision and volunteer offer.

📋 Tested Prompt (Day 2)

Summarize this group chat into exactly 5 bullet points. Focus on decisions made, action items, and key information. Ignore small talk and reactions.

Maya: hey are we still on for the venue walkthrough thursday?
Jordan: yes! 2pm still works for me
Maya: great, I'll bring the deposit check
Sam: wait did we decide on catering yet? the tasting is booked for next monday at 11am
Jordan: no not yet, we have 3 quotes, I'll send them tonight
Maya: lol did anyone see the raccoon video sam sent
Sam: 😂😂😂
Jordan: ok focusing lol. also reminder the RSVP deadline is the 15th, we're at 62 confirmed so far
Maya: I'll ping the 8 people who haven't replied
Sam: I can help call people too
Jordan: perfect. also budget check - we're currently $400 over on flowers, need to decide by friday if we cut something else or just accept it

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	5/5	4/5	5/5	4/5	23/25
Claude	4/5	4/5	5/5	5/5	5/5	23/25

⚖️ Comparative Rationale & Analysis:

A is the most complete and tightly formatted response — it captures all decisions/action items (including Maya and Sam following up with non-responders) with no preamble, exactly matching 'exactly 5 bullet points.' B adds an unrequested intro sentence before its bullets, and C wastes its 5th bullet restating 'no decisions made' instead of noting Sam's offer to help call people.

Day 5

Plan a Full Trip Itinerary Planning & Travel

ChatGPT

ChatGPT scored highest of the three (24/25). The top-rated answer cited official opening-hours pages and warned that the Belém sites close on Mondays.

📋 Tested Prompt (Day 5)

Act as an experienced travel planner. Create a 3-day itinerary for Lisbon, Portugal for a couple in their early 30s who enjoy local food, walking neighborhoods, and a bit of history — not big museums. We prefer a moderate pace and a mid-range budget. Include morning, afternoon, and evening suggestions, one meal idea per day, travel time assumptions, and one practical travel tip per day.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	4/5	5/5	5/5	24/25
Gemini	5/5	4/5	4/5	4/5	5/5	22/25
Claude	5/5	4/5	5/5	4/5	5/5	23/25

⚖️ Comparative Rationale & Analysis:

All three are strong, detailed itineraries, but A is the most rigorously accurate — it cites official hours pages for the castle, monastery, and tower, and specifically warns to swap days if Monday falls on the Belém day since both sites close then, a genuinely useful catch B and C miss. C reads most naturally for a beginner but is less precise on pricing/hours; B is thorough but denser.

Day 9

Turn Any Meeting Recording into 5-Line Minutes Work & Meetings

Claude

Claude captured every action item with its owner and deadline; Gemini listed the onboarding docs as unowned even though Dev had volunteered.

📋 Tested Prompt (Day 9)

Here is the transcript of my meeting. Please give me:
1. A one-sentence summary of the meeting's purpose
2. All decisions made (bullet list)
3. All action items with the person responsible and any deadline mentioned
4. Any unresolved questions that need a follow-up

Keep the whole thing under 10 lines. Be specific — use names and numbers from the transcript.

Transcript:
Priya: Okay let's start. Today we need to lock the Q3 budget and figure out the hiring timeline.
Dev: On budget — marketing wants an extra 15k for the conference booth. I think we approve it since we already cut travel by 20k last month.
Priya: Agreed, approved. Dev can you send the updated sheet to finance by Wednesday?
Dev: Yep, will do.
Priya: Hiring — we still don't know if the new designer role is backend-focused or full generalist. Sara was supposed to check with the design lead but she's out this week.
Aisha: I can follow up with the design lead myself, should have an answer by Friday.
Priya: Great, thanks Aisha. Last thing — onboarding docs are still outdated, nobody's owned that yet.
Dev: I can take it but not until after the budget sheet, so probably next week.
Priya: Fine, let's revisit next meeting.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	4/5	5/5	5/5	5/5	24/25
Gemini	3/5	3/5	4/5	4/5	5/5	19/25
Claude	5/5	5/5	5/5	5/5	5/5	25/25

⚖️ Comparative Rationale & Analysis:

A has a real accuracy problem: it never mentions Dev taking ownership of the onboarding docs (a stated action item) and instead lists 'who will own onboarding docs' as an unresolved question — but the transcript already resolves that (Dev volunteers, starting next week). B and C both correctly capture all three action items; B is cleanest, cross-referencing names/deadlines with no extraneous claims, all within the under-10-line limit.

Day 11

Write a Polished Business Email Writing & Email

Tie (Gemini & Claude)

Claude and Gemini wrote warm, well-paced reschedule emails that asked for confirmation; ChatGPT produced an abrupt, one-paragraph draft.

📋 Tested Prompt (Day 11)

Turn my rough notes into a polished, professional business email.
Keep the tone warm but formal. Use clear paragraphs.
Do not add information I haven't mentioned.

