Codex CLI Token Optimization: Token and Credit Burn
A source-backed engineering report on Codex CLI token economics: model selection, AGENTS.md layering, MCP allowlists, tool-output limits, and compaction thresholds, drawn from current OpenAI docs.
The highest-value controls are visible in the docs.
The report keeps only source-backed metrics and treats savings as workload-dependent unless the cited source gives a number.
Current Codex docs list GPT-5.5 and describe lower token use than GPT-5.4.
Plus tier local-message range per 5 hours for GPT-5.5 in the source table.
Pro 20x tier local-message range per 5 hours for GPT-5.5.
enabled_tools and disabled_tools limit exposed MCP surface.
Token waste usually comes from repeated context, not one bad prompt.
Codex token pressure comes from persistent instructions, AGENTS.md layers, tool schemas, MCP servers, conversation history, file excerpts, command output, image inputs, web-search context, and memory features.
Short durable context, scoped tools, measured output, and deliberate compaction beat blanket context hoarding.
Always-on instructions
Large memory files and broad project docs are paid for repeatedly.
Scoped surfaces
Expose only the tools, files, and context needed for the current workflow.
Measure first
Usage commands and context inspection prevent guesswork.
Model Choice Is a Control
The current Codex docs put GPT-5.5 in the default decision path.
OpenAI documentation lists GPT-5.5 in the Codex model set and describes lower token usage than GPT-5.4 for comparable results.
The practical levers are small enough to operate.
Usage-limit guidance names AGENTS.md size as a tactic.
Keep repo instructions short and scoped by directory.
Config exposes a byte cap for AGENTS.md ingestion.
Set a smaller cap for bloated project docs.
Cached input tokens are billed separately and reusable prefixes are documented.
Keep durable instructions stable and volatile notes later.
Plus tier local-message range per 5 hours for GPT-5.5 in the source table.
Do not convert this into a universal savings percentage. The report keeps the number tied to its original source and context.
Controls map to operating decisions.
| Lever | Documented behavior | Operating decision | Source |
|---|---|---|---|
| AGENTS.md size and layering | Usage-limit guidance names AGENTS.md size as a tactic. | Keep repo instructions short and scoped by directory. | [codex-pricing] |
| project_doc_max_bytes | Config exposes a byte cap for AGENTS.md ingestion. | Set a smaller cap for bloated project docs. | [codex-config] |
| Prompt caching | Cached input tokens are billed separately and reusable prefixes are documented. | Keep durable instructions stable and volatile notes later. | [codex-pricing], [openai-prompt-caching], [openai-responses] |
| model_auto_compact_token_limit | Config exposes an automatic compaction threshold. | Set predictable compaction points for long tasks. | [codex-config] |
| model_reasoning_effort | Supported models expose adjustable reasoning effort. | Lower effort for routine edits and raise it for high-risk work. | [codex-config] |
Lowest-waste default workflow
The operating pattern keeps useful context close and noisy context out of the recurring prompt path.
Choose model deliberately
Start with GPT-5.5 for hard work when available; use smaller models for routine edits and narrow changes.
Check status
Run /status before long tasks and after large file reads.
Compact at boundaries
Use /compact at task boundaries and /new after commits or when changing problem areas.
Keep AGENTS.md small
Keep global AGENTS.md small; place module-specific rules in nested AGENTS.md files.
Scope MCP tools
Scope MCP tools with enabled_tools and disable entire MCP servers outside projects that use them.
Claims Removed or Downgraded
The older report contained or implied model guidance limited to GPT-5.4, API dollar pricing without a live pricing source, undocumented internal ceilings, and issue-specific bug conclusions not present in official docs.
No fabricated stats, no fake case studies, no universal savings claims without a source-backed measurement path.
[codex-pricing]
OpenAI Codex pricing
[codex-config]
OpenAI Codex configuration reference
[codex-slash]
OpenAI Codex CLI slash commands
Make Codex token spend a measurable advantage.
Use the documented controls for model choice, AGENTS.md layering, MCP scope, prompt caching, and compaction before waste becomes an engineering tax.
A practical report for teams operating Codex CLI at scale.
Turn token pressure into operating decisions.
The full report gives leaders a source-backed control set and a workflow for applying it.
- Model and context cost signals
- AGENTS.md, MCP, and caching controls
- Decision matrix for production workflows