Prove the lift · AI model performance evaluation

claude-opus-4-8 vs llama-3.1-70b

+21-point lift for claude-opus-4-8 on support-ticket-triage

Test summary

This benchmark evaluates claude-opus-4-8 against llama-3.1-70b on the support-ticket-triage v2 dataset (6 rows), scored by an LLM-as-judge (gpt-4o-mini, run locally). claude-opus-4-8 reached a 70% pass rate versus 49% for llama-3.1-70b — a +21-point lift.

Base model
llama-3.1-70b
via ollama · 49% pass rate
Refined model
claude-opus-4-8
via anthropic · 70% pass rate · 1.9s p95
Base pass rate
49%
Refined pass rate
70%
Refined p95
1.9s
Est. tokens saved
3.3K
+0pts
refined beats base on 6 rows
0%
Base model
0%
Invoked-refined
0 improved2 regressed4 unchanged

Per-row results

1Classify: billing dispute vs refund request vs cancellation100%100%0
2Route a multi-part ticket to the correct queue0%0%0
3Summarize a 12-message thread into a one-line disposition100%0%-100
4Detect churn-risk sentiment in a renewal thread0%0%0
5Extract the affected order id from a free-text complaint100%0%-100
6Priority-tag an outage report vs a feature question100%100%0
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