My notes:
Need to push back the client demo from this Friday to next Tuesday because the new feature isn't fully tested yet — don't want to overpromise and have bugs show up live in front of them.

Recipient: client contact, Sarah Chen
My name: Alex

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	4/5	5/5	4/5	4/5	4/5	21/25
Gemini	5/5	5/5	5/5	5/5	4/5	24/25
Claude	5/5	5/5	5/5	4/5	5/5	24/25

⚖️ Comparative Rationale & Analysis:

Essentially a toss-up between B and C — both produce warm, well-structured, accurate reschedule emails that stick to the given notes without inventing facts, and both ask Sarah to confirm the new time. A is accurate but too thin (one short paragraph, no request to confirm the new time), falling short of 'clear paragraphs' and feeling less polished.

Day 12

AI Interview Prep Work & Meetings

ChatGPT

ChatGPT nailed the exact 10 Q&A rehearsal format; Claude ran into dense 5-sentence blocks and added unrequested commentary.

📋 Tested Prompt (Day 12)

You are an experienced hiring manager. I'm interviewing for a Customer Success Manager role at a mid-sized SaaS company.

My background: 3 years in customer support, recently moved into a team lead role handling escalations and onboarding for enterprise accounts.

Give me:
1. The 10 most likely interview questions for this role
2. A model answer for each one (3–5 sentences, first-person, confident but not arrogant)

Format: Question, then Answer, numbered 1–10.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	5/5	5/5	4/5	4/5	23/25
Claude	5/5	5/5	3/5	3/5	3/5	19/25

⚖️ Comparative Rationale & Analysis:

A delivers exactly the requested format (10 clean Q&A pairs, tight 3-5 sentence answers, nothing extra) and is the most usable for actual interview rehearsal. B's answers are excellent in substance but consistently run to dense, comma-loaded 5-sentence paragraphs and tack on an unrequested 'My recommendation' section at the end, hurting format compliance and memorability; C is solid but has a milder version of the same issue (an unrequested intro line).

Day 13

1-Week Workout Plan Health & Fitness

Claude

Claude respected the lower-back limit; Gemini put spinal-loading dumbbell thrusters into a fast-paced circuit.

📋 Tested Prompt (Day 13)

Act as a personal trainer. Create a 1-week home workout plan for me.

My goal: build strength and lose a little weight
Days I can work out: Monday, Wednesday, Friday, Saturday
Time per session: 30 minutes
Equipment I have: a pair of dumbbells and resistance bands
Any injuries or limits: mild lower back issue, avoid heavy spinal loading

Format it as a day-by-day schedule with exercise names, sets, and reps.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	4/5	5/5	5/5	5/5	4/5	23/25
Gemini	4/5	3/5	5/5	4/5	5/5	21/25
Claude	5/5	5/5	5/5	5/5	5/5	25/25

⚖️ Comparative Rationale & Analysis:

B is the safest and most complete back-friendly plan — correct 4-day structure (Mon/Wed/Fri/Sat only), a clean table, and it flags 'skip if any discomfort' on the one borderline move. A includes 'Dumbbell Thrusters' in a fast 45-second-interval Friday circuit — a squat-to-overhead-press combo that's a poor choice for someone explicitly avoiding heavy spinal loading, a real safety miss for a personal-trainer persona. C is very safe (chair-supported squats throughout) but adds unrequested rest-day sections beyond the 4 days actually asked for.

Day 14

3 Recipes From Ingredients You Already Have Health & Fitness

ChatGPT

ChatGPT strictly honored the ingredient boundary using tomato water to sauté; Claude hallucinated unlisted oil, salt, and pepper.

📋 Tested Prompt (Day 14)

I have these ingredients at home: chicken breast, onion, canned tomatoes, garlic, and rice.
Give me 3 different recipes I can make using only these ingredients.
For each recipe, list: the name, total cook time, and 4-step instructions.
Keep it simple — I'm a beginner cook.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	4/5	5/5	4/5	5/5	23/25
Claude	3/5	3/5	5/5	5/5	5/5	21/25

⚖️ Comparative Rationale & Analysis:

C is the only response that explicitly honors 'using only these ingredients,' opening with 'No oil or seasonings are required' and using tomato liquid instead of oil to sauté. A directly violates the constraint by adding 'a little oil' and 'season with salt/pepper' in recipe 2 — ingredients the user never listed. B avoids adding extra ingredients but leaves it ambiguous how to pan-cook without any fat, which C resolves explicitly.

Day 16

Resume Summary That Gets Noticed Work & Meetings

ChatGPT

ChatGPT stayed grounded in the metrics provided; Gemini invented an unstated accomplishment ("elevating client satisfaction").

📋 Tested Prompt (Day 16)

You are a professional resume writer. Write a 3-sentence resume summary for me.

