Too much latency with computer-use-preview model

Hi Shah Saurabh,

Thanks for explaining clearly.

Computer-Use Preview model is unstable

·       It’s a preview model (not production-grade).

·       Microsoft itself warns: "performance, latency, reliability of preview models are not guaranteed."

·       These models are running on shared compute resources (less prioritization → more queuing delays).

Heavy model routing internally

·       Even with "small" tasks, internally OpenAI may be doing heavy work (especially for computer-use models handling function execution).

·       Some overhead comes from the tool use orchestration logic inside OpenAI, not just model inference.

Azure load congestion

·       April was smoother because fewer users were using computer-use preview.

·       As of May-April end, more usage → shared clusters saturated → higher queuing delays even before your token hits inference.

Model Overheads (Tool-Use Latency)

·       Tool use (like "open browser" or "write to Excel") isn't just a model generation — it triggers internal API calls or backend systems.

·       These secondary actions significantly increase the time beyond simple "prompt completion".

What you can do to improve it:

1. Move to Production-Ready Model (if possible)

·       Microsoft usually stabilizes preview models into production within 3–6 months.

·       Check if a GA (Generally Available) version of computer-use is planned soon.

·       If yes, migrate to GA → you'll get:

o   Better dedicated compute

o   Lower latency SLAs

o   Prioritized inference queues.

Split Large Tasks into Smaller Subtasks

·       Instead of asking it to "open browser, go to URL, read data, write into XLS" all at once,

·       Break into small model calls:

o   1st call → open browser and url

o   2nd call → scrape data

o   3rd call → write into file

·       Result: Each model call is smaller → faster completions.

·       Avoid tool-chaining too heavily inside one prompt.

Hope this helps, do let me know if you have any further queries.

Thank you! 

Hi Shah Saurabh,

Following up to see if the above suggestion was helpful. And, if you have any further query do let me know.

Thank you!

@Prashanth Veeragoni ,

Obviously, the provided information was really in detail but most of them are not helpful in current scenario (e.g. usage of GA model which is not available at all at this stage nor we have a clarity on its availability in near future) and some of the points (e.g. Split Large Tasks into Smaller Subtasks) are already being taken care.

So no additional clue from provided information.

I am wondering, if preview model can not meet most of the basic requirement then why it is charged higher by Azure (computer-use-preview is 20% costlier than gpt4o).

Let me know if anything additional can be done till the time GA will be released to use this model for different PoC use cases which meets basic requirements (e.g. consistency, acceptable latency, clarity on usage, etc...).

It would be great if someone share formula to calculate usage/image tokens for computer-use-preview model, the way its given for vision GPT-4 series at https://learn.microsoft.com/en-us/azure/ai-services/openai/overview#image-tokens-gpt-4-turbo-with-vision.

Hi Shah Saurabh,

This is a fair concern, especially when preview models like computer-use-preview exhibit latency issues despite higher cost.

Next Action Plan

 1.Clarify the Nature of Preview Models

Preview models in Azure OpenAI (like computer-use-preview) are:

·       Not production SLAs: They often run on shared infrastructure without throughput guarantees.

·       Priced higher because they're experimental and early-access (potentially with richer capabilities) but not optimized for latency or performance yet.

·       Subject to throttling, queuing, or instability, particularly under high demand — and no capacity guarantees like those available with GA models or PTUs.

Why latency increases:

·       As usage grows (especially post-April), shared preview models can experience queue backlogs, resulting in 3+ minute responses.

·       Azure doesn’t scale preview capacity dynamically in the same way as GA models with PTU.

What Can Be Done:

Here are realistic actions that can be taken right now within the preview limitations:

1.Switch to gpt-4o or GPT-4 Turbo (if possible)

·       Even for vision and interaction tasks, gpt-4o (GA) is faster and cheaper than computer-use-preview.

·       But You're correct: there is no GA version or clear roadmap for computer-use-preview as of now.

·       If computer-use-preview was selected solely for browser-based automation, see if similar tasks can be broken down into reasoning + tool use patterns in gpt-4o.

2.Use Function Calling or Tool Use Frameworks

Instead of expecting the model to drive a full automation sequence in one pass, consider:

·       Splitting the automation workflow into modular tasks triggered via LangChain agents or OpenAI function calling (via FastAPI / Logic Apps / Power Automate).

·       E.g., model only decides the next UI step, and a background function executes it — reducing context/token load and speeding up responses.

3.Log Token Usage & Latency

Use Azure’s built-in monitoring (Application Insights or Log Analytics) to:

·       Measure token usage and response latency across prompts.

·       Identify whether long prompts, complex images, or large input payloads are causing delays.

Token Calculation Formula for computer-use-preview

Unfortunately, Azure has not yet published official token calculation details for computer-use-preview. However, here's how you can estimate:

Estimate Tokens for computer-use-preview (Unofficial Heuristic):

Text Tokens: Use tiktoken (GPT-4 tokenizer) → ~4 characters = 1 token.
Image Tokens: Based on image resolution + complexity (no exact formula yet).

In absence of official formula:

·       Use GPT-4 Vision estimate: Azure doc on Vision tokens.

o   Image tokens ≈ width * height / (128 * 128)

o   This gives approximate number of 128×128 patches (like CLIP).

 Workaround: You can manually test via Azure billing logs:

1.      Run a single image/text prompt.

2.      Check Prompt Tokens and Completion Tokens from the billing dashboard or usage logs.

3.      Back-calculate average token usage per image size.

Hope this helps.

Thank you!

Hello Shah Saurabh,

Welcome to the Microsoft Q&A and thank you for posting your questions here.

I understand that you are having too much latency with computer-use-preview model, in review the thread responses, also I discovered misconception GPT-4o viability and cost logic, new issue about pricing fairness, token tracking, and lack of roadmap.

To resolve all of these, I provide you a support escalation, hybrid execution, telemetry, and some cost workaround with references for more detailed information and steps:

1. Use Log Analytics to measure latency spikes across time of day. Try scheduling calls during low-load windows and avoid peak hours using telemetry. - https://learn.microsoft.com/en-us/azure/azure-monitor/logs/log-analytics-overview
2. Use a hybrid architecture (agent-based execution) by offload tool execution to local agents or Power Automate Desktop scripts triggered by LLM commands. This mimics computer-use-preview tasks but reduces token-based latency. - https://learn.microsoft.com/en-us/power-automate/desktop-flows/introduction
3. Use Azure Batch / Durable Functions for repeated PoCs, queue tasks using Durable Functions to decouple LLM latency from task execution. This also allows retries and better failure handling. - https://learn.microsoft.com/en-us/azure/azure-functions/durable/durable-functions-overview
4. Token Estimation Workaround Since no token formula is published for computer-use-preview, use: Text tokens: ~4 characters = 1 token Image tokens (heuristic): (width * height) / (128 * 128) You can monitor token cost per operation using Azure OpenAI Usage Logs - https://learn.microsoft.com/en-us/azure/ai-services/openai/monitor-usage
5. If cost vs. performance is not justified, I will advise you to raise a support ticket for billing inquiry and refund escalation. - https://learn.microsoft.com/en-us/azure/cost-management-billing/manage/troubleshoot-azure-subscription-billing provide usage data and argue that GA models (like GPT-4o) perform better at lower cost, highlighting discrepancy.
6. You can also request priority allocation or dedicated capacity via Azure Support for preview model. Mention latency regression since April. You can even contact Microsoft Priority Customer Support - (PCS) with all of your fact details.

I hope this is helpful! Do not hesitate to let me know if you have any other questions or clarifications.

Please don't forget to close up the thread here by upvoting and accept it as an answer if it is helpful.