[Windows][GPT-5.6 Sol] Chrome/browser tasks are extremely laggy and token-heavy on both Medium and xhigh

What version of the Codex App are you using (From “About Codex” dialog)?

26.810.52044

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

GPT-5.6 Sol becomes extremely slow and inefficient in Codex when a task involves Chrome/browser tooling.

I have reproduced the problem with both Medium and xhigh reasoning effort. This does not appear to be simply a case of xhigh taking longer because it reasons more: browser-driven sessions feel laggy on both settings, consume a disproportionate amount of tokens/usage, and make very little progress relative to the amount of work and context consumed.

Typical symptoms:

• Long pauses around browser/Chrome interaction.
• Browser-driven tasks take much longer than the underlying task complexity would suggest.
• Token/usage consumption rises very quickly once Chrome/browser work is involved.
• The model performs many browser/tool interactions without proportional progress toward the requested result.
• Medium does not make the workflow feel acceptably responsive; xhigh is even more expensive and slow.
• The combination makes ordinary browser-assisted development or research difficult to complete productively.

The important point is that Chrome/browser tooling itself is available and can work. This report is about performance and model/tool-use efficiency, not the separate failure mode where the Browser/Chrome plugin is unavailable.

As a user-side comparison, Claude's Chrome/browser workflow feels dramatically more responsive and effective for similar browser-assisted tasks. I am not presenting that as a controlled benchmark, but the usability difference is large enough to make Codex's GPT-5.6 Sol browser workflow feel abnormal rather than merely slower by design.

What steps can reproduce the bug?

1. Open Codex App on Windows.
2. Start a fresh task with GPT-5.6 Sol, Medium reasoning effort.
3. Give it a normal task that requires repeated Chrome/browser interaction, such as inspecting a web application, navigating pages, checking UI state, or using browser tooling as part of a development task.
4. Observe the latency between useful actions, the number of browser/tool interactions, and the amount of progress made.
5. Monitor token/context/usage consumption during the task.
6. Repeat the same type of task with GPT-5.6 Sol, xhigh.
7. Observe that both reasoning settings remain noticeably laggy and token-heavy, with xhigh generally making the cost/latency problem more pronounced.

I have encountered this repeatedly rather than as a single isolated browser action.

What is the expected behavior?

Browser-assisted tasks should remain responsive and their token/usage consumption should be reasonably proportional to the work completed.

In particular, GPT-5.6 Sol should:

• avoid unnecessary repeated browser observations or tool calls when page state has not materially changed;
• avoid carrying or regenerating excessive browser/tool state in active context;
• choose browser actions that advance the task rather than repeatedly re-inspecting the same state;
• keep Medium meaningfully faster and cheaper than xhigh for the same browser workload;
• make enough progress per browser/tool interaction that normal Chrome-assisted tasks remain practical;
• expose enough diagnostics to identify whether the latency and usage are coming from model reasoning, browser tool calls, screenshots/page state, tool-output context, retries, or compaction.

Additional information

This appears related in theme, but not identical, to:

• #33592 — Browser and Chrome plugins unavailable with gpt-5.6-sol. In my case Chrome/browser tooling is available; the problem is severe latency and usage inefficiency when it is used.
• #37606 — rapid context growth and repeated tool calls in fresh GPT-5.6 Sol Codex Desktop sessions. This report narrows the workload specifically to Chrome/browser-assisted tasks and reproduces the poor experience on both Medium and xhigh.
• #39059 — separate report about GPT-5.6 Codex expanding bounded codebase work into self-reinforcing verification/governance machinery. The browser problem is sufficiently distinct that I am reporting it separately.

I would especially like this investigated as a model/tool-use efficiency problem, not only as a Chrome UI performance issue. The browser can be functional while the overall agent loop is still too slow and token-expensive to be useful.