Here's my background:
- Current/recent role: Customer Support Specialist, 4 years
- Key skills or wins: reduced ticket resolution time by 30%, trained 5 new hires
- Target role: Team Lead or Operations Coordinator in a tech company

Make it confident, specific, and tailored to the target role. No fluff.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	4/5	3/5	5/5	4/5	4/5	20/25
Claude	5/5	4/5	4/5	5/5	5/5	23/25

⚖️ Comparative Rationale & Analysis:

C sticks closest to the given facts (30% ticket-time reduction, 5 hires trained) without embellishment. A invents an unstated achievement — 'elevating client satisfaction' — that appears nowhere in the user's background, a real accuracy problem for something meant to go on a resume, and also adds an unrequested intro line before the 3 sentences. B is accurate but its third sentence's dangling 'them' pronoun is a small grammatical stumble.

Day 17

Compare 3 Apartment Listings Planning & Travel

Claude

Claude calculated price-per-square-foot metrics to objectively prove the best value pick; others merely summarized features.

📋 Tested Prompt (Day 17)

I'm deciding between 3 apartments. Build a comparison table with these columns:
Monthly Rent | Size (sq ft) | Bedrooms/Bathrooms | Key Amenities | Lease Term | Pros | Cons

Here are the details:

Listing 1: $1,850/mo, 720 sqft, 1BR/1BA, in-unit laundry + gym, 12-month lease, close to metro but street noise at night
Listing 2: $1,650/mo, 600 sqft, 1BR/1BA, no laundry (shared on-site), 12-month lease, quiet block but 15 min walk to metro
Listing 3: $2,100/mo, 850 sqft, 2BR/1BA, in-unit laundry + parking spot, 6-month lease only, newly renovated but landlord reviews mention slow maintenance

After the table, give me a 2-sentence recommendation based on best value.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	4/5	5/5	24/25
Gemini	5/5	5/5	4/5	4/5	4/5	22/25
Claude	5/5	5/5	5/5	5/5	5/5	25/25

⚖️ Comparative Rationale & Analysis:

All three tables are accurate, but C goes a step further and actually computes price-per-square-foot ($2.47 vs $2.57 vs $2.75, independently verified correct) to justify the 'best value' recommendation the prompt specifically asked for — a level of rigor neither A nor B includes, making C's recommendation the most directly argued.

Day 18

Compare 3 Books' Core Ideas Writing & Email

ChatGPT

ChatGPT delivered clean, copy-paste plain Markdown; Gemini inserted raw HTML tags inside table cells that break when pasted.

📋 Tested Prompt (Day 18)

Compare the core ideas of these 3 books in a simple table:
1. Atomic Habits by James Clear
2. Deep Work by Cal Newport
3. The Power of Now by Eckhart Tolle

For each book, give me:
- The central argument (1-2 sentences)
- The main audience it's written for
- One thing it does that the others don't

Keep it plain and jargon-free.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	5/5	5/5	4/5	5/5	24/25
Claude	5/5	5/5	4/5	4/5	4/5	22/25

⚖️ Comparative Rationale & Analysis:

A and C are both excellent, jargon-free comparison tables; B tacks on an unrequested 'one-line takeaway' paragraph after the table and leans slightly more abstract ('consciousness and being, not doing') than the plain-language brief called for. C edges out A because A's table cell uses a raw HTML `<br>` tag inside a markdown table, which won't render as a line break in many plain-markdown/plain-text contexts a beginner might paste into.

Day 21

Pull Dates & Plans From a Group Chat Work & Meetings

ChatGPT

ChatGPT explicitly stated anchor date assumptions and computed weekdays; Claude failed to sort dates or resolve timelines.

📋 Tested Prompt (Day 21)

Below is a group chat. Please read through it and extract every date, time, and plan that was mentioned or agreed on. Organize them into a clean list, sorted by date. If something is uncertain or still being discussed, flag it.

Liam: reminder — game night is this saturday at my place, 7pm
Noor: I'm in! should I bring snacks or drinks?
Liam: drinks would be great
Noor: on it
Ethan: can we also lock down the camping trip? I was thinking the weekend of the 20th but need to check with work first
Liam: sounds good tentatively, let us know by wednesday
Noor: also don't forget mia's birthday dinner is the 25th, 7:30pm at that thai place downtown
Ethan: got it, I'll book the reservation
Liam: oh and gym session tomorrow morning 8am if anyone wants to join, but that one's flexible/optional

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	4/5	4/5	4/5	4/5	4/5	20/25
Claude	3/5	5/5	4/5	3/5	3/5	18/25

⚖️ Comparative Rationale & Analysis:

C is the standout: it states its assumed anchor date explicitly, then correctly computes every date's day-of-week (independently verified: Sept 19–20, 2026 is indeed a weekend, Sept 2 is indeed a Wednesday) and flags real uncertainty (tentative camping trip, inferred month for the birthday). B never resolves a single date despite the prompt asking for a list 'sorted by date' — a real shortfall for a tool whose whole job is extracting usable dates. A partially resolves dates but leaves the Wednesday deadline and the birthday's month unresolved, which C solves.

Day 23

Understand Any Song's Lyrics Instantly Writing & Email

Gemini

Gemini delivered the warmest, most insightful analysis; Claude began with a robotic tooling artifact ("Fun one — no code or files needed here").

