Codex is starting to feel less like one tool and more like a control panel: four model buttons, two usage buckets, and three token prices. The comparison matters more than the branding.
The model matrix
| Model | Context | Images | API input / output per 1M | Usage on my $200 Pro 20× plan | I would use it for | Main downside |
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 1,050,000 | Yes | $5 / $30 | Shared bucket | Long migrations, ambiguous debugging | Highest price |
| GPT-5.6 Terra | 1,050,000 | Yes | $2.50 / $15 | Shared bucket | Large repositories, everyday work | Middle choice still requires guessing |
| GPT-5.6 Luna | 1,050,000 | Yes | $1 / $6 | Shared bucket | Cheaper multimodal tasks | Uses the same shared allowance |
| GPT-5.3 Codex Spark | 128,000 | No | Not finalized | Separate preview bucket | Small edits, refactors, quick CSS work | Older model, 8.2× less context, no screenshots |
Spark’s separate bucket is the interesting part. During the research preview, it does not consume my shared 20× allowance, so I can preserve GPT-5.6 usage for larger jobs. Pricing and availability are still provisional.
The separate allowance is also a budget-design decision. A pooled cap encourages substitution; a separate bucket encourages experimentation. I saw the organizational version when my company combined several AI budgets into one pool.
What I would choose
| Task | First choice | Why it fits | Usage consequence | What would make me switch |
|---|---|---|---|---|
| Rename, focused refactor, quick CSS edit | Codex Spark | Fast feedback for a small text-only change | Preserves the shared 20× bucket | The change spreads across the repository |
| Fix a UI from a screenshot or mockup | GPT-5.6 Luna | Image input at the lowest GPT-5.6 API price | Uses the shared bucket | The debugging requires more reasoning |
| Explore an unfamiliar large repository | GPT-5.6 Terra | 1.05M context with middle-tier pricing | Uses the shared bucket | The task is either trivial or unusually difficult |
| Plan a migration across many files | GPT-5.6 Sol | Maximum reasoning headroom and full context | Uses the shared bucket fastest | A smaller model proves sufficient |
| Continue after the shared limit gets tight | Codex Spark | Separate research-preview allowance | Leaves the shared bucket alone | I need images or more than 128k context |
The trade is not subtle: Spark is fast and quota-friendly, but text-only with roughly one-eighth of GPT-5.6’s context. A screenshot or a large repository can send me straight back to another button.
AI pricing is becoming UX because choosing a model now means choosing which limitation I want to hit.
Quota shape matters as much as token price. My Grok Build trial made that concrete: a weekly ceiling changed whether I was willing to begin a task at all.