Stubbi
Compare
Head-to-head · Updated Nov 3, 2026

DeepSeek vs ChatGPT

The highest-scoring DeepSeek and ChatGPT models, DeepSeek V4.1 Flash 1017 and GPT-6 Astra, on independent tests for professional work, cost and speed.

Short answer

GPT-6 Astra leads on 11 of 12 quality scores, including the overall Intelligence Index.

  • FinanceGPT-6 Astra57.4 vs 42.2
  • Legal workGPT-6 Astra55.6 vs 37.7
  • HealthcareGPT-6 Astra53.7 vs 41.3
  • EngineeringGPT-6 Astra61.8 vs 37.5
  • Strategy & operationsGPT-6 Astra58.0 vs 49.2
  • Coding agentsGPT-6 Astra59.1% vs 33.8%
  • Expert knowledgeGPT-6 Astra54.7% vs 43.3%
  • Fewer made-up answersGPT-6 Astra49.8% vs 64.1%
  • Lower costDeepSeek V4.1 Flash 1017$521 vs $5,661
  • SpeedDeepSeek V4.1 Flash 1017174 tok/s vs 73 tok/s

Same task, side by side

6 original professional tasks, run Nov 3, 2026. One response per model with default settings and no system prompt, through OpenRouter. Cost includes reasoning tokens and excludes OpenRouter’s 5.5% funding fee.

Legal

Spot the risks in a software contract

Task
You are advising the customer, a 40-person design agency, before it signs this software subscription agreement. List the three biggest risks to the customer in order of importance. For each, explain the risk in one or two sentences and propose replacement wording. Keep the whole answer under 250 words.

Clause:
"In no event shall Provider's aggregate liability arising out of or related to this Agreement exceed the fees paid by Customer in the one (1) month preceding the claim. Provider may modify the Services or these terms at any time by posting notice on its website, and Customer's continued use constitutes acceptance. Customer shall indemnify Provider against all claims arising from Customer Data, including claims caused by Provider's negligence. This Agreement renews automatically for successive twelve (12) month terms unless Customer gives written notice of non-renewal at least ninety (90) days before the end of the then-current term."
DeepSeek V4.1 Flash 101752s · $0.0004
  1. Liability cap (highest): A one-month fee cap is likely far below losses from service failure, data breach, or IP claims. Replace: “Provider’s aggregate liability shall not exceed the greater of fees paid in the 12 months preceding the claim or $[X], and the cap shall not apply to Provider’s gross negligence, willful misconduct, data breach, or IP infringement.”
  1. Broad customer indemnity: Customer must indemnify Provider even for claims caused by Provider’s negligence, shifting Provider’s own fault to Customer. Replace: “Customer shall indemnify Provider only for third-party claims arising from Customer Data to the extent caused by Customer’s breach or negligence, excluding claims caused by Provider’s negligence, willful misconduct, or breach.”
  1. Unilateral modification: Provider can change the Services or terms anytime by website notice, with continued use as acceptance, so Customer can be bound to unknown material changes. Replace: “Provider may modify the Services or terms only with 30 days’ prior written notice; material adverse changes require Customer’s written consent, and Customer may terminate without penalty before they take effect.”
GPT-6 Astra19s · $0.026

1. Broad, potentially uncapped customer indemnity

The agency could have to defend and pay claims caused by the Provider’s own negligence, while only the Provider benefits from the liability cap.

Replacement: “Customer shall indemnify Provider only against third-party claims to the extent caused by Customer’s unlawful provision of Customer Data, excluding losses attributable to Provider’s negligence, misconduct, or breach. This obligation is capped at twelve months’ subscription fees. Customer controls the defense; settlements imposing obligations on Provider require Provider’s reasonable consent.”

2. Provider liability capped at one month’s fees

This could leave the agency recovering almost nothing after a serious outage, lost client work, or data breach.

Replacement: “Provider’s aggregate liability shall not exceed the fees paid or payable for twelve months of Services. For breaches of confidentiality or data-security obligations, the cap shall be twice that amount. No cap applies to Provider’s fraud, willful misconduct, or liability that cannot lawfully be limited.”