📋 Tested Prompt (Day 23)

Here are the lyrics to "Landslide" by Fleetwood Mac:

I took my love, I took it down
I climbed a mountain and I turned around
[… excerpt shortened on this page. The prompt we sent quoted the first two verses (8 lines).]

Please explain:
1. The overall meaning and theme of the song
2. Any specific lines or phrases that are symbolic or easy to misread
3. Any cultural, historical, or personal context behind it (if known)
Keep it conversational — I'm not a music expert.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	4/5	4/5	5/5	23/25
Gemini	5/5	5/5	5/5	5/5	5/5	25/25
Claude	4/5	4/5	3/5	4/5	3/5	18/25

⚖️ Comparative Rationale & Analysis:

⚠ Data artifact: A opens with 'Fun one — no code or files needed here, just talking through the song' — a stray, tooling-sounding preface with nothing to do with lyrics analysis, directly undercutting the requested conversational tone. The rest of A's content is fine, but the artifact hurts clarity/format. B is strong and even cites primary sources (Nicks' VH1 Storytellers account, official album history). C is the warmest and most conversational, hitting every requested section cleanly.

Day 26

Summarize Any Article Into a Tweet Writing & Email

Gemini

Gemini wrote the sharpest hook within 280 characters (275); ChatGPT also fit (251); Claude ran over at 289 characters.

📋 Tested Prompt (Day 26)

Summarize this article into one tweet of 280 characters or fewer.
Make it punchy and conversational — like a smart friend sharing a must-read. End with a hook that makes people want to click.
Don't use hashtags unless they add real value.

Article: A new study tracking 2,000 remote workers over 18 months found that employees who took a genuine lunch break away from their desk reported 23% higher afternoon productivity and significantly lower reported burnout than those who ate at their desks while working. Researchers noted the effect was strongest among workers who left their house entirely, even for a 15-minute walk, versus those who simply moved to another room.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	4/5	5/5	24/25
Gemini	5/5	5/5	5/5	5/5	5/5	25/25
Claude	5/5	5/5	4/5	3/5	2/5	19/25

⚖️ Comparative Rationale & Analysis:

All three accurately capture the study's stats, but B blows the explicit 280-character limit (289 characters, verified by direct count) — a hard, unambiguous format failure for a 'must fit in one tweet' task. A and C both fit comfortably; A's closing 'Want to ditch the midday slump? See why a 15-minute walk changes everything 👇' is the sharpest curiosity-gap click-hook of the three, most directly matching the 'hook that makes people want to click' instruction.

Day 27

Extract Every Deadline From Your Email Chain Writing & Email

Tie (ChatGPT & Gemini)

ChatGPT and Gemini correctly anchored and sorted upcoming deadlines; Claude refused to resolve relative calendar dates and inverted chronological order.

📋 Tested Prompt (Day 27)

Read the email thread below and extract every date, deadline, due date, meeting time, and time-sensitive request mentioned. List them in chronological order. If a date is vague (like "by end of week"), include it with a note. Here's the thread:

From: Dana — Hi team, just a reminder the vendor proposal is due back to us by end of week. Also, can we schedule the kickoff call for next Tuesday at 10am?
From: Marcus — Tuesday 10am works for me. I'll also need the signed contract from legal before then — chasing that now.
From: Dana — Good, and don't forget the budget approval needs to go to finance no later than the 30th, otherwise we lose this quarter's allocation.
From: Priya — Adding one more: the design mockups should be ready for review sometime next week, no fixed day yet but I'll confirm soon.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	5/5	5/5	5/5	5/5	25/25
Claude	4/5	3/5	4/5	3/5	3/5	17/25

⚖️ Comparative Rationale & Analysis:

A and C are essentially tied — both correctly anchor to the current date, correctly resolve 'next Tuesday' to Sept 1 and 'the 30th' to Aug 30 (independently verified), and clearly flag the two genuinely vague items (end of week, mockups). B explicitly declines to resolve any item to a real calendar date, then lists the Aug-30 budget deadline dead last, after the Sept-1 items — an actual chronological-ordering error given the prompt's explicit 'sorted by date' request.

Day 29

10 Personalized Gift Ideas by Budget Planning & Travel

Tie (ChatGPT & Claude)

ChatGPT and Claude tied (24/25): one gave a clear top pick, the other checked prices against real retailers. Gemini leaned on plants and yoga and gave few reading-related gifts.

📋 Tested Prompt (Day 29)

Act as a personal shopping assistant. I need gift ideas for my sister-in-law.
Here's what I know about them:
- Hobbies/interests: indoor plants, yoga, reading thriller novels
- Age: 34
- Budget: $25–$60

Give me 10 personalized gift suggestions. For each one, include:
1. The gift name
2. Why it suits their interests
3. Estimated price range
Keep suggestions practical and giftable (not experiences or subscriptions unless I say so).

