Codex Models, Token Usage, and Best-Use Scenarios
The simple rule: do not make the biggest model carry every log file, typo fix, and CSS tweak. Use the strongest model where judgment matters, then hand clear execution to cheaper tiers.
Disclosure: StackCapybara may earn a commission if you click some tool links. This does not affect our recommendations. We focus on tested workflows, practical fit, pricing reality, and clear limitations.
Quick answer: choose by risk
Model routing cards
GPT-5.5
Best for: risky architecture, incident diagnosis, final review.
Avoid for: repetitive extraction or tiny cosmetic fixes.
StackCapybara use: release-gate review before promotion.
GPT-5.4
Best for: important implementation with clear acceptance criteria.
Avoid for: high-ambiguity root-cause investigation.
StackCapybara use: medium feature implementation after design is set.
GPT-5.3-Codex
Best for: Codex-focused coding sessions under budget pressure.
Avoid for: broad strategy or uncertain cross-system failures.
StackCapybara use: bounded code tasks with known file scope.
GPT-5.4 mini
Best for: scoped fixes, log summaries, verification passes.
Avoid for: architecture-level judgment calls.
StackCapybara use: targeted CSS/PHP bug repairs and QA notes.
GPT-5.4 nano
Best for: extraction, tagging, and routing prep.
Avoid for: security-sensitive coding and production-risk reasoning.
StackCapybara use: metadata extraction from article drafts.
Task routing cards
WordPress plugin architecture
Use: GPT-5.5
Why: hidden side effects cost more than token savings.
Security-sensitive PHP
Use: GPT-5.5
Why: stronger risk detection and safer review quality.
CSS/layout bug
Use: GPT-5.4 mini
Why: scoped and reversible with lower spend.
Log summarization
Use: GPT-5.4 mini
Why: good signal compression without premium-token burn.
Content extraction
Use: GPT-5.4 nano
Why: high-volume structured work at low cost.
Final review
Use: GPT-5.5
Why: final gate is where quality matters most.
OpenAI pricing pressure
Use this to understand cost pressure, not as live billing advice. Re-check OpenAI pricing before high-cost runs.
GPT-5.5
Input: $5.00 / 1M
Cached input: $0.50 / 1M
Output: $30.00 / 1M
Use for: high-risk judgment and final review
GPT-5.4
Input: $2.50 / 1M
Cached input: $0.25 / 1M
Output: $15.00 / 1M
Use for: bounded implementation
GPT-5.3-Codex
Input: $1.75 / 1M
Cached input: $0.175 / 1M
Output: $14.00 / 1M
Use for: Codex coding passes with clear scope
GPT-5.4 mini
Input: $0.75 / 1M
Cached input: $0.075 / 1M
Output: $4.50 / 1M
Use for: scoped fixes and verification
GPT-5.4 nano
Input: $0.20 / 1M
Cached input: $0.02 / 1M
Output: $1.25 / 1M
Use for: extraction and classification
Legacy Codex models
GPT-5.2-Codex: $1.75 input / $0.175 cached / $14.00 output
GPT-5.1-Codex: $1.25 input / output $10.00
GPT-5.1-Codex mini: $0.25 input / output $2.00
Use for: fallback/reference only
Model notes: what changes in practice
GPT-5.5
Use when: risk is high and failure is expensive.
Avoid when: the task is mechanical and narrow.
Example: final review of a WordPress patch touching data behavior.
GPT-5.4
Use when: implementation is clear and bounded.
Avoid when: root cause is still unknown.
Example: implementing approved feature slices in known files.
GPT-5.3-Codex
Use when: coding depth is needed but budget is tighter.
Avoid when: judgment-heavy direction decisions are unresolved.
Example: medium-size coding pass after planning is complete.
GPT-5.4 mini
Use when: fix scope and rollback are straightforward.
Avoid when: architecture and risk boundaries are vague.
Example: layout regression repair with targeted QA validation.
GPT-5.4 nano
Use when: output is structured extraction or classification.
Avoid when: coding safety or incident analysis is required.
Example: extracting metadata from draft content sets.
Legacy models: GPT-5.2-Codex and GPT-5.1-Codex are mostly useful as fallback/reference models if newer options are unavailable.
Token cost examples: what changes with task size
Small patch
Size: 15K input + 3K output
Cheapest sensible model: GPT-5.4 mini
Upgrade when: patch touches risky data/auth/redirect behavior.
Takeaway: for small tasks, risk matters more than cents.
Medium feature
Size: 80K input + 15K output
Cheapest sensible model: GPT-5.4 or GPT-5.3-Codex
Upgrade when: root cause or scope is still unclear.
