gpt5.4 uses more limits than 5.3codex.

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #13568
• #13567
• #13317
• #13186
• #13167

Powered by Codex Action

Is this just based on vibes? Can you provide any concrete details or side-by-side comparisons?

Even more? This is crazy, thanks for the heads up. And yeah, it seems like they're so busy with launches that they don't have time to address the bugs that make the product unusable.

Is this just based on vibes? Can you provide any concrete details or side-by-side comparisons?

Any news on this? https://github.com/openai/codex/issues/13186, we can't work.

I am also seeing accelerated usage on the same workload from a couple hours ago.

Just did two test tasks with maybe total 300k tokens, of which 20k were output tokens and my pro usage jumped by 4% of my pro account. Usually that was <1%. No fast mode and no 1m+ context enabled.

I just confirmed that 5.4 does have a higher cost, and therefore a higher usage multiplier than 5.3-codex. Apologies — I wasn't aware of this until just now.

We've updated our documentation here to reflect this.

You can choose to stick with 5.3-codex if you find that 5.4 does justify the additional cost.

Vibes locked in.

I'm noticing the same issue. I'm on the Pro plan. I went through almost 70% of my usage in one day. I usually end the week with at least 10% or more remaining, and I don't feel I had an abnormal workflow that day in comparison to other days. My limits just reset on the 4th at 3:30 PM.

User ID: user-lkqg4raupjouznjF0npCN8cK

Based on an analysis I've done of the logs and the token usage, I believe this is largely related to the fork_context feature on spawn_agent.

---

What happened

A single session consumed 1.16B tokens (~50% of my weekly limit). The session started as "run e2e tests" (tests passed in 40s), then continued with follow-up tasks across 50+ user messages.

The model spawned 22 subagents, 12 with fork_context=true. The ones spawned with fork_context=true consumed 70-220M tokens each. The ones with fork_context=false were all under 2M.

Token breakdown

| Agent | fork_context | Tokens | Duration |
|-------|-------------|--------|----------|
| Parent thread | — | 230M | — |
| Dewey (worker) | true | 219M | 5 min |
| Avicenna (explorer) | true | 157M | 6 min |
| Arendt (explorer) | true | 142M | 5 min |
| Pasteur (explorer) | true | 78M | 5 hrs |
| Boyle (explorer) | true | 78M | 4.8 hrs |
| Aquinas (explorer) | true | 77M | 6 min |
| Raman (explorer) | true | 77M | 5 min |
| Nash (explorer) | true | 69M | 5 min |
| Godel (worker) | true | 23M | 13 min |
| 12 others | false | <2M each | — |
| Total | | 1.16B | |

The model's fork_context decisions

The first 3 spawns correctly used fork_context=false. Starting at spawn #4, the model began using fork_context=true for tasks that don't appear to need full session history:

[937]   fork_context=True  "Implement the dedicated organizations CRUD E2E setup helper..."
[3880]  fork_context=True  "Inspect apps/web/src/components/calculator/sales-entry-form.test.tsx..."
[3881]  fork_context=True  "Inspect apps/web/src/components/calculator/calculator-page.test.tsx..."
[7955]  fork_context=True  "Audit apps/web/src/components/calculator/calculator-page.pending-shield.test.tsx..."
[7958]  fork_context=True  "Audit apps/web/src/components/calculator/calculator-page.test.tsx..."
[11309] fork_context=True  "Audit the current Convex timeout cluster in this repo without editing files..."
[11310] fork_context=True  "Inspect the current quality-gates / Turbo orchestration..."

The forked subagents inherited the full session state (79 user messages, thousands of tool results). Dewey's rollout shows 84 turns, 16 compactions, and 1,061 tool calls in 5 minutes.

Config at the time

• Codex CLI v0.111.0
• Model: gpt-5.4, reasoning effort: high/xhigh
• max_threads = 8, max_depth = 2

Session details

• Session ID: 019cbfdd-a496-7b31-b017-a0c1e1467051
• Rollout available if needed for investigation