Reducing ChatGPT usage starts with choosing the right surface. Chat, ChatGPT Work, and Codex serve different jobs and can follow different usage rules.
Then reduce context that does not change the result. A short, complete request usually needs fewer follow-up turns than a vague request.
Choose the right ChatGPT surface
OpenAI’s ChatGPT guide, checked on October 14, 2026, describes three paths:
- Chat supports questions, brainstorming, drafting, summaries, and comparisons.
- ChatGPT Work carries a larger task to a reviewable result.
- Codex provides developer tools and technical detail for software work.
Use Chat for conversation. Use Work for a multi-step deliverable. Use Codex when the task needs repository access, commands, diffs, tests, or pull-request views.
OpenAI says ChatGPT Work and Codex share usage limits. Do not assume every ordinary Chat limit uses the same meter.
Keep one topic in each chat
Start a new chat when the subject changes. Old messages can add context that the new question does not need.
Continue the existing chat when later instructions depend on earlier decisions. The goal is relevant continuity, not the shortest possible transcript.
Include only context that changes the answer
State the goal, audience, constraints, source material, and required format. Remove background that does not affect those choices.
Attach the relevant pages from a long document when possible. Ask for a focused date range when current research does not need the full archive.
Do not paste the same source into several turns. Refer to the earlier attachment when it remains available.
Use Projects for repeated source material
Projects keep related chats, files, and instructions together. They work well for a recurring report, product area, client, or research topic.
Keep Project instructions short. Put task-specific requirements in the current chat, where they apply only to the work at hand.
Review Project files periodically. Remove obsolete drafts and duplicate exports that can confuse retrieval.
Control tools and retrieval
Web search, plugins, connected apps, code execution, and file analysis can add useful context. They can also add retrieved content and additional model steps.
Enable a tool because the task needs it. Ask ChatGPT not to search the web when the answer should come only from supplied documents.
For a comparison, name the sources or date range. For an internal document, say which files are authoritative.
Specify the output before generation
Define the length, format, and decision you need. Ask for five bullets when five bullets are enough.
Separate required content from optional ideas. This reduces long outputs and revision turns that do not improve the deliverable.
Review the first result before requesting several variants. Keep the variant that solves a real decision.
Treat coding as a workflow
Move software implementation to Codex when you need developer tools. A coding task often includes exploration, commands, tests, review, and retries.
Measure the complete workflow. A concise final answer can hide substantial work performed by tools or workers.
OpenAI’s pricing guide, checked on October 14, 2026, says local Codex messages and cloud chats share a plan allowance. Model, context, reasoning, tools, retrieval, and caching affect usage.
Read how to reduce Codex usage for coding context, model effort, tools, and worker accounting.
Where JEV and Halv apply
Halv does not modify Chat or Work conversations on chatgpt.com. Its usage tools apply to supported coding-agent workflows installed on the user’s machine.
JEV classifies supported coding tasks by difficulty, reasoning needs, risk, and confidence. Halv then selects a configured model tier and reasoning effort.
This can reduce use of higher-cost models for bounded subtasks. It does not guarantee a fixed token or plan-usage reduction.
Read the benchmark carefully
Our SWE-rebench checkpoint compared complete Codex workflows across 14 tasks and three repetitions.
Both arms passed 25 of 42 runs. Halv recorded $144.67 in selected model cost, compared with $337.50 for vanilla Codex.
The Halv workflow used 3.49 times as many tokens. The 57.1% recorded cost result does not establish the same reduction in tokens, plan usage, or every task.
It also does not measure ordinary ChatGPT conversations.
A practical weekly review
Review chats that required many corrections. Identify missing instructions, unnecessary files, broad searches, and oversized output requests.
For a controlled check, record the applicable account meter and reset time. Complete one representative task without crossing a reset.
Record the meter after an accepted result. Include retries, Work or Codex workers, and failed attempts when they share that allowance.
Compare the percentage of allowance consumed, when the interface reports it. Do not convert API token prices into a subscription-usage percentage.
Check current limits and credits in the account usage dashboard. Do not convert API token rates into a promised number of subscription messages.
Continue with the detailed ChatGPT and Codex token guide, Codex usage limits, or how to reduce Claude usage.
Sources
Sources checked on October 14, 2026: