Codex usage grows from more than your latest prompt. Repository instructions, files, tool results, reasoning, workers, and retries can all consume the plan allowance.
The best reduction strategy is operational. Give Codex less irrelevant work, then use enough model capability to pass the task’s checks.
Halv uses JEV to classify supported tasks. The Halv app then selects the model tier and reasoning effort from its configured routing ladder.
A strong model can coordinate and review while bounded subtasks use suitable tiers. JEV routing is included in Halv without a separate routing fee.
Separate usage from tokens and price
These four measures answer different questions.
| Measure | Question it answers |
|---|---|
| Tokens | How much input and output did the models process? |
| Credits or cost | How was that work priced? |
| Plan usage | How much included allowance remains? |
| Subscription | What recurring price does the plan charge? |
OpenAI’s pricing guide, checked on October 14, 2026, says model, context, reasoning, tools, retrieval, caching, and execution location all affect usage.
The same guide says API token prices are separate from subscription usage. Therefore, an API rate cannot predict a fixed number of included Codex tasks.
1. Give each task a clean context
Start a new chat when you move to unrelated work. A coding conversation can accumulate file contents, diffs, command output, and rejected approaches.
Use compaction when the same task needs continuity. Compaction processes the existing history, so it also consumes usage.
For more detail, read Codex compact versus new chat.
2. Scope the request before Codex explores
Name the target file, required behavior, constraints, and verification command. A precise brief reduces broad searches and clarification turns.
Keep optional improvements separate. Otherwise, a bounded fix can become a wide review of the repository.
3. Keep repository instructions focused
Codex loads applicable AGENTS.md instructions into context. Keep shared rules
short, and place directory-specific rules near the code they control.
Remove duplicate prose. Preserve instructions that affect safety, tests, style, or release behavior.
See how AGENTS.md affects token usage for a focused audit process.
4. Limit tools and command output
Enable only the MCP servers and tools required for the task. Tool definitions, retrieved content, and results can all add context.
Filter large logs to the failing command, relevant stack trace, and final summary. Keep the full log in a file for later inspection.
Read how much MCP context adds before removing a tool that the workflow needs.
5. Match model and effort to difficulty
OpenAI’s subagent guidance, checked on October 14, 2026, recommends different models and effort levels for different tasks. Higher effort can improve difficult work, but it uses more tokens and time.
Use lower effort for clear, mechanical work. Increase it when the task requires planning, edge-case analysis, or deeper validation.
JEV applies this principle automatically in supported Halv coding workflows. It classifies task difficulty, reasoning needs, risk, and confidence.
Halv then selects a configured model tier and reasoning effort. This can reduce use of higher-cost models when a lower tier fits the subtask.
6. Count every worker and retry
Subagents can protect the main conversation from large logs. Each worker still sends its own model requests.
Measure coordinator sessions, workers, retries, and failed runs. A smaller main chat does not prove lower total usage.
7. Watch the shared allowance
OpenAI says local messages and cloud chats share the Codex plan allowance. ChatGPT Work and Codex also share usage, credits, and limits.
Check the account usage page for current reset times and available credits. Published message estimates are not fixed quotas.
Our Codex usage limits guide explains the difference between windows, credits, API billing, and subscription pricing.
Measure allowance use on a real task
Record the account usage meter and reset time before the task. Complete one representative task without crossing a reset.
Record the meter again after the result passes your checks. Also record retries, worker sessions, and failed attempts.
Compare the percentage of plan allowance consumed. Do not substitute an API token rate, because API pricing does not define subscription allowance usage.
What the Halv benchmark establishes
Our SWE-rebench checkpoint covered 14 tasks with three repetitions each. Both workflows passed 25 of 42 runs.
Halv recorded $144.67 in selected model cost. Vanilla Codex recorded $337.50. That is 57.1% lower recorded cost for this complete toolkit and sample.
The Halv workflow used 3.49 times as many tokens. The result does not prove a 57.1% plan-usage reduction, fewer tokens, or universal savings.
Halv does not modify conversations on chatgpt.com. It supports coding-agent workflows where Halv is installed and configured.
Continue with how to reduce ChatGPT usage, how to reduce Claude usage, or the detailed ChatGPT and Codex token guide.
Sources
Sources checked on October 14, 2026: