Inconsistent Presentation of Model Capabilities in Agent Mode on OpenRouter Compared to Official Website #4144

Closed
opened 2026-02-21 18:04:56 -05:00 by yindo · 2 comments
Owner

Originally created by @XiaoLey on GitHub (Jun 17, 2024).

Self Checks

  • This is only for bug report, if you would like to ask a question, please head to Discussions.
  • I have searched for existing issues search for existing issues, including closed ones.
  • I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
  • Please do not modify this template :) and fill in all the required fields.

Dify version

0.6.11

Cloud or Self Hosted

Self Hosted (Docker)

Steps to reproduce

I have observed that the presentation of the model's capabilities in Agent mode on OpenRouter is significantly different from the presentation of the same model provided on the official website. Specifically, OpenAI's GPT-4o can recognize images and correctly call Agent tools to provide meaningful results, whereas OpenRouter's GPT-4o fails to do so.

  1. Use GPT-4o in Agent mode on OpenRouter.
  2. Attempt to recognize images or call Agent tools.
  3. Compare its responses with those from the official website for similar tasks.
  4. Noticeable discrepancies in capability presentation can be observed.

✔️ Expected Behavior

The model should present its capabilities consistently across different platforms, ensuring that features like image recognition and tool usage work similarly in both OpenRouter and the official website.

Actual Behavior

Snipaste_2024-06-18_11-23-57

Snipaste_2024-06-18_11-26-06

Snipaste_2024-06-18_11-27-04

Originally created by @XiaoLey on GitHub (Jun 17, 2024). ### Self Checks - [X] This is only for bug report, if you would like to ask a question, please head to [Discussions](https://github.com/langgenius/dify/discussions/categories/general). - [X] I have searched for existing issues [search for existing issues](https://github.com/langgenius/dify/issues), including closed ones. - [X] I confirm that I am using English to submit this report (我已阅读并同意 [Language Policy](https://github.com/langgenius/dify/issues/1542)). - [X] Please do not modify this template :) and fill in all the required fields. ### Dify version 0.6.11 ### Cloud or Self Hosted Self Hosted (Docker) ### Steps to reproduce I have observed that the presentation of the model's capabilities in Agent mode on OpenRouter is significantly different from the presentation of the same model provided on the official website. Specifically, OpenAI's GPT-4o can recognize images and correctly call Agent tools to provide meaningful results, whereas OpenRouter's GPT-4o fails to do so. 1. Use GPT-4o in Agent mode on OpenRouter. 2. Attempt to recognize images or call Agent tools. 3. Compare its responses with those from the official website for similar tasks. 4. Noticeable discrepancies in capability presentation can be observed. ### ✔️ Expected Behavior The model should present its capabilities consistently across different platforms, ensuring that features like image recognition and tool usage work similarly in both OpenRouter and the official website. ### ❌ Actual Behavior ![Snipaste_2024-06-18_11-23-57](https://github.com/langgenius/dify/assets/46273173/3d1227b0-a46f-456c-bf03-6d0c692a15ba) ------ ![Snipaste_2024-06-18_11-26-06](https://github.com/langgenius/dify/assets/46273173/8f612813-84c1-45d9-8df8-68dc0ec8295c) ------ ![Snipaste_2024-06-18_11-27-04](https://github.com/langgenius/dify/assets/46273173/4a8f33c8-b684-4c77-9cb8-970c48e445ff)
yindo added the 🤖 feat:agent label 2026-02-21 18:04:56 -05:00
yindo closed this issue 2026-02-21 18:04:56 -05:00
Author
Owner

@dosubot[bot] commented on GitHub (Jun 17, 2024):

To resolve the inconsistency in the presentation of GPT-4o's capabilities between OpenRouter and the official website, specifically regarding image recognition and tool usage, you can refer to the features listed in the provided YAML configuration. The features include:

  • multi-tool-call
  • agent-thought
  • stream-tool-call
  • vision

The inclusion of vision indicates that GPT-4o supports image recognition. Additionally, the features multi-tool-call, agent-thought, and stream-tool-call suggest capabilities related to tool usage [1][2].

Additionally, the OpenRouter configuration schema includes a specific field for vision support, which can be set to either 'support' or 'no_support'. This further confirms that GPT-4o can be configured to support image recognition [3].

This confirms that GPT-4o has both image recognition and tool usage capabilities.

To continue talking to Dosu, mention @dosu.

@dosubot[bot] commented on GitHub (Jun 17, 2024): <!-- Answer --> To resolve the inconsistency in the presentation of GPT-4o's capabilities between OpenRouter and the official website, specifically regarding image recognition and tool usage, you can refer to the features listed in the provided YAML configuration. The features include: - `multi-tool-call` - `agent-thought` - `stream-tool-call` - `vision` The inclusion of `vision` indicates that GPT-4o supports image recognition. Additionally, the features `multi-tool-call`, `agent-thought`, and `stream-tool-call` suggest capabilities related to tool usage <sup>[[1]](https://github.com/langgenius/dify/blob/main/api/core/model_runtime/model_providers/openai/llm/gpt-4o-2024-05-13.yaml)</sup><sup>[[2]](https://github.com/langgenius/dify/blob/main/api/core/model_runtime/model_providers/openai/llm/gpt-4o.yaml)</sup>. Additionally, the OpenRouter configuration schema includes a specific field for vision support, which can be set to either 'support' or 'no_support'. This further confirms that GPT-4o can be configured to support image recognition <sup>[[3]](https://github.com/langgenius/dify/blob/main/api/core/model_runtime/model_providers/openrouter/openrouter.yaml)</sup>. This confirms that GPT-4o has both image recognition and tool usage capabilities. <!-- Dosu Comment Footer --> *To continue talking to [Dosu](https://dosu.dev), mention @dosu.*
Author
Owner

@XiaoLey commented on GitHub (Jun 18, 2024):

I spent quite a bit of time tracking these issues. So far, I've only managed to fix one issue (which I'll submit shortly): resolving the problem where cot_agent_runner was not correctly analyzing user-uploaded images.

However, I have also discovered another issue: the openai_api_compatible model is not fully compatible with the Agent. This is evident in problems such as "responses containing formatting artifacts" and "memory function incorrectly recalling assistant's replies."

Snipaste_2024-06-18_17-29-14

@takatost

@XiaoLey commented on GitHub (Jun 18, 2024): I spent quite a bit of time tracking these issues. So far, I've only managed to fix one issue (which I'll submit shortly): resolving the problem where `cot_agent_runner` was not correctly analyzing user-uploaded images. However, I have also discovered another issue: the `openai_api_compatible` model is not fully compatible with the Agent. This is evident in problems such as "responses containing formatting artifacts" and "memory function incorrectly recalling assistant's replies." ![Snipaste_2024-06-18_17-29-14](https://github.com/langgenius/dify/assets/46273173/e4645f9b-6548-4fed-9179-7275f06c5080) @takatost
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: langgenius/dify#4144