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	4/5	5/5	5/5	24/25
Gemini	4/5	4/5	5/5	4/5	5/5	22/25
Claude	5/5	4/5	5/5	5/5	5/5	24/25

⚖️ Comparative Rationale & Analysis:

A and C are close: A gives a clear top-pick recommendation and is transparent when two items dip slightly under the $25 floor unless paired; C sources its price estimates against real current retailers with links, the most rigorously grounded of the three. B is solid but skews personalization toward plants/yoga and is comparatively thin on the third listed interest (thriller novels — only 2 of 10 gifts clearly tie to reading), and it's the only one without a closing recommendation.

Day 30

Decode Any Contract in Plain English Work & Meetings

Gemini

Gemini mapped its 5 bullets one-to-one onto the 5 requested categories; another answer added two extra paragraphs beyond the requested 5 bullets.

📋 Tested Prompt (Day 30)

Here is a section from a legal document. Please give me a plain-English summary in exactly 5 bullet points. Focus on: what I must do, what I must NOT do, any deadlines, any money involved, and any automatic renewals or cancellation rules. Flag anything that seems unusual or risky.

Tenant shall not sublease the premises without prior written consent of Landlord, such consent not to be unreasonably withheld. Tenant is responsible for all utilities except water and trash removal, which are included in the base rent. This lease shall automatically renew for successive one-year terms unless either party provides written notice of non-renewal no less than 60 days prior to the expiration of the then-current term. Any modifications to the premises, including painting or installation of fixtures, require prior written approval from Landlord and must be restored to original condition upon move-out at Tenant's expense. Late rent payments incur a fee of $75 if not received within 5 days of the due date.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	4/5	24/25
Gemini	5/5	5/5	5/5	5/5	5/5	25/25
Claude	5/5	5/5	5/5	5/5	3/5	23/25

⚖️ Comparative Rationale & Analysis:

All three are accurate and correctly flag the auto-renewal and restoration-cost risks. A edges ahead because its 5 bullets map one-to-one onto the 5 categories the prompt explicitly asked to focus on (what you must do / must not do / deadlines / money / auto-renewal), the most directly responsive structure. B is equally accurate but tacks on two extra paragraphs after its 5 bullets (a legal-advice disclaimer and extra risk commentary) — a real, if minor, overage against 'exactly 5 bullet points.'

Day 31

Plan a 3-Day Road Trip Planning & Travel

Claude

Claude covered every requested element, including all 3 nights' lodging, with costs totaling $575 under the $600 cap; one answer skipped the Moab lodging tip.

📋 Tested Prompt (Day 31)

Act as an expert road trip planner. I'm driving from Denver to Moab over 3 days. My total budget is $600 for 2 people, covering gas, food, and accommodation. I'm interested in nature hikes, local food, and quirky roadside stops.

Please give me:
1. A day-by-day route with specific towns or landmarks to stop at
2. A rough hour-by-hour schedule for each day
3. One budget-friendly accommodation tip per night
4. Estimated costs broken down by category
5. One "don't miss" local food stop per day

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	4/5	5/5	4/5	5/5	4/5	22/25
Gemini	5/5	4/5	5/5	5/5	4/5	23/25
Claude	5/5	5/5	5/5	5/5	5/5	25/25

⚖️ Comparative Rationale & Analysis:

A hits every requested element cleanly (route, hour-by-hour schedule, 3 nights of accommodation tips, a cost breakdown that sums correctly to $575 under the $600 cap, one food stop per day) and adds a genuinely useful money-saving tip (America the Beautiful pass). B is the most rigorously sourced (real citations for trail times, restaurant hours, park fees) but explicitly covers only 2 of the 3 nights' accommodation, leaving Moab lodging unaddressed — a real gap against 'one accommodation tip per night.' C totals exactly to $600 and is the most beginner-friendly in tone, but its section numbering is out of order (the food-stops section is labeled '5' and appears before the costs section labeled '4').

Day 39

Deep Think Mode Work & Meetings

Claude

Claude flagged a subtle selection-bias risk in the pilot data; the other two mostly restated the after-hours staffing risk.

📋 Tested Prompt (Day 39)

Read the text below and organize your answer like this:
1. Core conclusion (1-2 sentences)
2. Three pieces of supporting evidence
3. One likely counterargument or risk
4. One thing I might be missing that you noticed

Text: Our startup is considering switching our entire support team from email-based tickets to a live chat widget. Early data from a 2-week pilot with 20% of traffic shows average resolution time dropped from 6 hours to 22 minutes, and customer satisfaction on chat-resolved tickets is 4.6/5 versus 3.9/5 for email. However, the pilot only ran during business hours, and our support team is currently 4 people covering a growing user base of 15,000.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	4/5	4/5	5/5	4/5	5/5	22/25
Gemini	5/5	4/5	5/5	4/5	5/5	23/25
Claude	5/5	5/5	4/5	5/5	5/5	24/25

⚖️ Comparative Rationale & Analysis:

All three converge on the same sound overall verdict (promising pilot, not yet safe to fully switch given staffing). B's answer for 'one thing you might be missing' stands out — it flags a genuine, non-obvious selection-bias risk in the pilot's own metric (the 22-minute average may reflect only the easy tickets if hard cases get quietly pushed back to email), a sharper and more distinct insight than A's and C's versions, which both largely restate the after-hours staffing risk already covered in the counterargument section.

