Skip to content

[Feature Request] Restore full context window for gpt-5.6-sol to improve performance and answer quality #36145

Description

@youlongsam

What variant of Codex are you using?

Chat GTP APP 26.721.81911

What feature would you like to see?

Problem Description

When using Codex with the gpt-5.6-sol model at both medium and high reasoning effort settings, I am experiencing significant performance issues:

  1. Slow response times – Even for simple, routine coding tasks, the model takes noticeably longer to respond compared to other tools like Claude Opus 4.8. The delay is consistent across all types of questions and persists even when using the "medium" reasoning setting, which I would expect to be reasonably fast.

  2. Poor answer quality – The generated code and solutions are often incomplete, logically flawed, or miss key context from the conversation. Using the "high" reasoning setting does not noticeably improve quality, which is particularly frustrating given the longer wait time.

My primary concern is the overall user experience: the combination of slow responses and low-quality outputs makes Codex significantly less effective than alternatives. Even for straightforward coding tasks, I find myself having to rephrase questions multiple times or switch to other tools to get useful results.

I've noticed that the model frequently loses track of earlier conversation context, even when the total discussion length is moderate. This leads me to suspect that the effective context available in Codex may be more limited than what the gpt-5.6-sol model is officially capable of, which could be contributing to both the slow reasoning and the poor output quality.

Proposed Solution

I would like to request the following:

  1. Investigate and optimize response speed – Identify the bottlenecks causing consistent delays across all question types and reasoning effort levels.

  2. Restore the full context capability for gpt-5.6-sol in Codex, so that users can actually benefit from the model's advertised capabilities – this may help address both the speed and accuracy issues.

  3. Provide transparency – if the model's reasoning budget, context window, or other performance-related parameters are being modified, users should be notified in advance.

  4. Publish official performance benchmarks for gpt-5.6-sol within Codex at different reasoning effort levels, so users can make informed decisions about which setting to use for different tasks.

Why This Matters

As a developer, I rely on Codex to handle daily coding tasks efficiently. The current performance makes the tool significantly less useful than alternatives like Claude Code (with Opus 4.8). Improving response speed and answer quality would restore my confidence in Codex as a primary development tool.

Additional information

I'm comparing Codex with Claude Code (the CLI tool, accessed via claude command in iTerm2) using the Claude Opus 4.8 model on the same coding tasks.

Metadata

Metadata

Assignees

No one assigned

    Labels

    appIssues related to the Codex desktop appcontextIssues related to context management (including compaction)enhancementNew feature or requestmodel-behaviorIssues related to behaviors exhibited by the modelperformance

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions