| Claude Haiku 4.5 | Llama 3.3 70B Instruct | |
|---|---|---|
| Input, per 1M tokens | $1.00 | $0.10 |
| Output, per 1M tokens | $5.00 | $0.32 |
| Context window | 200K | 131K |
| Classification or routing (600 tokens in, 10 out), per 1,000 | $0.65 | $0.0632 |
| Chat reply (2,000 in, 400 out), per 1,000 | $4.00 | $0.328 |
| Retrieval-augmented answer (8,000 in, 500 out), per 1,000 | $10.50 | $0.96 |
| Confidence signal | no | yes |
| Tool calling | yes | yes |
| BANKING77, correct of 400 | 301 | 305 |
Which one should carry your traffic?
If Llama 3.3 70B Instruct answers your requests as well as Claude Haiku 4.5, moving that traffic saves 92% on chat-shaped requests. Averages cannot tell you whether it does for your requests. Llama 3.3 70B Instruct exposes a confidence signal, so Gimbix can let it answer first and hand unsure cases to Claude Haiku 4.5.
Claude Haiku 4.5 profile · Llama 3.3 70B Instruct profile · Measure it on your traffic