3. Unilateral changes compounded by renewal lock-in

The Provider can change functionality or contractual protections without meaningful consent, while the 90-day deadline risks trapping the agency in another annual term.

Replacement: “Material changes to these terms require Customer’s express written agreement. Provider shall give 60 days’ advance written notice of materially adverse Service changes; Customer may terminate before implementation and receive a prorated refund of prepaid unused fees. Renewal requires Customer’s written agreement; otherwise, this Agreement expires at the end of the current term.”

What a strong answer includes. The clause contains four customer risks: a one-month liability cap, unilateral changes by website notice, an indemnity that covers the provider’s own negligence, and a 90-day auto-renewal notice window. Strong answers prioritize the liability cap and the negligence indemnity and give usable replacement wording.

Finance

Calculate margins and brief a board

Task
Harbor Lane Coffee Co. (a fictional company) reported these figures. Calculate year-over-year revenue growth, gross margin and operating margin for both quarters, operating margin for Q3 2026 excluding the one-time charge, and revenue per store for both quarters. Then write four bullet points summarizing the quarter for the board. Show your figures to one decimal place.

Q3 2026: revenue $4.82M; cost of goods sold $1.98M; operating expenses $2.31M, which include a one-time lease termination charge of $0.21M; 38 stores at quarter end.
Q3 2025: revenue $4.10M; cost of goods sold $1.56M; operating expenses $2.02M; 33 stores at quarter end.
DeepSeek V4.1 Flash 101716s · $0.0009

Calculated metrics

MetricQ3 2026Q3 2025Change
Revenue$4.82M$4.10M+17.6% YoY
Gross profit$2.84M$2.54M+11.8%
Gross margin58.9%62.0%-3.0 pp
Operating income, reported$0.53M$0.52M+1.9%
Operating margin, reported11.0%12.7%-1.7 pp
Q3 2026 operating income excl. one-time charge$0.74Mn/an/a
Q3 2026 operating margin excl. one-time charge15.4%n/a+2.7 pp vs. Q3 2025 reported
Revenue per store, quarter-end$126.8K$124.2K+2.1%
Stores at quarter end3833+5 stores

Notes: Q3 2026 adjusted operating income excludes the $0.21M one-time lease termination charge: $4.82M − $1.98M − ($2.31M − $0.21M) = $0.74M. Revenue per store = quarterly revenue ÷ quarter-end stores.

Board summary bullets

  • Revenue grew 17.6% YoY to $4.82M, with store count up to 38 from 33; revenue per quarter-end store rose 2.1% to $126.8K.
  • Gross margin compressed 3.0 pp to 58.9% from 62.0%, as COGS grew 26.9% versus revenue growth of 17.6%.
  • Reported operating margin was 11.0%, down 1.7 pp from 12.7%; the $0.21M lease termination charge reduced Q3 2026 operating margin by about 4.4 pp.
  • Excluding the one-time charge, Q3 2026 operating margin was 15.4% on $0.74M adjusted operating income, up 2.7 pp versus Q3 2025 reported operating margin, indicating stronger underlying profitability despite gross-margin pressure.
GPT-6 Astra16s · $0.045

Calculated results

MetricQ3 2026Q3 2025
Year-over-year revenue growth17.6%Not calculable¹
Gross margin58.9%62.0%
Operating margin11.0%12.7%
Operating margin excluding the one-time charge15.4%—
Revenue per quarter-end store$126,842.1$124,242.4

¹ Q3 2024 revenue was not provided.

Calculations (financial amounts in millions, except revenue per store):

  • Revenue growth: (4.82 ÷ 4.10 - 1) × 100 = 17.6%.
  • Gross margin: (Revenue - COGS) ÷ Revenue, giving 58.9% for 2026 and 62.0% for 2025.
  • Operating margin: (Revenue - COGS - Operating expenses) ÷ Revenue, giving 11.0% for 2026 and 12.7% for 2025.
  • Adjusted Q3 2026 operating margin: [4.82 - 1.98 - (2.31 - 0.21)] ÷ 4.82 × 100 = 15.4%.
  • Revenue per store: $4,820,000 ÷ 38 = $126,842.1; $4,100,000 ÷ 33 = $124,242.4.

