Google's multimodal preview model for robotics, supporting text, image, video, and audio inputs with up to 65,536 output tokens.
google/gemini-robotics-er-1.6-preview is a preview model from Google's Gemini family optimized for robotics and embodied reasoning. It is a multimodal model that can process text, images, video, and…
The model is designed for robotics-related multimodal reasoning. It can interpret images and video to identify objects, environments, and human actions. It can process audio commands and extract information from spoken language. Combining these modalities, it can generate instructions for robotic manipulation, navigation, or interaction. For example, given a video of a kitchen and a spoken request to 'fetch the apple from the counter,' the model can output a sequence of actions. The large output context enables it to produce detailed plans or code for robot controllers. Note that the model's performance on these tasks is not yet benchmarked publicly for this preview version.
If your application does not require multimodal inputs (e.g., you only use text), a smaller text-only model would be more cost-effective. The robotics-er model is priced relatively high for input tokens ($1.00 per million) compared to many general-purpose models. For simple question answering or text generation without visual or audio context, consider using a lighter model available through OrcaRouter, such as Gemini 1.5 Flash or a smaller open-source model. Also, if the maximum output of 65,536 tokens is not necessary, a model with a smaller output context may offer lower latency and cost. Evaluate whether the multimodal capabilities justify the price for your use case.
The 65,536-token output limit is particularly valuable for generating long-form content such as robotic motion sequences (e.g., Python code using ROS or control libraries), detailed environment descriptions for 3D reconstruction, or multi-step task plans that require extensive reasoning. It can also be used for in-context learning, where the model is given many examples within a single prompt and expected to produce a comprehensive output. For robotics research, this enables generating long-horizon plans that cover many possible contingencies. However, note that longer outputs increase cost and latency, so consider whether a shorter response would suffice.
Audio and video inputs are processed as part of the multimodal prompt. When you send a video, the model likely treats it as a sequence of frames or uses a video understanding encoder (exact methods are Google's internal). Audio can be included as a URL to an audio file. The model can transcribe or understand spoken commands and generate text responses accordingly. The quality of audio and video processing depends on the sampling and format supported by Google's Gemini API; consult the provider documentation for supported file types and maximum durations. OrcaRouter simply relays the input in the OpenAI-compatible format.
As a preview release, Google has not published specific benchmark scores for the gemini-robotics-er-1.6-preview model. General Gemini 1.6 models have shown strong performance on multimodal tasks, but this robotics-specific variant may have different strengths. Users should evaluate the model's performance on their own domain-specific tasks before relying on it. OrcaRouter does not provide benchmark numbers beyond what the provider publishes. For real-world reliability, conduct A/B testing with your own data and compare latency, accuracy, and cost against other models.
Latency for google/gemini-robotics-er-1.6-preview is not specified by the provider. In general, multimodal models that process video and audio tend to have higher latency than text-only models due to the encoding overhead. Additionally, the large output context can increase response time, especially for long generations. Actual latency depends on request size, complexity, and server load on both Google's infrastructure and OrcaRouter's proxy. For best performance, minimize unnecessary input data and set appropriate max_tokens. OrcaRouter's API returns standard timing headers; monitoring those will give you empirical data for your use case.
The model's primary strength is its support for multiple input modalities (text, image, video, audio) combined with an exceptionally large output context of up to 65,536 tokens. This makes it well-suited for complex robotics tasks that require understanding both visual and auditory information and generating extensive, detail-rich plans. The preview nature suggests it is among the first Gemini models fine-tuned for embodied reasoning. Another strength is the straightforward access via OrcaRouter's OpenAI-compatible API, which lowers the integration barrier for teams familiar with OpenAI's interface.
As a preview model, its stability and performance may not match production-ready models. The lack of published benchmarks means its accuracy on standard robotics tasks is unknown. It may have higher failure rates or unexpected behaviors compared to established models. The model does not generate images or audio outputs, only text. Pricing per output token is relatively high ($5.00 per million) compared to many models. Additionally, the large output context may encourage verbose responses that are not always necessary. Users should prototype extensively before committing to this model for any serious application.
Billing for google/gemini-robotics-er-1.6-preview on OrcaRouter is straightforward: $1.00 per 1 million input tokens and $5.00 per 1 million output tokens. Both input and output tokens are counted as per Google's tokenization, which may differ from other models. OrcaRouter charges the exact provider rate with zero markup. There are no additional fees for the API proxy service. The cost is based on the total number of tokens processed in the request and response, including multimodal input tokens (e.g., image patches may be counted as tokens). Always check the usage returned in the API response to track costs.
The output token cost ($5.00 per million) is significantly higher than the input cost ($1.00 per million). This means making the model generate long responses increases cost far more than providing longer input. For use cases where output length can be kept short (e.g., a one-sentence command), the cost may be acceptable. However, if you leverage the full 65,536 output context, a single request could cost up to $0.33 just for output (65k tokens at $5/M). Compare this to models with lower output pricing or smaller context windows. Also, multimodal inputs can be expensive if they require many tokens (e.g., high-resolution images or long videos). Consider token compression or using lower resolution when possible.
