GPT-5.6 Terra and Sol exhaust the 5-hour Codex usage limit within minutes #32606

What version of Codex CLI is running?

codex-cli 0.144.1

What subscription do you have?

ChatGPT Plus with additional purchased Codex credits

Which model were you using?

gpt-5.6-terra (xhigh) and gpt-5.6-sol; rolled back to gpt-5.4 (xhigh)

What platform is your computer?

Windows 11 Core 64-bit, version 10.0.26200, locale it-IT

What terminal emulator and version are you using (if applicable)?

Windows Terminal with PowerShell

Codex doctor report

Codex doctor overallStatus: warning
Codex version: 0.144.1
Platform: Windows 11 Core 64-bit, version 10.0.26200
Install method: npm
Authentication: configured, ChatGPT auth
Network: OpenAI API and Responses WebSocket reachable
Config: loaded successfully
Databases: integrity OK
Updates: current version is latest
Warnings:
1. One old empty rollout session file from 2026-02-03 caused a rollout/state inventory warning.
2. Terminal width is 76 columns; recommended >=80.
No installation, authentication, network, database, or update issue was detected that explains the abnormal GPT-5.6 usage.

What issue are you seeing?

GPT-5.6 models are consuming Codex usage and purchased credits at an abnormally high rate.

With GPT-5.6 Terra using xhigh reasoning, a routine repository pre-audit/pre-report phase consumed almost the entire 5-hour usage window within only a few minutes, before completing meaningful work.

GPT-5.6 Sol was able to continue the task more reliably than Terra, but its usage consumption was also extremely high. I purchased approximately €40 in additional Codex credits during the same day in an attempt to continue the workflow, but the available usage and credits were consumed too quickly for the model to remain practically usable.

I ultimately had to roll back to GPT-5.4 xhigh. GPT-5.4 can work for long periods on the same repository, and has previously handled substantially larger and more complex repositories, without comparable usage consumption.

This appears to be abnormal usage accounting, excessive hidden reasoning, repeated context processing, a tool/sub-agent loop, or a GPT-5.6 integration regression rather than normal consumption proportional to the task.

The current session may show GPT-5.4 because I already rolled back, but the attached feedback logs were uploaded from the session where the GPT-5.6 consumption issue occurred.

What steps can reproduce the bug?

Uploaded thread: 019f4b1e-9a58-7120-9647-34a6b386b124

What is the expected behavior?

GPT-5.6 usage should remain proportional to the complexity and duration of the task.

A routine repository audit or pre-report phase should not consume almost an entire 5-hour usage window within a few minutes, especially before producing meaningful work.

Purchased credits should provide a practically usable amount of work. GPT-5.6 Terra and GPT-5.6 Sol should not consume dramatically more usage than GPT-5.4 xhigh on the same repository and comparable workflows unless this difference is explicitly documented and expected.

Additional information

Environment:

• Codex CLI: 0.144.1
• Subscription: ChatGPT Plus with additional purchased Codex credits
• Platform: Windows 11 Core 64-bit, version 10.0.26200
• Terminal: Windows Terminal with PowerShell
• Affected models: GPT-5.6 Terra xhigh and GPT-5.6 Sol
• Comparison model: GPT-5.4 xhigh

Observed behavior:

• GPT-5.6 Terra consumed almost the entire 5-hour usage window within only a few minutes during a routine pre-audit/pre-report phase, before completing meaningful work.
• GPT-5.6 Sol continued more reliably than Terra but also consumed usage and purchased credits at an unsustainable rate.
• Approximately EUR 40 in additional Codex credits were purchased during the same day in an attempt to continue the workflow, but the available usage was still exhausted too quickly.
• I had to roll back to GPT-5.4 xhigh.
• GPT-5.4 xhigh can work for long periods on the same repository and has previously handled substantially larger and more complex repositories without comparable consumption.

Reproduction summary:

1. Open an existing repository in Codex CLI 0.144.1.
2. Select GPT-5.6 Terra with xhigh reasoning.
3. Request a routine repository pre-audit and report.
4. Observe almost the entire 5-hour usage window being consumed within approximately five minutes, before meaningful work is completed.
5. Run a comparable workflow with GPT-5.6 Sol and observe similarly excessive usage and purchased-credit consumption.
6. Roll back to GPT-5.4 xhigh and observe substantially more sustainable usage.

The codex doctor report was generated after rolling back to GPT-5.4, so the configured model shown there is GPT-5.4. The affected GPT-5.6 session logs were uploaded through /feedback using the thread ID already included above.

Codex doctor otherwise reports healthy installation, authentication, network connectivity, databases, and update status. The only warnings concern one old empty rollout file from February 2026 and a terminal width below 80 columns; neither appears related to GPT-5.6 usage accounting.