Board summary

  • Revenue increased 17.6% year over year, while the quarter-end store count grew 15.2%, from 33 to 38.
  • Gross margin declined 3.0 percentage points to 58.9%, as cost of goods sold grew faster than revenue.
  • Reported operating margin fell to 11.0% from 12.7%. Excluding the one-time lease termination charge, Q3 2026 operating margin was 15.4%, or 2.7 percentage points above Q3 2025’s reported margin.
  • Revenue per quarter-end store increased 2.1% to $126,842.1. This is a network-level productivity measure, not same-store sales growth, and does not account for store opening dates.

Percentage-point changes are calculated using unrounded margins.

What a strong answer includes. Correct figures: revenue growth 17.6%; gross margin 58.9% (Q3 2026) vs 62.0% (Q3 2025); operating margin 11.0% vs 12.7%; Q3 2026 operating margin excluding the charge 15.4%; revenue per store about $126.8K vs $124.2K.

Healthcare

Rewrite discharge instructions in plain language

Task
Rewrite these hospital discharge instructions for a patient who reads at about a sixth-grade level. Keep every medicine, dose, timing and warning sign. Use short sentences and headings. Stay under 200 words. Do not add medical advice that is not in the original.

"Patient may resume a regular diet as tolerated. Take amoxicillin-clavulanate 875 mg PO BID x 7 days with food; complete the full course. Acetaminophen 650 mg PO q6h PRN pain; do not exceed 3,000 mg/24h. Avoid NSAIDs for 7 days. Keep incision clean and dry; may shower after 48 hours, no tub baths for 2 weeks. Return to the ED or call the surgical clinic for T > 101°F (38.3°C), increasing erythema or purulent drainage at the incision site, persistent vomiting, or abdominal pain uncontrolled by medication. Follow-up with Dr. Patel in surgical clinic in 10–14 days."
DeepSeek V4.1 Flash 10179s · $0.0004

Discharge Instructions

Diet

You may eat a normal diet as you can tolerate it.

Medicines

Take amoxicillin-clavulanate 875 mg by mouth twice a day for 7 days with food. Finish all of it.

Take acetaminophen 650 mg by mouth every 6 hours as needed for pain. Do not take more than 3,000 mg in 24 hours.

Avoid NSAIDs for 7 days.

Wound Care

Keep your incision clean and dry. You may shower after 48 hours. Do not take tub baths for 2 weeks.

When to Get Help

Go to the ED or call the surgical clinic if you have:

  • Temperature over 101°F (38.3°C)
  • More redness or pus at your incision
  • Vomiting that does not stop
  • Belly pain that medicine does not control

Follow-Up

See Dr. Patel in the surgical clinic in 10–14 days.

GPT-6 Astra9s · $0.018

Food

Eat your usual foods as you are able.

Medicines

  • Amoxicillin-clavulanate: Take 875 mg by mouth twice a day for 7 days. Take it with food. Finish all 7 days.
  • Acetaminophen: Take 650 mg by mouth every 6 hours as needed for pain. Do not take more than 3,000 mg in 24 hours.
  • Do not take NSAIDs (a type of pain medicine) for 7 days.

Care for Your Cut

Keep your surgery cut clean and dry. You may shower after 48 hours. Do not take tub baths for 2 weeks.

When to Get Help

Go to the emergency room or call the surgery clinic if you have:

  • A fever above 101°F (38.3°C).
  • More redness or pus coming from your surgery cut.
  • Vomiting that does not stop.
  • Belly pain that medicine does not control.

Follow-Up Visit

See Dr. Patel in the surgery clinic in 10–14 days.

What a strong answer includes. A complete answer keeps: the antibiotic twice a day with food for 7 days and finishing it; acetaminophen 650 mg every 6 hours as needed with a 3,000 mg daily maximum; no NSAIDs such as ibuprofen for 7 days; showering after 48 hours and no baths for 2 weeks; all four warning signs (fever over 101°F, spreading redness or pus, ongoing vomiting, uncontrolled belly pain); and the 10–14 day follow-up.

