ModelVerdict

Daniel Stewart

English (US)Email9 fields

Document en-p0017 · layout thread · 9 fields · public part of the dataset · Same invoice in Czech → · Report an error in this example

Email text (the model got exactly this text)

From: Daniel Stewart <daniel.stewart@williams-ross.example>
To: support@johnson-singleton.example
Date: Tue, Apr 21, 2026 at 8:51 AM
Subject: Re: Cancellation

Hello,

Following up on our previous email below, we're still waiting for a solution. We need an answer by the end of the month.

Best regards,
Daniel Stewart
Williams, Ross and Donaldson

On Fri, Apr 17, 2026 at 10:35 AM, support@johnson-singleton.example wrote:

> Hello, thank you for your message. We've logged your request as ticket #48521 and are looking into it.
>
> Johnson-Singleton

> On Fri, Apr 17, 2026 at 5:51 AM, Daniel Stewart <daniel.stewart@williams-ross.example> wrote:

> > Hello,
> >
> > We would like to cancel our order, we no longer need the goods. Order #10296889.
> >
> > Best regards,
> > Daniel Stewart
> > Williams, Ross and Donaldson

Verdict

Cheapest without an error: Gemini 3.5 Flash Lite at $0.46 per 1,000 documents.

  • Gemini 3.8 Flash is also error-free but 7× more expensive.
  • Gemma 4 31B is 3.4× cheaper, but: wrong value (Priority).
  • Mistral Small 4 is 3× cheaper, but: wrong value (Priority).
  • GPT-6 Luna is 3× cheaper, but: wrong value (Sentiment).

Prompt given to every model

You triage the support inbox. Read the email below and return only JSON with these keys:

category        one of: billing (invoices, payments, charges), delivery (where is the order, shipping,
                address change), complaint (damaged, wrong or defective goods, warranty claims),
                technical (login, app or system errors, integrations), cancellation (cancel an order,
                contract or subscription), sales (quotes, prices, interest before buying)
priority        one of: urgent, high, normal, low (rules below)
sentiment       one of: positive, neutral, negative (the customer's tone)
customer_name   full name of the customer who wrote the request
order_number    the customer's order number, if mentioned
invoice_number  invoice number, if mentioned
amount          the amount of money the request is about (number)
currency        ISO 4217 code of that amount (USD, EUR, CZK, …)
deadline        the date by which the customer wants a resolution or answer (YYYY-MM-DD)

Priority rules:
- urgent: the customer cannot work right now, or the deadline is at most 1 day after the email date
- high: the deadline is at most 7 days after the email date, or the customer threatens to cancel
  or escalate (lawyer, ending the business)
- low: the customer says explicitly that it is not urgent
- normal: everything else

Rules:
- A thread: judge the newest message; order numbers, invoice numbers and amounts may come from the
  quoted earlier messages. A forwarded email: the customer is the original sender, not the
  colleague who forwarded it.
- Resolve relative deadlines ("by Friday", "by tomorrow", "by the end of the month") against the
  date of the message that states them.
- Support ticket numbers, tax IDs and phone numbers are not order or invoice numbers.
- If a value is not in the email, return null. Do not guess.
- Amounts as numbers with a decimal point, no thousands separators and no currency symbol.

How we score

Exact match per field after normalising dates, amounts and whitespace. Missing diacritics count as an error. Cost is the real token spend of this run, converted at the rate shown in the footer.

What each model extracted

FieldCorrect valueGemma 4 31B7 / 9 fieldsMistral Small 47 / 9 fieldsGPT-6 Luna8 / 9 fieldsDeepSeek V4.1 Flash8 / 9 fieldsGemini 3.5 Flash Lite9 / 9 fieldsQwen3.8 Flash8 / 9 fieldsMistral Medium 3.58 / 9 fieldsGPT-6 Sol9 / 9 fieldsGemini 3.8 Flash9 / 9 fieldsClaude Sonnet 58 / 9 fields
Categorycancellationcancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)cancellation (correct)
Prioritynormal highwrong value highwrong valuenormal (correct)normal (correct)normal (correct)normal (correct) highwrong valuenormal (correct)normal (correct)normal (correct)
Sentimentneutral negativewrong value negativewrong value negativewrong value negativewrong valueneutral (correct) negativewrong valueneutral (correct)neutral (correct)neutral (correct) negativewrong value
Customer nameDaniel StewartDaniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)Daniel Stewart (correct)
Order number1029688910296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)10296889 (correct)
Invoice number—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Amount—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Currency—— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)— (correct)
Deadline2026-04-302026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)2026-04-30 (correct)
Tokensinput / output733 / 76695 / 76810 / 141955 / 351721 / 961,052 / 945695 / 76810 / 92721 / 7091,016 / 316
of which hidden reasoningbilled as output00822970848033613201
of which from cacheinput billed at a discount4530058607680000
Cost of this documentthis run$0.0001$0.0001$0.0002$0.0003$0.0005$0.0005$0.0016$0.0025$0.0032$0.0052
Per 1,000 documentsfor the same kind of document$0.13$0.15$0.15$0.32$0.46$0.50$1.61$2.54$3.20$5.19
Latencythis run2.6 s0.7 s2.8 s5.9 s0.6 s11.8 s1.3 s2.8 s3 s5.4 s