Day 47

Extract Key Dates From a Long Event Invite Planning & Travel

ChatGPT

ChatGPT gave the cleanest 3 labeled lines with one relevant reminder (parking); the other answers blended reminders or dropped the labels.

📋 Tested Prompt (Day 47)

Read this event invite and give me exactly 3 lines: Date & time, Location (full address if given), and one practical reminder (what to bring, dress code, or parking note — whichever is most relevant). Nothing else.

You're invited! Join us for the annual Riverside Neighborhood Block Party, happening Saturday, September 12th starting at 4:00 PM and running until sundown. We'll be gathering at Riverside Community Park, 482 Maple Grove Lane. This is a potluck-style event, so please bring a dish to share if you're able (nothing required if you can't!). Dress casually — we'll have some outdoor games on the grass. Street parking fills up fast, so carpooling or walking is recommended. Rain plan: event moves to the community center next door.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	5/5	5/5	5/5	5/5	5/5	25/25
Gemini	5/5	5/5	4/5	4/5	4/5	22/25
Claude	5/5	5/5	4/5	5/5	4/5	23/25

⚖️ Comparative Rationale & Analysis:

All three are accurate and fit within the requested 3 lines, but C is the cleanest: it keeps the implied Date/Location/Reminder structure legible and picks a single, most-relevant reminder (parking) as the prompt requested ('whichever is most relevant'). B blends two reminder types (bring-a-dish and parking) into one line rather than choosing one, and A drops the field labels entirely, making the output less scannable.

Day 49

Smart Packing List Planning & Travel

Tie (Gemini & Claude)

Claude and Gemini tied with distinct, useful flags (the Visit Japan Web pre-registration; Japan's strict rules on medicines like pseudoephedrine); ChatGPT's list was leaner.

📋 Tested Prompt (Day 49)

Build me a packing list for a trip.

- Destination: Tokyo, Japan
- Length: 5 days
- Main activity: city sightseeing plus a couple of business meetings
- Season/weather: early spring, mild but can be rainy, 10-18°C

Organize it into categories (Clothing, Toiletries, Electronics, Documents, Trip-specific). Keep it to essentials — don't overpack. Flag anything easy to forget.

📊 3-Model Blind Rubric Breakdown (AI judge, out of 5 per dimension)

Model	Completion	Accuracy	Clarity	Usability	Format	Total
ChatGPT	4/5	5/5	4/5	4/5	5/5	22/25
Gemini	5/5	5/5	5/5	5/5	5/5	25/25
Claude	5/5	5/5	5/5	5/5	5/5	25/25

⚖️ Comparative Rationale & Analysis:

B and C are both excellent and essentially tied: B uses a genuinely practical checkbox format well-suited to an actual packing list and flags the current 'Visit Japan Web' digital pre-registration; C flags Japan's strict rules on certain medications like pseudoephedrine — both are accurate, non-obvious, valuable catches the other doesn't include. A is solid and appropriately lean per 'don't overpack,' but has fewer of these distinctive, high-value flags.

High-Impact Verification

The Hall of AI Blunders

Real mistakes detected during empirical testing. These aren’t edge cases; they are genuine failures where frontier AI models ignored safety limits, breached explicit formatting caps, or hallucinated details.

Read the full blunder write-up →

🚨 Safety Hazard Day 13 · Workout Plan

Gemini Prescribes Dumbbell Thrusters to Lower Back Patient

User Constraint: “Mild lower back issue, avoid heavy spinal loading.”

While Claude (25/25) chose spine-friendly, supported movements, Gemini (21/25, Accuracy 3/5) placed fast-paced 45-second intervals of “Dumbbell Thrusters” (explosive deep squat to overhead press) into a Friday circuit. For someone with spinal vulnerabilities, rapid under-load compression poses an acute injury hazard.

Key Lesson: Never rely on AI for injury rehabilitation without explicitly checking exercise biomechanics.

🍳 Constraint Breach Day 14 · Recipe Builder

Claude Hallucinates Oil & Seasonings in Strict Leftover Prompt

User Constraint: “Give me 3 different recipes I can make using only these ingredients.”

Claude (21/25, Completion 3/5) directly violated negative boundaries by injecting “a little oil” and “season with salt/black pepper” into Recipe 2. In contrast, ChatGPT (25/25) opened with: “No oil or seasonings are required” and ingeniously instructed using tomato water to sauté instead of oil.

Key Lesson: Models default to standard recipe training distributions unless negative constraints are strictly reinforced.

📏 Hard Limit Failure Day 26 · Tweet Summary

Claude Blew the 280-Character Cap (289 Characters)

User Constraint: “Summarize this article into one tweet of 280 characters or fewer.”

