I have searched for any existing and/or related issues.
I have searched for any existing and/or related discussions.
I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!).
I am using the latest version of Open WebUI.
Installation Method
Docker
Open WebUI Version
v0.9.5
Ollama Version (if applicable)
No response
Operating System
Garuda Linux
Browser (if applicable)
Firedragon
Confirmation
I have read and followed all instructions in README.md.
I am using the latest version of both Open WebUI and Ollama.
I have included the browser console logs.
I have included the Docker container logs.
I have provided every relevant configuration, setting, and environment variable used in my setup.
I have clearly listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
I have documented step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation. My steps:
Start with the initial platform/version/OS and dependencies used,
Specify exact install/launch/configure commands,
List URLs visited, user input (incl. example values/emails/passwords if needed),
Describe all options and toggles enabled or changed,
Include any files or environmental changes,
Identify the expected and actual result at each stage,
Ensure any reasonably skilled user can follow and hit the same issue.
Expected Behavior
With Grok model, prompt should use the prompt cache to reduce token costs. In response Info, cached_tokens should be high for consecutive requests within 5 minutes.
Actual Behavior
cache_tokens stays 64 or 128 or other such values, billing me the message history at full cost on every request. Sometimes it caches, often it doesn't. Leaving temperature and top_p to default values seems to help; it's worse with custom values.
I'm pretty sure server-affinity header is the issue, since it works but inconsistent. It depends what servers the request gets routed to.
Originally created by @mysteryx93 on GitHub (May 16, 2026).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/24804
### Check Existing Issues
- [x] I have searched for any existing and/or related issues.
- [x] I have searched for any existing and/or related discussions.
- [x] I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!).
- [x] I am using the latest version of Open WebUI.
### Installation Method
Docker
### Open WebUI Version
v0.9.5
### Ollama Version (if applicable)
_No response_
### Operating System
Garuda Linux
### Browser (if applicable)
Firedragon
### Confirmation
- [x] I have read and followed all instructions in `README.md`.
- [x] I am using the latest version of **both** Open WebUI and Ollama.
- [x] I have included the browser console logs.
- [x] I have included the Docker container logs.
- [x] I have **provided every relevant configuration, setting, and environment variable used in my setup.**
- [x] I have clearly **listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup** (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
- [x] I have documented **step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation**. My steps:
- Start with the initial platform/version/OS and dependencies used,
- Specify exact install/launch/configure commands,
- List URLs visited, user input (incl. example values/emails/passwords if needed),
- Describe all options and toggles enabled or changed,
- Include any files or environmental changes,
- Identify the expected and actual result at each stage,
- Ensure any reasonably skilled user can follow and hit the same issue.
### Expected Behavior
With Grok model, prompt should use the prompt cache to reduce token costs. In response Info, `cached_tokens` should be high for consecutive requests within 5 minutes.
### Actual Behavior
`cache_tokens` stays 64 or 128 or other such values, billing me the message history at full cost on every request. Sometimes it caches, often it doesn't. Leaving temperature and top_p to default values seems to help; it's worse with custom values.
### Steps to Reproduce
1. Add Connection to xAI (https://api.x.ai/v1)
2. Use model `grok-4.3`
3. Set `temperature=.57` and `top_p=0.84`
4. Enable `Usage` display on model
5. Send multiple prompts and check `cached_tokens`
### Logs & Screenshots
I think the issue has been diagnosed here:
Grok prompt caching likely misses xAI server-affinity header
https://github.com/NousResearch/hermes-agent/issues/22705
### Additional Information
I'm pretty sure server-affinity header is the issue, since it works but inconsistent. It depends what servers the request gets routed to.
GiteaMirror
added the bug label 2026-05-18 06:12:08 -05:00
@owui-terminator[bot] commented on GitHub (May 16, 2026):
🔍Related Issues Found
I found some existing issues that might be related. Please check if any of these are duplicates or contain helpful solutions:
🟣#22791issue: Open WebUI API: incomplete / dropped output with latest Grok models via /api/v1/chat/completions, while older Grok and OpenAI models work This is the closest Grok/xAI-related issue in Open WebUI. It also concerns latest Grok models behaving inconsistently through Open WebUI’s chat completions path, so it may share the same provider-specific request handling or routing behavior. by Patrick-0815 · bug
🟣#12058issue: Cached tokens out of nowhere This issue is about unexpected cached token behavior and inflated token usage, which overlaps with the core symptom here even though it was reported with a different backend. It’s relevant as a prior caching-related bug in Open WebUI. by davidvpe · bug
💡 If your issue is a duplicate, please close it and add any additional details to the existing issue instead.
This comment was generated automatically. React with 👍 if helpful, 👎 if not.
