OpenAI GPT-4o-mini preview with extended context for search-oriented text generation at low cost.
This model is a preview variant of GPT-4o-mini from OpenAI, optimized for search-oriented text generation. It accepts only text inputs and offers a 128,000-token context window, allowing it to…
The model excels at understanding and generating text based on large contexts (up to 128k tokens) and producing outputs up to 16,384 tokens. It retains the core strengths of GPT-4o-mini: fast inference, good instruction following, and cost efficiency. The search-preview tuning likely improves its ability to extract relevant snippets, compare information, and answer questions that require scanning lengthy passages. It can be used for summarization, question answering, information extraction, and multi-turn conversation as long as the input remains text. It does not support function calls or tool use by default, though you can incorporate external search results into your prompt.
Based on its design, the model is best for tasks where a user provides a long text block that contains the information needed to generate an answer. This includes scenarios like: analyzing a set of search results (provided as text), comparing product descriptions, extracting key facts from research papers, or summarizing meeting transcripts. The 128k context allows processing entire books or multi-document collections in one request. However, it may not be optimal for creative writing, code generation, or tasks that require multimodal understanding. For those, consider the standard GPT-4o-mini or GPT-4o models.
If your task is very simple—such as a short Q&A or classification—and does not require the large context window or search-specific tuning, a cheaper model like GPT-4o-mini (without preview) or even GPT-3.5 Turbo might suffice. The preview model's pricing is identical to GPT-4o-mini, so cost is not a differentiator. However, if latency is the priority, GPT-4o-mini is already one of the fastest models; the preview version should have similar speed. If you do not need the search focus, the standard GPT-4o-mini may be more broadly tested and stable.
The maximum output tokens per request is 16,384. The input context can be up to 128,000 tokens. These are generous limits that allow for long responses and very long contexts. However, because the model is text-only, all tokens must be text. There is no native support for structured data like JSON schema enforcement, though you can instruct the model to output JSON. The preview may have rate limits; when using OrcaRouter, those limits depend on your account tier and the backend provider's availability.
OpenAI has not published separate benchmark scores for this specific preview model. However, the base GPT-4o-mini is known to achieve strong results across many NLP benchmarks while being significantly faster and cheaper than GPT-4o. Users should expect similar general performance, with potentially better results on tasks that involve information retrieval or long-context reasoning due to the search tuning. Without official benchmarks, it is recommended to evaluate the model on your own test set to gauge its effectiveness for your domain.
GPT-4o-mini is one of the fastest models in the GPT-4 family, and this preview version should have comparable latency. Typical time-to-first-token is low, making it suitable for interactive applications. Exact latency depends on request size, output length, and current provider load. The 128k context may increase time-to-first-token compared to shorter contexts, but once streaming begins, tokens are produced quickly. OrcaRouter does not add significant overhead; the API calls are optimized for minimal latency.
The main strength is the combination of high speed, low cost, and the ability to handle very long contexts. For tasks where the answer is present in the provided text, this model can extract it efficiently without the higher cost of GPT-4o. It also inherits the strong instruction adherence of GPT-4o-mini. However, it may not match the nuanced reasoning, creativity, or factual accuracy of larger models like GPT-4o or GPT-4 Turbo. In complex reasoning tasks, the larger models often perform better, but at higher latency and cost.
As a preview, the model may have undiscovered bugs or inconsistent behavior. It has not been as widely tested as stable releases. The search tuning might cause unexpected outputs when the input does not contain enough information, potentially leading to hallucination. The model cannot use external tools or browse the web unless you provide the relevant content. If your task requires up-to-date web search results, you would need to inject them into the prompt manually. Additionally, it does not support images or audio, limiting its use in multimodal scenarios.
OrcaRouter passes through the provider pricing with zero markup. For openai/gpt-4o-mini-search-preview-2025-03-11, the cost is $0.15 per 1 million input tokens and $0.60 per 1 million output tokens. Input tokens include the entire prompt (system message, user messages, and any history). Output tokens are the generated text. There are no additional fees from OrcaRouter. This makes the model very affordable for large-volume usage, especially when processing long contexts.
No specific caching discounts are mentioned for this model on OrcaRouter. The pricing is per-token, pay-as-you-go. Some providers offer prompt caching that reduces cost for repeated prefixes, but it is not confirmed for this preview. Typically, OpenAI may offer caching for certain models in their API, but this preview may not be eligible. Check OpenAI's documentation for the latest caching policies. On OrcaRouter, you are billed the same rate regardless of usage patterns.