Takeaway: scope clarity saves more than model downgrading.
Heavy repo task
Size: 250K input + 50K output
Cheapest sensible model: GPT-5.3-Codex (when risk is controlled)
Upgrade when: outcome affects release gating or rollback strategy.
Takeaway: isolate context before you scale model spend.
Show API-style example costs
| Model | Small patch | Medium feature | Heavy repo task |
|---|---|---|---|
| GPT-5.5 | $0.165 | $0.850 | $2.750 |
| GPT-5.4 | $0.083 | $0.425 | $1.375 |
| GPT-5.4 mini | $0.025 | $0.128 | $0.413 |
| GPT-5.4 nano | $0.007 | $0.035 | $0.113 |
| GPT-5.3-Codex | $0.068 | $0.350 | $1.138 |
| GPT-5.1-Codex mini | $0.010 | $0.050 | $0.163 |
Long-context warning: GPT-5.5 long-context example is $5.700 and GPT-5.4 long-context example is $2.850. Reserve long context for tasks that truly need whole-system visibility.
What wastes Codex allowance
- Vague prompts: open-ended requests create expensive exploration loops.
- Too much repo context: unrelated files burn input tokens before execution starts.
- Long-running chats: stale context increases cost turn by turn.
- Repeated failed loops: retries without diagnosis burn output fast.
- Huge command output: full dumps cost more than targeted slices.
Copy-paste prompts for coding-agent workflow
Choose model
Classify this task before acting.
Project: [project name]
Environment: dev unless explicitly stated otherwise
Task: [paste task]
Return:
1. Risk class A/B/C/D/E
2. Recommended model
3. Why this model is enough
4. Required context
5. Context to exclude
6. Stop conditionReduce context
Before editing, reduce context.
Identify:
1. Exact files likely relevant
2. Commands needed to confirm the issue
3. Minimum useful log slices
4. Files that should not be read
5. Whether this needs GPT-5.5 or can be handled by GPT-5.4 mini
Do not edit files in this step.Safe implementation handoff
You are working on dev only.
Objective:
[one clear objective]
Allowed scope:
[files/directories]
Forbidden:
- Do not touch production
- Do not modify unrelated templates
- Do not install plugins
- Do not make broad rewrites
Preflight:
- confirm pwd/site root
- run git status --short
- confirm backup or create one
- identify files before editing
Implementation:
- make the smallest patch that solves the objective
- if PHP changes, run php -l on every changed PHP file
Verification:
- test exact URL(s)
- include desktop/mobile checks if UI changed
- report commands, changed files, URLs tested, blockers, rollback notes, and final git status
Stop condition:
- stop after this task; do not continue into adjacent improvements.Copy-paste setup: risk ratings for your agent docs
Risk rating system
Before Codex starts
- Project
- Environment
- Site root
- Target URL
- Allowed files
- Forbidden files/actions
- Backup path
- Risk rating
- Model and effort
- Verification commands
- Stop condition
Copy raw Markdown template
## Agent Task Risk Rating
A - High risk:
Security, database, auth, redirects, production, payments, affiliate tracking.
Default model: GPT-5.5.
Requires backup, rollback notes, and final review.
B - Medium risk:
Multi-file feature, uncertain bug, plugin/theme interaction.
Default model: GPT-5.3-Codex or GPT-5.4 with clear scope.
Escalate if root cause is unclear.
C - Low risk:
Known file, reversible CSS/PHP/content fix.
Default model: GPT-5.4 mini.
Requires targeted verification.
D - Prep work:
Extraction, tagging, formatting, summaries.
Default model: GPT-5.4 nano or mini.
No code edits unless explicitly approved.
E - Final gate:
Pre-release review, production promotion, rollback decision.
Default model: GPT-5.5.
## Before Codex Starts
- Project:
- Environment:
- Site root:
- Target URL:
- Allowed files:
- Forbidden files/actions:
- Backup path:
- Risk rating:
- Model and effort:
- Verification commands:
- Stop condition:Future article idea: a dedicated coding-agent safety runbook with risk ratings, backup rules, and model routing examples.
For broader context, see Best AI Stack for Building a WordPress Affiliate Site and Review Methodology.
Final recommendation
- GPT-5.5 for high-risk judgment and final review.
- GPT-5.4 / GPT-5.3-Codex for bounded implementation.
- GPT-5.4 mini for scoped fixes, logs, and QA.
- GPT-5.4 nano for extraction/routing prep only.
Cut waste by tightening scope first, then selecting model tier.
If your task starts getting bigger mid-run, pause, re-rate risk, and reset scope before continuing.