GPT-5.4 sits in no-man's land: too expensive to be cheap, too weak to justify the premium

GPT-5.4 (OpenAI) occupies an awkward position in the current model landscape. At 56.8 quality and $5.63/M tokens, it delivers marginally less than Gemini 3.1 Pro's quality while costing 25% more, and sits 3.4 points below GPT-5.5 without offering meaningful savings relative to that gap. For most production workloads, GPT-5.4 is the model you skip.

The pricing squeeze

The problem is straightforward. Gemini 3.1 Pro Preview (Google) scores 57.2 quality at $4.50/M tokens. GPT-5.4 scores 56.8 at $5.63/M tokens. That's 0.4 points less quality for $1.13 more per million tokens. In any scenario where you're processing volume — batch summarization, RAG pipelines, document classification — that $1.13 compounds into real budget pressure with no quality upside.

The speed differential makes it worse. Gemini 3.1 Pro runs at 142 tok/s versus GPT-5.4's 85 tok/s. That's 67% faster inference. For latency-sensitive applications like interactive agents or streaming UIs, Gemini completes responses in roughly 60% of the wall-clock time.

Where GPT-5.4 might still make sense

I can construct one narrow case. If you're locked into OpenAI's API ecosystem — fine-tuned models, existing prompt libraries, specific function-calling behavior — and GPT-5.5's $11.25/M is too expensive for your throughput, then GPT-5.4 is your best option within that vendor. It's the cheapest OpenAI model above 55 quality.

That's not nothing. Vendor switching costs are real. Prompt engineering that exploits OpenAI-specific behaviors (tool use formatting, system message handling) doesn't port cleanly. If your team has invested months tuning prompts for OpenAI's instruction-following style, paying the $1.13 premium over Gemini might be cheaper than rewriting and revalidating.