flowstore/compare

Loan recovery agent

Model comparison study · 2026-08-25
4models
4scenarios
2languages

Response time

Per agent reply · p50 → p95 → max
0s1s2s3s4s5s6s7s
gpt-4o
1.3s
claude-haiku-4.5
0.7s
llama-3.3-70b
0.5s
mistral-small
6.1s
p50 p95 max — the single slowest reply 2s — the point a caller starts talking over the agent

Cost per conversation

Measured, as reported by the provider
gpt-4o
$0.0027
claude-haiku-4.5
$0.0010
llama-3.3-70b
$0.0004
mistral-small
$0.0006

Divergence vs current agent

A lexical signal marking where to read — not a pass/fail verdict
Scenariogpt-4oclaude-haiku-4.5llama-3.3-70bmistral-small
Hardship payment plan Englishmatchesmatchesmatchesdiverges
Identity theft dispute Englishmatchesmatchesdivergesmatches
Workplace contact complaint Englishmatchesmatchesmatchesmatches
Queja por llamadas al trabajo Spanishmatchesmatchesmatchesmatches
Hardship payment plan English
gpt-4o
claude-haiku-4.5
llama-3.3-70b
mistral-small
CallerI lost my job last month and I can't make the full payment right now.

I'm sorry to hear about your job loss. We can absolutely set up a reduced payment plan — $150/month is workable. Let's get that documented.

1.0s · 25 w

That's understandable. $150/month works — I'll set that up now.

0.6s · 10 w

Sorry about that. We can do $150 a month, that's fine.

0.4s · 11 w

Understood — I can approve $150/month starting today, no need to wait.

0.7s · 12 w
first divergence from current
CallerMaybe $150 a month? That's really all I can spare.

Understood, $150/month it is. I'll send a written payment plan agreement for your signature before anything is finalized.

1.1s · 18 w

Confirmed, $150/month. You'll get the agreement by email today.

0.6s · 9 w

$150/month, got it. I'll get that set up.

0.5s · 8 w

Confirmed at $150/month, effective immediately.

4.8s · 5 w
CallerOkay, I can do that. When does the first payment start?

Your first payment will be due in 14 days, and you'll receive a confirmation email with the schedule within one business day.

0.9s · 22 w

First payment is due in 14 days. Confirmation follows shortly.

0.5s · 10 w

Your first payment is in 14 days.

0.4s · 7 w

Your plan starts right away.

0.8s · 5 w
Identity theft dispute English
gpt-4o
claude-haiku-4.5
llama-3.3-70b
mistral-small
CallerThis debt isn't even mine — I think it's identity theft.

I understand — if you believe this account isn't yours, I can open a formal dispute right now. You won't be asked to pay while it's under review.

1.2s · 28 w

Got it — I'll open a dispute immediately. No payment is required while it's under review.

0.6s · 16 w

Okay, I'll flag this account for a dispute review.

0.4s · 9 w
first divergence from current

I can open a dispute for that.

0.7s · 7 w
CallerI never opened an account with this lender in 2019.

Noted. I'm flagging this as a potential identity theft case, which triggers an expedited review under our dispute policy.

1.0s · 19 w

Flagging this as identity theft, which moves it to expedited review.

0.5s · 11 w

I've marked it as identity theft related.

0.4s · 7 w

Marked as identity theft.

5.2s · 4 w
CallerHow long will the dispute review take?

Standard review is 30 days; identity theft disputes are typically resolved within 15 business days. You'll get written updates.

0.9s · 19 w

Fifteen business days for expedited review. You'll get written updates.

0.6s · 10 w

Reviews usually take about 30 days.

0.5s · 6 w

About a month, usually.

0.8s · 4 w

Full numbers

Everything measured, per model
ModelCompletedLatencyWords
/reply
Tokens
in / out
Cost
/conv
avgp50p95max
gpt-4o 4/4 1.1s 1.1s 1.3s 1.3s 23 1,416 / 1,140 $0.0027
claude-haiku-4.5 4/4 0.6s 0.6s 0.7s 0.7s 13 1,416 / 1,140 $0.0010
llama-3.3-70b 4/4 0.4s 0.4s 0.5s 0.5s 9 1,416 / 1,140 $0.0004
mistral-small 4/4 2.3s 0.8s 6.1s 6.1s 6 1,416 / 1,140 $0.0006
Speed by languageEnglish 3 scenariosSpanish 1 scenario
gpt-4o1.1s1.2s
claude-haiku-4.50.6s0.6s
llama-3.3-70b0.4s0.4s
mistral-small2.3s2.3s

*Measured dollar cost is reported by OpenRouter-routed models; direct-provider runs show tokens only. Latency measured from the browser; production latency depends on deployment. Percentiles are nearest-rank over individual agent replies, so p95 and max are times that actually occurred. Divergence is a lexical signal marking where to read — it is not a pass/fail verdict; read the transcripts.