<!-- gh-comment-id:4465439653 -->
@owui-terminator[bot] commented on GitHub (May 16, 2026):
<!-- terminator-bot:related-issues-reply -->
🔍 **Related Issues Found**
I found some existing issues that might be related. Please check if any of these are duplicates or contain helpful solutions:
1. 🟣 [#22791](https://github.com/open-webui/open-webui/issues/22791) **issue: Open WebUI API: incomplete / dropped output with latest Grok models via /api/v1/chat/completions, while older Grok and OpenAI models work**
*This is the closest Grok/xAI-related issue in Open WebUI. It also concerns latest Grok models behaving inconsistently through Open WebUI’s chat completions path, so it may share the same provider-specific request handling or routing behavior.*
*by Patrick-0815 · `bug`*
2. 🟣 [#12058](https://github.com/open-webui/open-webui/issues/12058) **issue: Cached tokens out of nowhere**
*This issue is about unexpected cached token behavior and inflated token usage, which overlaps with the core symptom here even though it was reported with a different backend. It’s relevant as a prior caching-related bug in Open WebUI.*
*by davidvpe · `bug`*
---
💡 If your issue is a duplicate, please close it and add any additional details to the existing issue instead.
*This comment was generated automatically.* React with 👍 if helpful, 👎 if not.
<!-- gh-comment-id:4465493613 -->
@mysteryx93 commented on GitHub (May 16, 2026):
Here is a diagnostic and fix that was implemented in another project
https://github.com/NousResearch/hermes-agent/pull/22809
Even on providers where caching works, sometimes you still get a cache miss due to location differences.
Open WebUI is fully chat completions standard conform - if you get cache misses, then that's an upstream issue of either the provider or the provider demanding special headers that open webui doesn't send (i.e. you forgot to configure).
Also check if grok even does caching if you do off-default temperature and top p values. Many providers nowadays will refuse to do reasoning or answer at all with off-default values
<!-- gh-comment-id:4466366203 -->
@Classic298 commented on GitHub (May 16, 2026):
very likely provider or config issue.
Even on providers where caching works, sometimes you still get a cache miss due to location differences.
Open WebUI is fully chat completions standard conform - if you get cache misses, then that's an upstream issue of either the provider or the provider demanding special headers that open webui doesn't send (i.e. you forgot to configure).
Also check if grok even does caching if you do off-default temperature and top p values. Many providers nowadays will refuse to do reasoning or answer at all with off-default values
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Originally created by @mysteryx93 on GitHub (May 16, 2026).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/24804
Check Existing Issues
Installation Method
Docker
Open WebUI Version
v0.9.5
Ollama Version (if applicable)
No response
Operating System
Garuda Linux
Browser (if applicable)
Firedragon
Confirmation
README.md.Expected Behavior
With Grok model, prompt should use the prompt cache to reduce token costs. In response Info,
cached_tokensshould be high for consecutive requests within 5 minutes.Actual Behavior
cache_tokensstays 64 or 128 or other such values, billing me the message history at full cost on every request. Sometimes it caches, often it doesn't. Leaving temperature and top_p to default values seems to help; it's worse with custom values.Steps to Reproduce
grok-4.3temperature=.57andtop_p=0.84Usagedisplay on modelcached_tokensLogs & Screenshots
I think the issue has been diagnosed here:
Grok prompt caching likely misses xAI server-affinity header
https://github.com/NousResearch/hermes-agent/issues/22705
Additional Information
I'm pretty sure server-affinity header is the issue, since it works but inconsistent. It depends what servers the request gets routed to.
@owui-terminator[bot] commented on GitHub (May 16, 2026):
🔍 Related Issues Found
I found some existing issues that might be related. Please check if any of these are duplicates or contain helpful solutions:
🟣 #22791 issue: Open WebUI API: incomplete / dropped output with latest Grok models via /api/v1/chat/completions, while older Grok and OpenAI models work
This is the closest Grok/xAI-related issue in Open WebUI. It also concerns latest Grok models behaving inconsistently through Open WebUI’s chat completions path, so it may share the same provider-specific request handling or routing behavior.
by Patrick-0815 ·
bug🟣 #12058 issue: Cached tokens out of nowhere
This issue is about unexpected cached token behavior and inflated token usage, which overlaps with the core symptom here even though it was reported with a different backend. It’s relevant as a prior caching-related bug in Open WebUI.
by davidvpe ·
bug💡 If your issue is a duplicate, please close it and add any additional details to the existing issue instead.
This comment was generated automatically. React with 👍 if helpful, 👎 if not.
@mysteryx93 commented on GitHub (May 16, 2026):
Here is a diagnostic and fix that was implemented in another project
https://github.com/NousResearch/hermes-agent/pull/22809
@Classic298 commented on GitHub (May 16, 2026):
very likely provider or config issue.
Even on providers where caching works, sometimes you still get a cache miss due to location differences.
Open WebUI is fully chat completions standard conform - if you get cache misses, then that's an upstream issue of either the provider or the provider demanding special headers that open webui doesn't send (i.e. you forgot to configure).
Also check if grok even does caching if you do off-default temperature and top p values. Many providers nowadays will refuse to do reasoning or answer at all with off-default values