OpenAI's December change: one shared token pool for ChatGPT and Codex

OpenAI is changing what a ChatGPT subscription buys. Starting in December, Plus, Pro, and Business plans will measure usage in tokens and draw one shared allowance down across both ChatGPT and Codex, according to OpenAI's help center. The change lands in the same month as a deprecation wave that removes six GPT-5 and o3 snapshots from the API on December 11, 2026. For anyone who sized a workload on a flat-rate plan, the unit of measurement just changed from "a month of chatting" to tokens.

Key facts

• Plus, Pro, and Business plans move to a usage allowance shared across ChatGPT and Codex starting in December 2026, measured in tokens (OpenAI Help Center)
• Six model snapshots, gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, and o3-pro, are removed from the API on December 11, 2026; fine-tuned variants follow the same schedule (OpenAI deprecations page)
• Instant chat does not draw from the shared allowance; heavier models and more complex tasks consume more of it (OpenAI Help Center)
• The same deprecations page lists further removals: GPT-4o and GPT-4.1 snapshots on January 14, 2027, tts-1 on January 6, 2027, gpt-image-2 on February 10, 2027, and gpt-5.1 and gpt-5.3-codex on April 1, 2027 (OpenAI deprecations page)
• When the pool runs out, the documented options are wait for a reset, upgrade the plan, or buy credits (OpenAI Help Center)

What happened

The help-center page states the policy plainly: "Starting in December, Plus, Pro, and Business plans' usage allowance will be shared across ChatGPT and Codex as we bring the agentic capabilities of Work into the main ChatGPT experience." Usage will be "measured in tokens, the units of information a model processes and generates," and how much of the allowance a session consumes "will depend on the model you choose and the task you ask it to do."

Two carve-outs matter. Instant chat will not draw from the usage pool at all. And OpenAI says combined remaining usage and reset times will be visible in Settings, then Usage, so subscribers can at least see the meter they are now on. The company also frames the transition defensively: "We expect most users won't need to wait for their usage to reset more often than they do now." That is OpenAI's own expectation, not a measured result, and the practical test arrives when heavy Codex users hit the pooled limit mid-workday for the first time.

The December deprecation wave is a separate but same-month squeeze on the API side. On June 11, 2026, OpenAI notified developers that older GPT-5 and o3 snapshots would be removed on December 11, 2026. The six snapshots cover the main tiers of the previous generation: gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, and o3-pro, with the gpt-5.6 family (sol, terra, luna) listed as the recommended replacements. Fine-tuned versions of these models follow the same shutdown schedule.

The removals do not stop there. The deprecations page shows a rolling schedule through early 2027: GPT-4o and GPT-4.1 snapshots leave on January 14, 2027, the tts-1 text-to-speech model on January 6, the gpt-image-2 model on February 10, and gpt-5.1, gpt-5.3-codex, and gpt-5.4-nano on April 1. OpenAI's stated notice periods are at least six months for generally available models and at least three for specialized and image variants, with preview models retireable on roughly two weeks' notice. Any agent pinning an old snapshot name will start failing on the shutdown date unless it migrates.

Why it matters

The practical definition of a $20-or-$200 ChatGPT subscription is moving from flat-rate usage toward a metered pool, and coding-agent traffic now draws from the same pool as chat. Codex loops are token-hungry: an agent session that reads files, plans, edits, and retries can consume far more tokens in an afternoon than a month of conversation. Folding that traffic into the ChatGPT allowance is the structural change, and it means the answer to "how much does my plan include?" now depends on the model mix and task complexity of the work itself.

For builders, three checks follow from the documentation. First, anyone budgeting against a subscription tier should model usage in tokens rather than sessions and watch the combined meter once it appears. Second, audit which model snapshots your agents pin: anything on the December 11 or January 14 lists needs a migration path now, and the shutdown dates are queryable programmatically through the shutdown_date field on the models endpoint. Third, note that the escape hatches are all paid ones: waiting, upgrading, or buying credits.

The comparison worth making is with billing models that never had a subscription pool to reprice. Per-token API billing and self-hosted open weights charge for what you actually run, so a provider redefining its allowance terms does not change your unit costs. That is not an argument that one route is better; it is a reason the October deprecation wave at hyperscaler endpoints and this allowance change both land hardest on workloads that assumed provider terms were static. Our migration guide off the OpenAI API covers the mechanical side of keeping options open, and self-hosting versus API cost breaks down when owning the hardware pencils out.

Background

Subscription usage limits on ChatGPT have always been metered in some form, but the meter was coarse and mostly invisible: message caps on specific models, reset windows, and little else. Two product moves made metering in tokens the natural unit. Agentic coding pushed per-user consumption up by orders of magnitude, since a Codex session burns tokens continuously rather than one message at a time. And the move to bring "the agentic capabilities of Work into the main ChatGPT experience," as the help center puts it, means one product surface where chat and agent loops coexist, which is awkward to meter with separate allowances.

Provider terms changing under a shipped workload is a recurring pattern in this beat, not a one-off. We tracked the hyperscaler endpoint retirements in October's deprecation wave and Perplexity's chat-completions retirement, both of which hit builders who had assumed endpoint stability. OpenAI's own deprecations page now carries a shutdown-date field in its API so teams can monitor removals programmatically, which is a concession to how often this happens.

The policy change is verified against OpenAI's own documentation. What remains unconfirmed is any specific allowance size or credit price for the new pooled meter; OpenAI has not published those numbers in the pages cited here, and community reports about effective degradation for heavy users are anecdotal. We will update when OpenAI publishes the actual allowance figures.

What's next

December brings both changes at once: the shared allowance takes effect for Plus, Pro, and Business, and the December 11 snapshot removals hit the API. January 14, 2027 removes the GPT-4o and GPT-4.1 snapshots, with gpt-image-2 following on February 10 and the gpt-5.1 and gpt-5.3-codex removals on April 1. The number to watch first is the published allowance size for each plan tier, since that single figure determines whether the pooled meter is a rounding error or a real quota for Codex-heavy workflows.