OrcaRouter currently applies the provider rate with zero markup. There is no built-in caching that reduces cost for repeated prompt prefixes, as it is a passthrough proxy. However, you can implement caching at the application level: if you reuse identical input contexts (e.g., static system prompts or reference images), store the model's response and reuse it, avoiding repeated API calls. Discounts or volume pricing may be available through OrcaRouter's enterprise plans; contact the OrcaRouter team for details. Standard pricing applies to all usage unless a special agreement is in place.
Use the standard OpenAI chat completions endpoint. Set the 'model' parameter to 'google/gemini-robotics-er-1.6-preview'. The base URL is https://api.orcarouter.ai/v1. Include your OrcaRouter API key in the Authorization header. For text-only messages, the request body is identical to an OpenAI ChatCompletion request: { "model": "google/gemini-robotics-er-1.6-preview", "messages": [{"role": "user", "content": "Your message"}], "max_tokens": 2000 }. For multimodal inputs, structure the content as an array with type and data fields as per the provider schema. OrcaRouter translates the request to Google's format internally.
The API supports the same parameters as the OpenAI /v1/chat/completions endpoint: model, messages, max_tokens (up to 65536), temperature (0 to 2, default 1), top_p (0 to 1, default 1), frequency_penalty, presence_penalty, stop sequences, stream (boolean), logprobs, and user. Not all parameters may be effective on Google's backend; for instance, frequency_penalty and presence_penalty may have limited effect. For streaming, set 'stream: true' to receive token-by-token responses. The response format mirrors OpenAI's, including choices, usage (prompt_tokens, completion_tokens, total_tokens), and id. Check OrcaRouter documentation for any model-specific parameter overrides.
Since OrcaRouter provides an OpenAI-compatible API, migration is usually a matter of changing the base URL and API key. Replace 'https://api.openai.com/v1' with 'https://api.orcarouter.ai/v1' and set the model to 'google/gemini-robotics-er-1.6-preview'. If your app uses the OpenAI Python client, update the base_url and api_key. For multimodal inputs, note that the OpenAI format for images uses 'image_url', but Google's format may require different field names (like 'video_url'). OrcaRouter's API accepts a superset of OpenAI's multimodal schema; refer to their documentation for the exact structure. Test with a simple request before migrating complex logic.
Gemini 1.5 Pro is a general-purpose multimodal model with a strong balance of performance and cost. The robotics-er preview model is specialized for robotics and embodied reasoning, as suggested by its name. While Gemini 1.5 Pro supports up to 8,192 output tokens typically, this model offers up to 65,536 output tokens. Pricing differs: Gemini 1.5 Pro costs $1.25 per million input tokens and $5.00 per million output tokens (approximately), so this model is slightly cheaper on input but same on output. However, the preview model may have less general knowledge and more focus on physical world understanding. Benchmark comparisons are not available; users should test on their own tasks.
GPT-4o is a multimodal model from OpenAI with support for text, images, and audio (non-video) and a 16,384 token output limit. The robotics-er model has a much larger output context (65,536) and explicitly supports video input, which GPT-4o does not natively. Pricing differences: GPT-4o charges $2.50 per million input and $10.00 per million output tokens for standard usage, making the robotics model cheaper per token. However, GPT-4o is a mature production model with extensive benchmarks, while this model is a preview. For robotics-specific tasks, the Google model may be designed for better performance, but without public benchmarks, this is speculative.
Llama 3 models are text-only (or soon multimodal) but available for self-hosting, which can be cheaper at scale. The robotics-er model offers native multimodal support (including video and audio) out of the box, which open-source alternatives may not provide without additional components. The per-token pricing of the preview model can be expensive for high-volume use, whereas an open-source model run on your own hardware has predictable infrastructure costs. However, the Google model benefits from cloud-scale infrastructure and updates. For prototyping, the preview model's ease of use (via API) may outweigh the cost. Evaluate your total cost of ownership including development time.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1https://api.orcarouter.aiimport 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="google/gemini-robotics-er-1.6-preview",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)max_tokensresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_pinclude_reasoningreasoning| Input / 1M tokens | $1.00 |
| Output / 1M tokens | $5.00 |
| 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/google/gemini-robotics-er-1.6-previewOpen @misc{orcarouter_gemini_robotics_er_1_6_preview,
title = {google/gemini-robotics-er-1.6-preview API},
author = {google},
year = {n.d.},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/google/gemini-robotics-er-1.6-preview}
}google. (n.d.). google/gemini-robotics-er-1.6-preview API. OrcaRouter. https://www.orcarouter.ai/models/google/gemini-robotics-er-1.6-preview