Compact SLM battle: Microsoft Phi-4 14B ($0.10/$0.30) vs OpenAI GPT-5.4 nano ($0.20/$1.25).
microsoft/phi-4 · microsoft
gpt-5.4-nano · openai
Benchmark
| Workload Scenario | Microsoft Phi-4 (14B) | GPT-5.4 nano | Price Delta |
|---|---|---|---|
| 1M input tokens (raw text) | $0.10 | $0.20 | −$0.10 |
| 1M output tokens (generation) | $0.30 | $1.25 | −$0.95 |
| 1M tokens · 70% input / 30% output mix | $0.16 | $0.515 | −$0.355 |
| Standard chat turn (4K in / 800 out, 50% cached) | $0.00064 | $0.00144 | −$0.0008 |
| Monthly scale (10K requests / day) | $192.00 | $432.00 | −$240.00 |
Negative difference = Microsoft Phi-4 (14B) is cheaper. Positive = GPT-5.4 nano is cheaper.
Scaling Curve
Total cost of a token volume at 70/30 input/output split (uncached). Log-log scale.
Verdict
Microsoft Phi-4 (14B) is cheaper on both input and output rates, so it costs less at every input/output mix. Price alone still isn't the whole decision: capability, latency and context limits (Microsoft Phi-4 (14B): 16.4K, GPT-5.4 nano: 128K) may justify the premium for your task.
FAQ
On input, Microsoft Phi-4 (14B) is cheaper ($0.10/M vs $0.20/M). On output, Microsoft Phi-4 (14B) is cheaper ($0.30/M vs $1.25/M). The same model is cheaper on both sides, so it wins at every mix.
A chat-style request (4,000 input + 800 output tokens, 50% cached) costs $0.00064 on Microsoft Phi-4 (14B) and $0.00144 on GPT-5.4 nano — Microsoft Phi-4 (14B) is 2.3× cheaper for that workload.
Microsoft Phi-4 (14B) supports 16,384 tokens (4.1K max output); GPT-5.4 nano supports 128,000 (16.4K max output). GPT-5.4 nano fits 7.8× more context, which matters for long documents and agents.