Pick the model for the job, then stay on the subscription
Labs now publish a cheap everyday model, a mid coding model, and an expensive long-agent model. Match the job inside an official plan. The API rate card is why the top model eats quota — not a reason to skip Plus.
The usual mistake is buying the logo, then leaving the picker on the most expensive name. Labs now ship a ladder. The cheap rung is for classification and short chat. The middle rung is daily coding. The top rung is long agents. You pay 40–80× more to sit on the top rung by default — which is why a $20 plan is cheaper than running Astra or Fable on the API.
Takeaway
Subscribe first. Then name the job: short Q&A, repo edit, or a long agent run. Pick the cheapest SKU inside the official app that still does that job. The input / output rates below explain how fast the top model burns the plan’s window. They are not a shopping list for skipping Plus. For why the plan beats a raw key, see subscription vs API.
Scope
The model picker inside an official chat or coding plan. API rates are listed so you can see the overflow price. No gray-market model names.
The ladders on 20 November 2026
USD per million tokens, Standard, short context — the API meter, which is what you pay if you leave the plan. Open the source before you pay.
| Job | OpenAI | Anthropic | xAI | |
|---|---|---|---|---|
| Cheap / high volume | GPT-6 Luna $0.10 / $0.50 | Haiku 5.5 $0.10 / $0.50 (prompts ≤100K; 5× above that) | Gemini 3.8 Flash $0.75 / $3.75 through 31 Dec 2026 | Use Grok 4.7 only if you need the frontier SKU; the models page does not list a separate nano price |
| Daily coding / agents | GPT-6.2 Sol $2 / $10 | Sonnet 5.5 $2 / $10 or Opus 5.5 $4 / $20 | Gemini 4 Hydrogen $1.50 / $7.50 | Grok 4.7 $2 / $6 under 200K prompt tokens |
| Long / hardest agents | GPT-6.2 Astra $8 / $40 | Fable 5.2 $8 / $40 | Gemini 4 Carbon or Argon $2 / $10 through 30 Nov, then $4 / $20 | Same Grok 4.7; prompts ≥200K bill $4 / $12 for the whole request |
xAI’s models page says: for code and chat, use Grok 4.7. Images, video, and voice are separate SKUs.
On a ChatGPT Plus subscription the same ladder shows up as a quota, not a token invoice. OpenAI’s Learn page estimates local Codex messages per five-hour window: Luna 350–3,000, GPT-6.1 Sol 15–160, GPT-6.2 Astra 6–50. Weekly limits may also apply. Pro currently has no five-hour limit on that page.
Claude paid plans share one pool across web, desktop, mobile, and Claude Code. Fable on Pro is usage credits. Max and Ultra include Fable up to half the weekly cap.
How to choose in one sitting
- Buy the official entry plan for the product you actually open (chat or coding).
- Write three real tasks from last week.
- Run the cheap SKU first. If it fails the job, step up one rung.
- If the five-hour or weekly cap is gone every week on the top model, step up the plan (Max / Ultra / ChatGPT Pro), not to a gray key.
- Only open an API key if you are scripting outside the app, or spending bundled Pro/Max/Ultra credits. Cache, Batch, and long-context cliffs matter on that invoice: Anthropic cache reads are a fraction of input; OpenAI long context is >272K input; xAI doubles the whole request at 200K prompt tokens; Claude Haiku 5.5 is 5× above 100K.
FAQ
Is the most expensive model “the best”? Vendors describe Astra, Fable, and Carbon as the long-agent / hardest-reasoning SKUs. They are the wrong default for summarization and classification. Luna, Haiku, and Flash exist for that — use them on the plan so the expensive names last.
Can I trust a third-party dropdown that prints the same name? A printed name is a vendor claim. Official identity is the model id on your own account. See relays and official token providers.
Should I start with a subscription or an API key? Subscription. Direct API is the overflow meter and the automation meter, not a cheaper Plus. See subscription vs API.
What about $10 plans that advertise $100 of usage? Those are coding-agent passes. The $100 is a named Flash / open-weight allowance, not Astra or Fable.
Last checked 2026-11-20 against the links below. Open the vendor page in your region before you pay — list prices move.
- Official · OpenAI API pricing · Open source · GPT-6 Luna $0.10 / $0.50. GPT-6.2 Sol $2 / $10. GPT-6.2 Astra $8 / $40. Short context ≤272K tokens.
- Official · Claude pricing — latest models · Open source · Haiku 5.5 $0.10 / $0.50 (prompts ≤100K). Sonnet 5.5 $2 / $10. Opus 5.5 $4 / $20. Fable 5.2 $8 / $40.
- Official · Gemini Developer API pricing · Open source · Hydrogen $1.50 / $7.50. 3.8 Flash $0.75 / $3.75 through 31 Dec 2026. Carbon/Argon $2 / $10 through 30 Nov 2026.
- Official · xAI models · Open source · Grok 4.7 is the lab’s recommended model for code and chat. $2 / $6 under 200K prompt tokens; double at or above 200K.
- Official · ChatGPT Learn pricing · Open source · Plus local Codex estimates per five-hour window: Luna 350–3,000, Sol 15–160, Astra 5–50. Pro currently has no five-hour limit there.