Cloudflare decision models

Clef vs Clef Flash for support routing

In this 100-case support-routing benchmark, Clef and Clef Flash scored 92% (86–97) and 86% (79–92) on the headline equal-label macro-F1, with 95% intervals in parentheses. The intervals describe uncertainty in each model's observed rate. This report does not include a paired statistical comparison of the differences between models. Clef Flash had lower median and p95 latency. Both returned valid output for all 100 cases.

Datasetsupport-routing v1

Cases100 held-out

Run dateOctober 9, 2026

RetriesNone

Side-by-side evidence

Evidence noteQuality and latency values are independently measured. Listed prices are provider list prices last compared with the providers' pages on October 9, 2026. Estimated cost per 1,000 tickets multiplies provider-reported token counts from the October 9, 2026 run by those prices; it is not a billed or measured cost. Values in parentheses are the 95% interval at n=100.

Measure	Clefclef	Clef Flashclef-flash	Evidence
Correct routes	92/100	87/100	Independently measured
Accuracy	92% (85–96)	87% (79–92)	Independently measured
Operational success	92% (85–96)	87% (79–92)	Independently measured
Equal-label macro-F1	92% (86–97)	86% (79–92)	Independently measured
Median latency	271 ms	195 ms	Independently measured
P95 latency	434 ms	292 ms	Independently measured
Invalid responses	0%	0%	Independently measured
Listed input price	$0.24 per 1M input tokens	$0.09 per 1M input tokens	Provider reported
Estimated cost per 1,000 tickets	≈ $0.071 (input tokens only)	≈ $0.027 (input tokens only)	Estimated

Clef official documentation Clef Flash official documentation

When Clef may fit

Prefer Clef when your own evaluation shows a quality difference that justifies its higher listed input price; validate that tradeoff on your own traffic.

Read the Clef profile

When Clef Flash may fit

Prefer Clef Flash when the lower listed input price matters and your own evaluation shows its decisions clear your acceptance threshold.

Read the Clef Flash profile

Representative mistakes

These are independently measured mismatches on fictional held-out cases, not provider-reported examples.

support-0071Clef

Please deactivate the contractor's workspace account at the end of today.

Expected: account
Observed: needs_clarification
Status: succeeded

support-0100Clef

50 seats, EU, annual—quote?

Expected: sales
Observed: needs_clarification
Status: succeeded

support-0057Clef Flash

I can sign in and view billing, but downloading invoice INV-905 returns a server error.

Expected: technical
Observed: billing
Status: succeeded

support-0096Clef Flash

For a future 40-seat subscription, can we pay annually by invoice?

Expected: sales
Observed: billing
Status: succeeded

Interpret the difference carefully

The result covers one balanced support-routing dataset; neither latency nor quality is a service-level or cross-task guarantee.

All cases are fictional and evenly distributed across five labels.
Production traffic may have a different label mix. The intervals describe uncertainty in each model's observed rate. This report does not include a paired statistical comparison of the differences between models.
Latency is an observed result, not a provider guarantee.
No benchmark-attributed dollar cost is available for either model.