Quick answer
Codex access comes through ChatGPT plans with usage allowances, billed in token-based credits since April 2026: work consumes credits by model and volume. Cloud tasks and the local CLI draw on the same plan, heavier tiers burn faster, and the best lever is tier routing, GPT-6.1 Sol by default, GPT-6 Astra where it pays, GPT-6 Luna for mechanical work.
- Access
- Via ChatGPT plans; cloud tasks and local CLI share one allowance
- The meter
- Moving to token credits in 2026, consumed by model and volume
- Burn rate
- Scales with tier; GPT-6 Astra fastest, GPT-6.1 Sol middle, GPT-6 Luna slowest
- Community read
- The $20 tier's Codex allowance is widely considered generous
- Exact numbers
- Change often, check OpenAI's Codex pricing docs for current terms
01
The Stable Mechanics
Codex metering has changed shape more than once, plan-based allowances, then a shift toward token credits, and pages that lead with specific quota figures go stale fast. What has stayed stable is the architecture underneath, and it is what you need to plan around.
Access rides on ChatGPT plans
Codex is included with paid ChatGPT plans, each carrying a usage allowance that scales with the tier. You are not buying a separate Codex subscription; you are spending an allowance attached to the plan you already have.
Token credits since April 2026
Since April 2, 2026, Codex on ChatGPT plans is metered in token-based credits: work consumes credits in proportion to the model used and the tokens processed, and extra credits can be purchased beyond the plan's allowance. Metering by what work actually costs, rather than counting tasks or messages.
One plan across every surface
Cloud tasks kicked off from the web or IDE and local Codex CLI sessions draw on the same allowance. Delegating to the cloud moves the compute, not the bill, there is no separate cloud budget to arbitrage.
Heavier tiers burn faster
Credit consumption scales with the model tier: a GPT-6 Astra session costs meaningfully more of your allowance than the same work on GPT-6.1 Sol or GPT-6 Luna. Under credit billing, model choice is a spend decision, not just a quality one.
One data point worth having, attributed as what it is: community sentiment across dev forums has generally rated Codex allowances at the $20 tier as generous, with many engineers reporting moderate daily use rarely hits the ceiling. That is a sentiment about a moving target, not a spec, but it is consistent enough to be worth knowing.
Your AI workspace for shipping software.
Download AIDEN free and point it at your existing Claude Code or Codex setup. No credit card, running in minutes.
Download AIDEN freeFree to start · macOS 12+ · No credit card required
02
Tier Burn Rates: The Price Ratio Is the Proxy
OpenAI's API token prices are the honest public proxy for how fast each tier consumes credits, they track what each model costs to run. The ratios are the point: routing a session from GPT-6 Astra down to GPT-6.1 Sol cuts its burn 5x, and down to GPT-6 Luna 100x.
| Model | API price (burn proxy) | Relative burn | Route here |
|---|---|---|---|
| GPT-6 Luna | $0.10 in / $0.50 out per MTok | 1x input / 1x output (baseline) | Quick mechanical edits, triage, high-volume cheap tasks |
| GPT-6.1 Sol | $2 in / $10 out per MTok | 20x input / 20x output | The default for complex coding and agentic work |
| GPT-6 Astra | $10 in / $50 out per MTok | 100x input / 100x output | The hardest end-to-end tasks, when a failed run costs more than the premium |
Prices verified against OpenAI's docs on September 30, 2026. Which tier earns its burn for which task shape, and which reasoning effort to start at, is covered in the best Codex model and the head-to-head GPT-6 Astra vs Sol vs Luna.
03
Three Strategies for Budgeting Codex Work
Treat the allowance as a budget and the levers become clear. Three of them do most of the work.
1. Route tiers by task
GPT-6.1 Sol as the default, GPT-6 Astra only where the task genuinely needs the top tier, GPT-6 Luna for mechanical high-volume work. Under credit billing this is the single biggest lever on how long an allowance lasts. Decision rules in best Codex model.
2. Add a second pool
A Claude plan is an independent budget, its limits never touch OpenAI's. Engineers who run both CLIs route work across two pools and hit ceilings far less often; AIDEN runs both on one board, which keeps the routing visible. The workflow is in Claude Code and Codex together.
3. Know when the API wins
Plan credits suit steady daily use. Per-token API billing wins at the extremes: bursty occasional use that never justifies a monthly fee, or sustained automation where an uncapped, deliberately budgeted per-token spend beats bouncing off plan ceilings.
QUESTIONS
FAQ
Do Codex cloud tasks and the local CLI share the same limits?
Yes. Cloud tasks delegated from the web or IDE and sessions run through the local Codex CLI draw on the same ChatGPT plan allowance. Offloading work to the cloud changes where it runs, not whose budget it spends. One plan, one budget, across every surface.
What are Codex credits?
Credits are OpenAI's token-based unit for metering Codex work. Since April 2, 2026 (April 23 for existing Enterprise and Edu plans), Codex usage on ChatGPT plans is priced by input, cached input and output tokens instead of per message. Work consumes credits in proportion to the model used and the volume of tokens processed, so a long GPT-6 Astra session costs meaningfully more credits than the same session on GPT-6.1 Sol or GPT-6 Luna. Plans include an allowance, and additional credits can be purchased when you need more.
Which Codex model burns through limits fastest?
Heavier tiers burn faster, and API token prices are the best public proxy for the ratio. GPT-6 Astra ($10 in / $50 out per MTok) is the fastest burn of the current family, 5x GPT-6.1 Sol ($2 in / $10 out per MTok), which in turn costs 20x GPT-6 Luna ($0.10 in / $0.50 out per MTok) on both input and output. Defaulting to GPT-6.1 Sol, OpenAI's recommended Codex model for complex coding, and reserving Astra for genuinely hard work is the biggest single lever on how far an allowance goes.
Is the $20 ChatGPT plan enough for Codex?
Community sentiment, and it is sentiment, not a guarantee, has generally rated Codex allowances at the $20 tier as generous relative to competitors, with many engineers reporting they rarely hit the ceiling on moderate daily use. Heavy parallel agent use is a different feature on any plan. Exact allowances change, so check OpenAI's official Codex pricing docs for current terms.
Should I use plan credits or API billing for Codex?
Plan credits win for steady, predictable daily use: a flat monthly cost with a known allowance. Per-token API billing wins at the extremes, occasional bursty use that never justifies a subscription, or sustained heavy automation where you would rather budget an uncapped per-token spend than repeatedly hit plan ceilings. Most individual engineers land on a plan; most CI-style automation lands on the API.