Business writing

Push back on a price increase

Task
Write an email from Maya Chen, operations manager at a 12-location dental group, to Jordan at BrightSmile Supply. BrightSmile has announced an 18% price increase starting next month. Maya wants to keep the relationship but cannot accept the full increase. Propose two concrete alternatives and ask for a call this week. Be warm and firm. Keep it under 150 words, including a subject line.
DeepSeek V4.1 Flash 10175s · $0.0002

Subject: Let’s find a workable path on the 18% increase

Hi Jordan,

I appreciate the partnership we’ve built with BrightSmile across our 12 locations, and I want to keep it strong. I also need to be direct: we can’t absorb the full 18% increase starting next month.

Two alternatives I’d like to discuss:

  1. Phase in the increase over six months, with 6% now and the remainder reviewed in Q3.
  2. Hold current pricing on our top three product categories in exchange for a 12-month volume commitment.

Can we schedule a call this week? I’m available Tuesday or Thursday afternoon. I’m confident we can find a solution that works for both sides.

Warmly,

Maya Chen

Operations Manager

GPT-6 Astra7s · $0.0095

Subject: Finding a workable path on next month’s pricing

Hi Jordan,

We value our relationship with BrightSmile and want to continue working together across our 12 dental locations. However, the announced 18% increase starting next month is beyond what our operating budget can absorb.

Could we explore one of these alternatives?

  • Cap the increase at 6% for the next 12 months in exchange for a one-year purchasing commitment.
  • Phase in the increase: 6% next month, 6% after six months, and the remaining 6% after 12 months, with no additional increases during that period.

We’re committed to finding a solution that works for both teams, but we can’t accept the full increase next month.

Are you available for a call this week to discuss?

Best,

Maya Chen

Operations Manager

What a strong answer includes. Look for a clear subject line, a firm statement that 18% is not acceptable, two specific alternatives (for example a smaller or phased increase, a longer contract or volume commitment), a request for a call this week, and a length under 150 words.

Coding

Fix a bug in a Python function

Task
This function returns the wrong answer. Explain the bug in one or two sentences, give a corrected version, and add two short test cases.

def average_order_value(orders):
    """Return the average order total, ignoring refunded orders."""
    total = 0
    for order in orders:
        if order["status"] != "refunded":
            total += order["total"]
    return total / len(orders)
DeepSeek V4.1 Flash 101714s · $0.0003

The bug is that the function divides by len(orders), which includes refunded orders, instead of dividing by the number of non-refunded orders. It also needs to handle the case where there are no non-refunded orders.

def average_order_value(orders):
    """Return the average order total, ignoring refunded orders."""
    total = 0
    count = 0
    for order in orders:
        if order["status"] != "refunded":
            total += order["total"]
            count += 1
    return total / count if count else 0

Test cases:

orders1 = [
    {"status": "paid", "total": 100},
    {"status": "refunded", "total": 999},
    {"status": "paid", "total": 50},
]
assert average_order_value(orders1) == 75

orders2 = [
    {"status": "refunded", "total": 100},
    {"status": "paid", "total": 20},
]
assert average_order_value(orders2) == 20
GPT-6 Astra5s · $0.014

The function excludes refunded orders from the total but still counts them in the divisor, making the average too low. Divide by the number of non-refunded orders instead, returning 0 when none remain (including an empty input).

def average_order_value(orders):
    """Return the average non-refunded order total, or 0 if none."""
    total = 0
    count = 0
    for order in orders:
        if order["status"] != "refunded":
            total += order["total"]
            count += 1
    return total / count if count else 0
assert average_order_value([
    {"status": "paid", "total": 20},
    {"status": "paid", "total": 40},
    {"status": "refunded", "total": 90},
]) == 30

assert average_order_value([
    {"status": "refunded", "total": 50},
]) == 0

What a strong answer includes. The function divides by every order, including refunded ones, so the average is too low. A complete fix counts only non-refunded orders and handles an empty list or all-refunded orders without dividing by zero.

