| Llama 3.3 70B Instruct | Phi 4 | |
|---|---|---|
| Input, per 1M tokens | $0.10 | $0.07 |
| Output, per 1M tokens | $0.32 | $0.14 |
| Context window | 131K | 16K |
| Classification or routing (600 tokens in, 10 out), per 1,000 | $0.0632 | $0.0434 |
| Chat reply (2,000 in, 400 out), per 1,000 | $0.328 | $0.196 |
| Retrieval-augmented answer (8,000 in, 500 out), per 1,000 | $0.96 | $0.63 |
| Confidence signal | yes | no |
| Tool calling | yes | no |
| BANKING77, correct of 400 | 305 | not measured |
Which one should carry your traffic?
If Phi 4 answers your requests as well as Llama 3.3 70B Instruct, moving that traffic saves 40% on chat-shaped requests. Averages cannot tell you whether it does for your requests. Phi 4 does not expose a confidence signal through OpenRouter, so Gimbix would test it as a full replacement rather than a first answer.
Llama 3.3 70B Instruct profile · Phi 4 profile · Measure it on your traffic