At $0.15 input / $0.60 output per million tokens, it is identical to standard GPT-4o-mini. That makes it significantly cheaper than GPT-4o ($2.50 input / $10 output) and GPT-4 Turbo ($10 input / $30 output). It is more expensive than older models like GPT-3.5 Turbo ($0.50/$1.50 per million?). However, the value comes from the large context and search tuning. If you need the full context, the cost advantage over larger models is substantial.
To use the model, set your API base URL to https://api.orcarouter.ai/v1, your API key (from OrcaRouter), and the model ID to exactly "openai/gpt-4o-mini-search-preview-2025-03-11". Standard OpenAI Python client example: client = OpenAI(api_key="your_orcarouter_key", base_url="https://api.orcarouter.ai/v1") response = client.chat.completions.create( model="openai/gpt-4o-mini-search-preview-2025-03-11", messages=[{"role": "user", "content": "Your prompt"}] ) All standard chat completions parameters are supported, including temperature, max_tokens, top_p, frequency_penalty, presence_penalty, and stop sequences. Streaming is available using stream=True.
The model supports the same parameters as the OpenAI Chat Completions API. Key parameters: temperature (default 1.0, range 0-2), top_p (default 1.0), max_tokens (default varies, can set up to 16384), stop (list of strings), frequency_penalty (default 0), presence_penalty (default 0), and logit_bias. You can also pass user_id for monitoring. There is no native support for response_format schema yet; if you need structured output, use prompt instructions. The model respects system messages for setting context.
Migration requires no code changes to the request structure beyond updating the model ID. In your API call, replace the model string with "openai/gpt-4o-mini-search-preview-2025-03-11". Ensure your base URL points to OrcaRouter (https://api.orcarouter.ai/v1) if you are not already using it. Because it is a preview, test thoroughly: responses may differ in style or accuracy. You can conditionally switch between models in your application by storing the model ID in configuration. Monitor performance and costs to ensure the preview meets your needs before full rollout.
Both models share the same base architecture and pricing. The standard GPT-4o-mini is a general-purpose model with no specific search tuning. The preview is designed to improve performance on tasks that involve extracting and synthesizing information from large text contexts, such as search result analysis. In practice, the preview may handle long prompts with many facts more accurately, while the standard version may be better for open-ended conversation, creative writing, or tasks where search behavior is not needed. If your application already works with standard GPT-4o-mini, testing the preview is low risk due to identical API.
GPT-4o is OpenAI's flagship multimodal model with higher reasoning ability, image understanding, and a larger context window (128k tokens as well). However, it is significantly more expensive ($2.50/$10 per million tokens) and slower. The search preview of GPT-4o-mini trades off some reasoning depth for speed and cost. For tasks that are purely text-based and do not require complex multi-step reasoning, the preview may be sufficient at a fraction of the cost. For tasks needing multimodal input or higher accuracy, GPT-4o remains superior.
Claude 3 Haiku (Anthropic) and Gemini 1.5 Flash (Google) are also fast, cheap models with large contexts. Haiku has 200k tokens context, Flash up to 1M. Pricing varies. The search preview of GPT-4o-mini is unique because it is a preview of a search-tuned variant from OpenAI. Without direct benchmark comparisons, it is hard to say which is best. All three are excellent for information retrieval tasks. The choice may depend on ecosystem compatibility, prompt style, and specific performance on your data. OrcaRouter provides unified access to all of them, allowing easy A/B testing.
Choose this model if your primary requirement is cost efficiency, fast response times, and you are working with text-only inputs that do not require advanced reasoning. It is ideal for high-volume applications where each request's context is large (e.g., processing long documents) but the output is relatively short. If you need the best possible accuracy, especially on ambiguous or complex queries, or if you need to process images, then GPT-4o is a better choice. The preview model is also suitable for prototyping and testing before committing to a higher-cost model.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key="$ORCAROUTER_API_KEY",
)
response = client.chat.completions.create(
model="openai/gpt-4o-mini-search-preview-2025-03-11",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)max_tokensresponse_formatstreamstructured_outputsweb_search_options| Input / 1M tokens | $0.150 |
| Output / 1M tokens | $0.600 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/openai/gpt-4o-mini-search-preview-2025-03-11Open @misc{orcarouter_gpt_4o_mini_search_preview_2025_03_11,
title = {openai/gpt-4o-mini-search-preview-2025-03-11 API},
author = {openai},
year = {n.d.},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/openai/gpt-4o-mini-search-preview-2025-03-11}
}openai. (n.d.). openai/gpt-4o-mini-search-preview-2025-03-11 API. OrcaRouter. https://www.orcarouter.ai/models/openai/gpt-4o-mini-search-preview-2025-03-11