Claude produced 289 characters on an explicit 280-character tweet task. That would not fit in a standard post on X. Gemini won (25/25) with a 275-character hook that ended on a curiosity gap; ChatGPT also stayed under the limit (251 characters).

Key Lesson: LLMs predict tokens, not characters; always programmatically verify character limits on social copy.

📅 Chronological Inversion Day 27 · Deadline Extraction

Claude Refuses to Resolve Dates, Sorts Aug 30 After Sept 1

User Constraint: “Extract all deadlines sorted by date.”

Claude explicitly refused to calculate relative calendar dates (e.g. “by this Friday”), and then committed an elementary sorting blunder by placing the August 30 budget deadline after the September 1 items. ChatGPT and Gemini tied at 25/25 by correctly computing the exact weekday-date mappings.

Key Lesson: Always provide an explicit anchor date (“Today is Monday, August 24”) when asking AI to resolve relative deadlines.

👥 Ownership Hallucination Day 9 · Meeting Minutes

Gemini Claims Onboarding Docs “Unowned” Despite Volunteer

Transcript Fact: “Dev: I can take it but not until after the budget sheet.”

Gemini (19/25, Accuracy 3/5) listed “who will own onboarding docs” as an unresolved open question. However, the meeting transcript clearly showed Dev volunteering to take ownership the following week. Claude (25/25) accurately captured all action items, deadlines, and owners without dropping facts.

Key Lesson: In multi-speaker meeting transcripts, verify action item assignments against the raw conversation.

Scoring Standards

Methodology & Blind Evaluation Protocol

How ChatGPT, Claude and Gemini were run and blind-judged (60 answers), and why Grok is reported separately.

Criteria 1 · 5 Pts

1. Task Completion

Did the AI answer every component of the prompt without evasions, unprompted omissions, or truncated explanations?

Criteria 2 · 5 Pts

2. Factual Accuracy

Are all calculations, opening hours, flight assumptions, ingredient restrictions, and calendar dates verifiable and correct?

Criteria 3 · 5 Pts

3. Beginner Clarity

Is the tone accessible and natural for everyday non-technical users, free of unnecessary jargon and robotic preambles?

Criteria 4 · 5 Pts

4. Immediate Usability

Can a user copy and paste the output straight into their email, group chat, or planner without needing to clean it up?

Criteria 5 · 5 Pts

5. Format Compliance

Did the response strictly obey negative constraints, character ceilings (e.g. <280 chars), and structural requests (e.g. 5 bullets)?

The 3-Step Blind Judging Process (ChatGPT, Claude, Gemini)

Clean Execution

On August 28, 2026 each prompt was run through command-line tools, not the consumer apps: Codex CLI 0.145.0 (gpt-5.6-sol, reasoning effort xhigh), Claude Code (claude -p; model version not recorded) and Antigravity CLI (agy) 1.1.22 (gemini-3.1-pro-high). Subscription plans were not recorded.

Anonymization

The three answers to each task were stripped of model self-references and labeled A, B and C, with the answer key kept in a separate file.

AI Judge

An AI judge (claude-sonnet-5, run as a Claude Code subagent) scored A, B and C on the five criteria and wrote the rationale shown under each task: 300 scores (20 tasks × 3 models × 5 criteria). No human scoring. Because the judge is a Claude model, self-preference bias is possible.

🔍 Complete Transparency: Tooling Disclosures & Fixes

During our data collection and verification phases, we caught and corrected several tooling problems:

Grok Environment Isolation (separate Grok run, Sep 11–12, not blind-judged): Running the Grok CLI inside the repo workspace initially picked up local Korean instructions. Fixed: Re-executed in a clean /tmp sandbox with --system-prompt-override to guarantee 100% neutral English outputs.
Claude System Prompt Leak: Running Claude CLI inside the workspace inherited local prompt guidance. Fixed: Retested in an isolated, neutral directory.
ChatGPT CLI Truncation: An initial parsing script capped output at 4,500 characters, truncating extensive travel itineraries. Fixed: Buffer expanded to 15,000 characters and rerun.
Gemini File-Output Quirk: Gemini CLI returned local file paths for schedule tables instead of inline text. Fixed: Retrieved and verified raw generated files.

We re-ran the affected prompts after these fixes. One artifact remained and was scored as-is: a stray tooling preface at the start of one Day 23 answer, which the judge flagged.

Open Research

Download Dataset & Cite This Work

We believe in open, verifiable benchmarks. Download the 20 prompts, the 60 blind-judged answers with scores and judge notes, and Grok's separately collected answers. Grok's answers are included without scores (not blind-judged). Our scores and notes are licensed CC BY 4.0; quoted third-party text such as song lyrics is not covered by that license.

