ChatGPT and Codex Shared Allowance in December: How to Plan

ChatGPT and Codex will share one usage allowance on Plus, Pro, and Business starting in December 2026. As of 2026-11-02, OpenAI’s announcement confirms token-based measurement and exempts Instant chat. Measure your current Codex use before upgrading, then choose model, reasoning, and speed settings that finish your work within budget.

The announcement does not specify an exact December start date, plan token budgets, or future reset windows. Those gaps prevent a reliable calculation of how much coding capacity you will gain or lose.

What changes in December, and what applies now?

The confirmed change extends sharing across ChatGPT and Codex; Work and Codex already share included usage. Use this distinction when comparing your current account with the December announcement.

Question	Current rules as of 2026-11-02	December announcement
What shares usage?	Work and Codex already share an allowance.	ChatGPT and Codex share usage on Plus, Pro, and Business.
How is consumption measured?	Model, context, reasoning, tools, and speed affect usage. Message counts are unreliable.	Usage is measured in tokens and depends on model and task.
What happens to Instant?	Check the rules for your current ChatGPT experience.	Instant chat does not consume the announced shared allowance.
Which reset windows apply?	Five-hour and weekly windows can apply, depending on plan.	The announcement does not specify the future window structure.
Where do you check?	Settings → Usage shows current allowance and resets.	Settings → Usage will show combined remaining usage and reset times.

The current Work and Codex guide says both windows must have allowance when both apply. Pro 100, Pro 200, and Pro 500 currently have no five-hour Work/Codex limit. Business Premium seats also currently have no five-hour limit. The December notice does not say whether these exceptions continue.

Free, Go, Enterprise, and Edu are not named in that notice. Do not apply its Plus, Pro, and Business announcement to those plans without separate guidance.

One earlier uncertainty is resolved: the October 29 release notes confirm that OpenAI canceled the planned Pro 200 allowance reduction. Existing higher allowances stay; lower allowances doubled to match. Our guide to the recent Pro allowance changes explains that separate update.

OpenAI expects most users will not wait for resets more often. That is a forecast, not an individual allowance guarantee.

How do you measure your current Codex usage?

Record changes in the account usage display across representative tasks, including retries. Start with your existing workload rather than converting a subscription price into an assumed token budget.

Open Settings → Usage in the account and workspace you use for Codex.
Record each displayed window’s remaining allowance and reset time.
In your active Codex CLI session, enter /status.
Record the model, reasoning level, speed, and whether the task runs locally or in the cloud.
Complete a familiar, bounded task, such as fixing one failing unit test.
Recheck the usage display after it updates. Record the result and all retries needed to accept it.

OpenAI documents checking limits in the dashboard and CLI. The /status reference also describes session configuration, token usage, and context capacity. Remaining context capacity measures how much fits in the session; it is not your remaining subscription allowance.

Avoid crossing a reset during a measurement. Pause unrelated work drawing from the same pool, including automations. If you cannot isolate that activity, label the change as combined usage rather than attributing it entirely to Codex.

For a hypothetical sample, suppose one window falls from 80% to 68% after three accepted tasks, including retries:

Average allowance used = (80 - 68) / 3 = 4 percentage points per task
Planning reserve = 12 percentage points
Estimated similar tasks remaining = floor((68 - 12) / 4) = 14

This is a planning estimate, not a quota or benchmark. Percentage rounding, context growth, and harder tasks can change the result. Calculate each applicable window separately and use the smaller estimate.

Save this baseline and repeat the measurement after your account transitions. Current percentages cannot reveal the unpublished December token budget.

Which settings make your allowance last longer?

Use Standard speed for routine work and increase model capability or reasoning only when the task benefits. Judge the result by allowance per accepted task, including failed attempts.

Start with these candidates from OpenAI’s model guidance:

GPT-6.1 Sol: start with your client’s default reasoning effort for regular coding work.
GPT-6 Luna: try it for focused, repeatable tasks with clear checks.
GPT-6 Astra: use it for difficult debugging or unfamiliar problems that cheaper attempts fail to resolve.

Astra’s Light setting maps to low in configuration. Higher effort can consume more tokens without improving every result. Before increasing it, check whether the agent has the necessary files, instructions, and permissions.

For a fixed Sol comparison, run this in Bash from your project directory:

codex --model gpt-6.1-sol --config 'model_reasoning_effort="medium"'

This selects a model and reasoning setting for that launch. It does not select Standard speed. Use /status to verify the session, and check the speed control separately.

The configuration guide documents these overrides and personal defaults in ~/.codex/config.toml. If Sol is unavailable, use /model to select an available model and supported effort. A model ID in a command does not grant access.

Check speed before reducing useful reasoning

Returning to Standard removes a documented usage multiplier without requiring a different model. As of 2026-11-02, OpenAI’s speed documentation gives these rates relative to Standard for the same model:

Speed	Included subscription allowance	Purchased credits
Standard	1×	1×
Fast, where supported	2.5×	2×
GPT-6 Astra or GPT-6.1 Sol Ultrafast, where eligible	8×	6×

These describe consumption rates, not guaranteed speed improvements. Ultrafast currently requires Pro 500 or an eligible Enterprise/Edu plan. These current rates do not establish December’s rules.

