OpenAI's GPT-5.4 mini variant with 128k max output tokens, GPQA Diamond 87.5, priced at $0.75/$4.50 per 1M tokens via OrcaRouter.
openai/gpt-5.4-mini-2026-03-17 is a smaller, more cost-effective variant of OpenAI's GPT-5.4 model, released on March 17, 2026. It is hosted on OrcaRouter and accessed via an OpenAI-compatible API.…
The model performs well on graduate-level reasoning as indicated by its GPQA Diamond score of 87.5. It handles complex question answering, code generation, mathematical problem solving, and multimodal reasoning. Its long output limit makes it suitable for generating entire reports, code files, or lengthy analyses in a single response.
Ideal use cases include: generating long-form content (articles, stories, documentation) up to 128k tokens; processing and answering questions about images or files; performing multi-step reasoning tasks such as debugging or data extraction; and any application where cost per token is a primary concern but high reasoning quality is still required.
While the mini variant scores 87.5 on GPQA Diamond, larger GPT-5.4 models likely achieve higher raw scores on the same benchmark. The mini model also has a fixed maximum output of 128k tokens; it cannot produce longer responses without iterative calls. It may also have lower performance on extremely niche or specialized domains where a larger model's broader knowledge base is advantageous.
For tasks that do not require high reasoning ability, such as simple text classification, translation, or basic summarization, a smaller or cheaper model (e.g., GPT-3.5 variants) may suffice. If the workload is purely text-only with short outputs, lower-cost models available on OrcaRouter could reduce expenses further. Evaluate whether the GPQA Diamond-level quality is necessary for your specific application.
GPQA Diamond is a benchmark consisting of graduate-level questions across physics, chemistry, biology, and other sciences. A score of 87.5 indicates that the model correctly answers 87.5% of the questions in that dataset. This places it in a strong reasoning tier, comparable to or exceeding some earlier large models. However, benchmark scores do not guarantee real-world performance and should be tested on domain-specific tasks.
Latency is not directly specified in the provided facts. As a mini variant, it is expected to be faster and more cost-efficient than full-size GPT-5.4, but actual speed depends on request size, output length, and OrcaRouter’s current load. Users should benchmark with their own payloads to estimate time per response. The model does not have a stated rate limit; that is managed by OrcaRouter.
The 87.5 on GPQA Diamond is the only benchmark figure provided. It suggests strong capability in graduate-level science reasoning. The model also supports multimodal input and long outputs, which are not measured by GPQA but are additional strengths for practical use. Its pricing is competitive for the quality level indicated by this single benchmark.
Benchmark scores may not cover all real-world scenarios. The GPQA Diamond test is multiple-choice and may not reflect open-ended generation quality. The model’s performance on creative writing, code generation, or safety-related tasks is not captured by this figure. Additionally, the mini variant may have reduced knowledge breadth compared to the full GPT-5.4. Always validate the model on your own data.
Pricing is $0.75 per million input tokens and $4.50 per million output tokens. These are the exact rates billed by OpenAI with zero markup from OrcaRouter. Input refers to the total token count of the user prompt (including system messages, images tokenized, and file text). Output is the token count of the model's response.
Compared to larger models like GPT-5.4 full, this mini variant is cheaper per token. However, it may require more conservative prompting or multiple retries for complex tasks, potentially increasing total token usage. For high-volume use, the lower per-token cost can lead to significant savings. But if output quality is critical and the task is very challenging, the full model might reduce total cost by requiring fewer calls.
The provided facts do not mention any caching or discount programs. OrcaRouter passes through OpenAI's pricing without markup. Any volume discounts or prompt caching must be offered by OpenAI directly; check OpenAI’s documentation for cached input rates if available. At this time, the flat rate applies per billing cycle based on total tokens consumed.
Optimize cost by keeping input prompts concise, using system messages to enforce topic limits, and avoiding unnecessary image or file inputs which can increase token count. For very long outputs, consider chunking content into multiple requests if the full 128k output is rarely needed. Monitor token usage via OrcaRouter’s logs and set max_tokens to the minimum acceptable length for each task.
Use the OpenAI-compatible endpoint at https://api.orcarouter.ai/v1 with the model id "openai/gpt-5.4-mini-2026-03-17". Authenticate with your OrcaRouter API key in the Authorization header. Send a chat completion request with the model field set to that exact string. Example: curl https://api.orcarouter.ai/v1/chat/completions -H "Authorization: Bearer $API_KEY" -d '{"model":"openai/gpt-5.4-mini-2026-03-17","messages":[{"role":"user","content":"Hello"}],"max_tokens":1000}'
The API supports standard OpenAI chat completion parameters: model (required), messages, max_tokens, temperature, top_p, frequency_penalty, presence_penalty, stop, stream, and others. For multimodal input, include content blocks with type "text" and "image_url" in user messages. Files can be added by providing their contents as text or embeddings. OrcaRouter translates these correctly to the underlying model.
If you are using a different OpenAI model via OrcaRouter, change the model string to "openai/gpt-5.4-mini-2026-03-17". Test on a sample of your prompts to ensure output quality meets expectations. Adjust max_tokens if you were using a smaller output limit; this model supports up to 128k. No other API changes are needed. Monitor pricing change from your previous model's rate.
Comparison benchmarks are not provided. gpt-4o is an earlier model from OpenAI with a different architecture. gpt-5.4-mini is a newer mini model. Based on GPQA Diamond scores (if gpt-4o scores less than 87.5), it may outperform gpt-4o on reasoning. However, gpt-4o may have advantages in speed or broader knowledge. Users should evaluate on their own tasks.
Among mini models, this one offers a large output limit (128k tokens) and multimodal input. Its GPQA Diamond score of 87.5 is a differentiator for reasoning quality. Other mini models from OpenAI or third parties may have lower reasoning ability or shorter output limits. Pricing at $0.75/$4.50 per million tokens is competitive for the quality level, but always compare total cost per completed task.
Open-source models may have lower or no API cost but require infrastructure setup and maintenance. This model offers 128k output limit and 87.5 GPQA Diamond performance, which may surpass many open-source models. However, open-source models can be fine-tuned at no per-token cost. The choice depends on whether you prefer pay-per-use convenience or self-hosted control.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key=os.environ["ORCAROUTER_API_KEY"],
)
response = client.chat.completions.create(
model="openai/gpt-5.4-mini-2026-03-17",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)include_reasoningmax_completion_tokensmax_tokensreasoningresponse_formatseedstreamstructured_outputstool_choicetools| Input / 1M tokens | $0.750 |
|---|---|
| Output / 1M tokens | $4.50 |
| Cache read / 1M | $0.075 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
What developers are saying this week
GET /api/public/models/openai/gpt-5.4-mini-2026-03-17Open @misc{orcarouter_gpt_5_4_mini_2026_03_17,
title = {openai/gpt-5.4-mini-2026-03-17 API},
author = {openai},
year = {n.d.},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/openai/gpt-5.4-mini-2026-03-17}
}openai. (n.d.). openai/gpt-5.4-mini-2026-03-17 API. OrcaRouter. https://www.orcarouter.ai/models/openai/gpt-5.4-mini-2026-03-17