Prove the lift · AI model performance evaluation

claude-haiku-4-5 vs gemini-1.5-flash

+18-point lift for claude-haiku-4-5 on insurance-extraction-eval

Test summary

This benchmark evaluates claude-haiku-4-5 against gemini-1.5-flash on the insurance-extraction-eval v2 dataset (8 rows), scored by an LLM-as-judge (llama-3.1-8b). claude-haiku-4-5 reached a 72% pass rate versus 54% for gemini-1.5-flash — a +18-point lift.

Base model
gemini-1.5-flash
via google · 54% pass rate
Refined model
claude-haiku-4-5
via anthropic · 72% pass rate · 1.2s p95
Base pass rate
54%
Refined pass rate
72%
Refined p95
1.2s
Est. tokens saved
12.6K
+0pts
refined beats base on 8 rows
0%
Base model
0%
Invoked-refined
6 improved0 regressed2 unchanged

Per-row results

1Aggregate deductible across a loss sequence under a per-occurrence + annual cap0%100%+100
2Coinsurance 80% clause recovery on an under-insured building0%100%+100
3Subrogation waiver interaction with additional-insured status0%0%0
4Business-interruption period-of-restoration calculation100%100%0
5Replacement cost vs ACV on a depreciated roof0%100%+100
6Named-peril coverage determination under a basic HO form0%100%+100
7Suit-limitation clause: is a late-filed claim contractually barred?0%100%+100
8Pro-rata distribution across layered carriers on a covered loss0%100%+100
Prove the lift on your own data.

Run any two models over your tasks, judged automatically, and see the improvement — in seconds.

Try Invoked
Generated with Invoked · Browse all benchmarks