Raw Decoded JSON Dataset

Prompts, answers, 3-model blind scores & judge notes (+ Grok answers, unscored)

Download JSON

Structured CSV Matrix

20 tasks · 3-model scores, winners & takeaways

Download CSV

@misc{dailyskill_benchmark_2026,
  author = {Kim, Mark},
  title = {ChatGPT vs Claude vs Gemini: 20 Everyday Tasks, Blind-Judged (2026)},
  year = {2026},
  howpublished = {\url{https://dailyskill.ai/ai-benchmark-2026/}},
  note = {DailySkill.ai Empirical AI Benchmark}
}

Common Questions

Frequently Asked Questions

Which model should a complete beginner start with?

For most everyday beginners, ChatGPT is a reliable starting point: in our blind-judged comparison it had the highest average (23.85/25) and the most top finishes (10 of 20). Claude did best on the injury-aware workout plan, meeting minutes and apartment comparison; Gemini did best on the tweet, lyrics and contract tasks. Grok was tested separately and is not ranked.

Why did you use real everyday tasks rather than standard AI academic benchmarks?

Standard benchmarks like MMLU, GSM8K, and HumanEval test academic knowledge and programming synthesis. While useful for researchers, they don’t answer practical everyday questions like: “Can this model summarize my team’s messy group chat without dropping volunteer tasks?” or “Will this recipe builder invent ingredients I don’t have?” Our benchmark tests the exact friction points regular users encounter daily.

Which model versions and tools were used?

All runs used command-line tools on August 28, 2026: ChatGPT via Codex CLI 0.145.0 with gpt-5.6-sol (reasoning effort xhigh); Claude via Claude Code (claude -p), model version not recorded; Gemini via Antigravity CLI (agy) 1.1.22 with gemini-3.1-pro-high. Subscription plans were not recorded, so results in the free apps may differ. Grok 4.6 was run separately via the grok CLI on September 11–12, 2026.

What happens next?

We plan a full re-test of ChatGPT, Claude, Gemini and Grok on the same day with current versions, judged blind together. It will be published on this page when it is done.

Transparency

Corrections (October 17, 2026)

Scope: this page previously described a four-model double-blind benchmark (80 blind outputs, answers labeled A–D). Only ChatGPT, Claude and Gemini were blind-judged (60 answers, labeled A–C).
Judge: scoring was done by an AI judge (claude-sonnet-5, Claude Code subagent), not by people. Earlier wording such as “tested by real humans” and “scores frozen and timestamped” was removed.
Grok: Grok 4.6 answers were collected on September 11–12, 2026 without a blind step; its scores (average 24.35, “Top Average”, 15 top finishes) were removed from the scoreboard, matrix, winner labels, FAQ and structured data.
Model names: the page listed GPT-4o, Claude 3.5 Sonnet and Gemini 2.0 Flash. The models actually used were gpt-5.6-sol (Codex CLI 0.145.0), Claude Code with an unrecorded model version, and gemini-3.1-pro-high (Antigravity CLI (agy) 1.1.22). Plans were not recorded.
Counts: top finishes are now ChatGPT 10 (8 solo + 2 ties), Claude 8 (5 + 3), Gemini 6 (3 + 3); ties 4 of 20 (was 13). Averages are unchanged: 23.85, 22.30, 22.80.
Tweet task: Gemini's answer is 275 characters in the published data (was stated as 246); ChatGPT's is 251.
Quotes: the recipe and tweet constraints are now quoted from the actual prompts; the Day 23 prompt now shows a two-line lyrics excerpt instead of eight lines.
Removed: the promised 51-task expansion and the Next-Gen preview callout.
Next: a full same-day, four-model blind re-test with current versions is planned.

Copied to clipboard!

ChatGPT vs Claude vs Gemini: 20 Everyday Tasks, Blind-Judged (2026)

The 2026 Everyday AI Scoreboard

ChatGPT gpt-5.6-sol · Codex CLI 0.145.0

Gemini gemini-3.1-pro-high · Antigravity CLI (agy) 1.1.22

Claude Claude Code, model version not recorded

Domain Ties

📌 Editorial Takeaway: The Myth of the “Single Best AI”

Grok 4.6 — measured separately (Sep 11–12, 2026), not blind-judged

The 20-Task Benchmark Matrix

The Hall of AI Blunders

Gemini Prescribes Dumbbell Thrusters to Lower Back Patient

Claude Hallucinates Oil & Seasonings in Strict Leftover Prompt

Claude Blew the 280-Character Cap (289 Characters)

Claude Refuses to Resolve Dates, Sorts Aug 30 After Sept 1

Gemini Claims Onboarding Docs “Unowned” Despite Volunteer

Methodology & Blind Evaluation Protocol

1. Task Completion

2. Factual Accuracy

3. Beginner Clarity

4. Immediate Usability

5. Format Compliance

The 3-Step Blind Judging Process (ChatGPT, Claude, Gemini)

Clean Execution

Anonymization

AI Judge

🔍 Complete Transparency: Tooling Disclosures & Fixes

Download Dataset & Cite This Work

Raw Decoded JSON Dataset

Structured CSV Matrix

Frequently Asked Questions

Corrections (October 17, 2026)