In the CLI, /fast toggles Fast mode. If Fast is already enabled, use it to switch off; check the resulting state. Running the command blindly can enable Fast.

For routine ChatGPT questions, choose Instant when it provides sufficient answers. The December exemption specifically names Instant chat. Recheck the selected mode when you move into reasoning or agent work.

Also narrow the task: identify relevant files, define the required output, and avoid unrelated repository-wide cleanup. Compare lower effort with the current setting on similar tasks. If extra retries erase the savings, restore the stronger setting. Our Sol and Astra comparison helps choose that next trial.

What should you do when the allowance runs out?

Choose the route that matches how often you are blocked and which features you need. Switching models does not refill an exhausted shared pool.

Wait for the reset when the deadline allows it. Check the exhausted window’s actual reset time. A refreshed five-hour window will not resolve an exhausted weekly allowance where both apply.

Buy ChatGPT credits for occasional overflow when your account offers them. The personal-plan credits guide says included usage is consumed first, followed by credits. Check Settings → Usage or Codex Usage & Billing. Credits cover eligible features; they do not increase every ChatGPT feature’s limit. Business users need available workspace credits and spending controls that permit their use.

Upgrade when repeated overflow costs justify a higher included allowance. Compare the actual upgrade charge with a representative month’s overflow spending. Check what your account’s upgrade offers before paying; the December notice gives no numerical capacity increase to use in that calculation.

Use an API key when local coding can run against a separate usage bill. This changes the billing source and introduces the API’s own balance and rate limits. It does not restore ChatGPT allowance.

How do you switch local Codex to API billing?

Use API-key authentication for local Codex CLI work after confirming model access and API funding. OpenAI’s authentication guide documents this sign-in route. Codex Cloud requires ChatGPT sign-in.

For an installed CLI configured for the default OpenAI provider, collect your own OpenAI API key in Bash:

read -rs OPENAI_API_KEY && export OPENAI_API_KEY
codex login --with-api-key <<< "$OPENAI_API_KEY"
codex login status
codex --model gpt-6.1-sol

Paste the key when read waits; it will not echo. The here-string supplies it to the documented stdin interface. codex login status reports the authentication mode.

Run one small task, inspect its output, and check API Platform usage before moving a long job. If access fails, check the key’s organization, model access, balance, and rate limits. The CLI and IDE extension share cached authentication, so changing the stored login can affect your next extension session.

Keep three balances distinct: ChatGPT continuation credits, purchased API balance, and promotional API credits included with eligible subscriptions. OpenAI’s subscriber API-credit rules exclude Codex sessions authenticated with API keys. See our API-credit eligibility and exclusions guide before counting that grant toward overflow.

BetterToken also lists gpt-6.1-sol and gpt-6-astra as of 2026-11-02. Its Codex guide uses https://www.bettertoken.ai/v1 and wire_api = "responses" with a BetterToken key. This is separately billed API access.

How do credits, an upgrade, and API costs compare?

Compare the same completed workload using each route’s actual billing rules. Current API prices cannot reveal how many tokens a ChatGPT subscription includes.

The following is an illustrative model-token estimate, not an executed task. Assume Standard global processing and these aggregate quantities across requests:

0.6M ordinary input tokens, 0.3M cached input tokens, and 0.1M input tokens written to a 30-minute cache.
0.2M output tokens, including reasoning tokens.
Every request has at most 272K input tokens.
No 12-hour cache writes, paid hosted tools, or hosted compute; amounts exclude applicable taxes.

Using OpenAI API rates as of 2026-11-02:

gpt-6.1-sol:
0.6 × $2.00 + 0.3 × $0.10 + 0.1 × $2.50 + 0.2 × $10.00 = $3.48

gpt-6-astra:
0.6 × $10.00 + 0.3 × $1.00 + 0.1 × $12.50 + 0.2 × $50.00 = $17.55

Each quantity is in millions of tokens; each rate is USD per 1M tokens. Cache-write rates replace ordinary input charges for the affected tokens, so do not charge both. Reasoning tokens count as output. The model pages for Sol and Astra specify higher rates above 272K input tokens.

For the same hypothetical token quantities, Codex’s Standard credit rates have no separate cache-write charge:

GPT-6.1 Sol:
(0.6 + 0.1) × 50 + 0.3 × 2.5 + 0.2 × 250 = 85.75 credits

GPT-6 Astra:
(0.6 + 0.1) × 250 + 0.3 × 25 + 0.2 × 1,250 = 432.5 credits

Convert the credit totals using your checkout’s USD price per credit. Keep consumption cost separate from cash required for a minimum credit pack. Add applicable taxes and mandatory fees to each route, then compare with the incremental upgrade charge.

An upgrade also buys other features, while API access has different feature boundaries. Identical token quantities do not establish identical completed work. No route wins on price from these figures alone.

Before December, record a normal week, choose a tested default model and speed, and set an overflow budget. After the transition, verify the new windows and repeat the same sample before changing your subscription.