Windows desktop / GPT-6.1 Sol: slow responses (~22.5 effective tokens/s) and ~179 s compaction #52014

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

Exact desktop build not captured. This observation was made in the native Windows Codex desktop app on October 8, 2026.

What subscription do you have?

Not included in this public report.

What platform is your computer?

Windows x64, OS build 26200.

What issue are you seeing?

The desktop app remains very slow during an ongoing conversation. I measured 12 consecutive completed, non-compaction model responses during approximately 07:36–07:47 UTC on October 8, 2026.

Verified request settings:

• Model: gpt-6.1-sol
• Reasoning effort: xhigh
• Service tier: default

The responses produced 8,635 output tokens over 384.515 seconds, or 22.46 effective output tokens/s. The median per-response rate was 25.89 tokens/s. Two responses took approximately 114 seconds each for about 2,300 output tokens. A separate automatic context-compaction operation took 178.7 seconds.

These waits seriously disrupt responsiveness. This is an observation of current performance, not a controlled before/after benchmark or proof of intentional throttling.

What steps can reproduce the bug?

This was observed in an existing conversation; I do not have a deterministic, minimal reproduction from a fresh chat.

1. Use the Windows Codex desktop app with the model/settings above.
2. Continue a conversation long enough for substantial context and automatic compaction.
3. Measure the intervals between the preceding tool result (or preceding response/compaction completion) and the next model-response completion, and separately measure compaction duration.

Input context in the measured responses ranged from approximately 41.5k to 235.4k tokens. Many input tokens were cached. The compaction request had approximately 238k input tokens and produced 6,080 output tokens; subsequent input context fell to approximately 41.5k.

Measurement limitation: These are effective response rates calculated from completion timestamps and reported output-token counts. The intervals exclude external tool execution and exclude compaction from the aggregate, but include request/first-token waiting, input processing, transport and client overhead. Output counts include reasoning tokens and tool-call arguments. This is not a measurement of pure model decoding speed, and the data do not isolate the bottleneck.

What is the expected behavior?

Responses and automatic compaction should complete with reasonable latency so an active conversation remains usable. Please investigate where the time is spent: request waiting, input processing, reasoning/generation, transport/client orchestration, or compaction. Please also clarify whether these latencies are expected for xhigh on the default service tier.

Additional information

Anonymized measurements:

Related reports:

• #49287 — broad slowdown and slow compaction, reported for the CLI and a different model.
• #51656 — desktop delays after tool work, reported on macOS.

This report adds measurements from the native Windows desktop app with GPT-6.1 Sol; a shared root cause has not been established.

Privacy: Only anonymized timing, token counts and generic runtime settings are included. No project names or business details, prompts, code, file paths, screenshots, conversation excerpts, session/request identifiers, or raw logs are attached.