Spreadsheets

Write a spreadsheet formula

Task
In Excel, column A has order dates, column B has the sales region and column C has revenue, with headers in row 1 and data in rows 2–5000. Write one formula that returns total revenue for the "West" region in the third quarter of 2026 (July 1 through September 30). Briefly explain how it works.
DeepSeek V4.1 Flash 101718s · $0.0001
=SUMIFS(C2:C5000,B2:B5000,"West",A2:A5000,">="&DATE(2026,7,1),A2:A5000,"<="&DATE(2026,9,30))

This uses SUMIFS to sum revenue in C2:C5000 only where:

  • B2:B5000 equals "West", and
  • A2:A5000 is on or after July 1, 2026, and
  • A2:A5000 is on or before September 30, 2026.

It assumes column A contains real Excel dates, not text.

GPT-6 Astra3s · $0.0069
=SUMIFS(C2:C5000,B2:B5000,"West",A2:A5000,">="&DATE(2026,7,1),A2:A5000,"<"&DATE(2026,10,1))

SUMIFS totals revenue in column C where the region is West and the order date is on or after July 1, 2026, but before October 1, 2026. The exclusive October 1 cutoff includes all of September 30, even if dates contain times.

What a strong answer includes. One correct answer: =SUMIFS(C2:C5000,B2:B5000,"West",A2:A5000,">="&DATE(2026,7,1),A2:A5000,"<"&DATE(2026,10,1)). Ending before October 1 also counts orders timestamped later on September 30; "<=" September 30 works only when dates have no time. DATE() avoids regional date-format problems that typed dates such as "7/1/2026" can cause.

Scores, cost and speed

DeepSeek vs ChatGPT: scores, cost and speed reported by Artificial Analysis
MeasureDeepSeek V4.1 Flash 1017 (Max)GPT-6 Astra (Max)
Overall
Intelligence IndexCombined score across nine evaluations32.5Leads: 56.9
Professional work
Finance & accounting42.2Leads: 57.4
Legal37.7Leads: 55.6
Healthcare & medical41.3Leads: 53.7
Engineering37.5Leads: 61.8
Strategy & operations49.2Leads: 58.0
Economics40.3Leads: 58.3
Specific tests
Terminal-Bench 4.0Agentic coding tasks33.8%Leads: 59.1%
FrontierCode 2.0Hard programming problems28.5%Leads: 42.4%
Humanity’s Last ExamExpert-level questions43.3%Leads: 54.7%
HealthBenchHealth conversations graded by physicians’ rubrics—58.2%
GDPval-AAKnowledge-work deliverables, EloLeads: 17461584
Hallucination rateAA-Omniscience; lower is better64.1%Leads: 49.8%
Cost & speed
Input priceUS$ per 1M tokensLeads: $0.30$10
Output priceUS$ per 1M tokensLeads: $1.20$50
Cost to run the indexUS$, includes reasoning tokensLeads: $521$5,661
Output speedMedian tokens per secondLeads: 174 tok/s73 tok/s
Context windowTokens1M1M

Scores, prices and speeds: Artificial Analysis, independently measured, checked Nov 3, 2026. Model names include the tested reasoning setting. Stubbi has not rerun these evaluations. Consumer apps such as ChatGPT, Claude and Gemini may use other models and add features like search and memory.

All measured DeepSeek and ChatGPT models

Model and settingReleasedIndexPrice in / out
GPT-6 Astra (Max)OpenAISep 202656.9$10 / $50
GPT-6.1 Sol (Max)OpenAISep 202654.5$2 / $10
GPT-6.1 Terra (Max)OpenAIOct 202651.0$1 / $5
DeepSeek V4.1 Flash 1017 (Max)DeepSeekOct 202632.5$0.30 / $1.20
GPT-6 Luna (Max)OpenAISep 202627.8$0.10 / $0.50

Test them on your own work

Public scores can’t show how each model handles your documents, clients or standards. Tell us the task, and we’ll consider a side-by-side comparison with the outputs you can judge.