Codex Usage
Back to Codex Limits/GPT-5.3-Codex-Spark
Last updated 06 Sep 2026

GPT-5.3-Codex-Spark Usage Limits

GPT-5.3-Codex-Spark operates on its own dedicated rate-limit tier and does not count against your standard Codex 5-hour or weekly limits. Available as a research preview exclusively for ChatGPT Pro subscribers, Spark runs on specialized Cerebras Wafer-Scale Engine 3 hardware delivering over 1,000 tokens per second. OpenAI has not published a fixed numeric message or token quota; instead, limits adjust dynamically based on live infrastructure demand, with temporary queuing during peak global traffic.

Allowance Pool
Separate Preview Bucket

Independent rate limit; does not deplete standard 5-hour or weekly Codex quotas.

Plan Eligibility
ChatGPT Pro Only

Research preview on Codex app, CLI, and VS Code. Plus and Business excluded.

Quota Type
Dynamic & Demand-Based

No published fixed cap. Capacity may queue or throttle during peak traffic.

Context Window
128k Tokens (Text Only)

Optimized for tight, iterative pairing rather than long-context autonomous agents.

Spark vs. Standard Codex Quotas vs. Astra vs. Luna Reserve

OpenAI employs different quota architectures across model families. While GPT-6 Astra draws from your unified subscription allowance and Luna Reserve provides emergency fallback capacity, GPT-5.3-Codex-Spark is an independent preview tier.

Last updated: 06 Sep 2026. Sourced directly from official OpenAI research announcements and Codex product documentation.
DimensionStandard Codex QuotaGPT-5.3-Codex-SparkGPT-6 AstraLuna ReserveSource
Allowance PoolShared subscription pool (5-hour rolling + weekly window)Dedicated, separate preview rate-limit bucketDraws directly from existing Work & Codex allowance poolSeparate fallback reserve bucket (activates only when regular hits 0%)Official
Drawdown on Normal Codex QuotaYes — depletes both 5-hour and weekly allowancesNo — operates independently; does not draw from standard limitsYes — consumes standard 5-hour and weekly limits at an accelerated rateNo — consumes only the dedicated gpt-reserve fallback poolOfficial
Plan AvailabilityFree, Plus ($20/mo), Pro ($100 & $200), Business, EnterpriseChatGPT Pro only (research preview; select API design partners)Plus & Business Standard (limited); Pro & Business Premium (full)Select personal Plus and Pro accounts only (A/B rollout)Official
Published 5-Hour LimitsEst. 5–2,000 local msgs (Plus); 5x–20x on ProNot published by OpenAI (dynamic demand-based limits)Est. 5–45 local msgs (Plus); 25–225 (Pro $100); 100–900 (Pro $200)Not published by OpenAI (dynamic fallback quota)Official (estimate) / Not published
Published Weekly WindowNot published by OpenAI for any subscription tierNot published by OpenAINot published by OpenAINot published by OpenAINot published
Underlying InfrastructureNot specified by OpenAICerebras Wafer-Scale Engine 3 (WSE-3)Not specified by OpenAINot specified by OpenAIOfficial (Spark) / Not specified
Behavior at Limit / Peak LoadPrompts pause until 5-hour recovery or weekly reset anchorTemporary queuing or limited access during high demandActive turn completes subject to fair use; subsequent prompts pauseReturns to paused state until regular quota resets or credits appliedOfficial

How Spark Rate Limits and Allowances Work

Understanding Spark requires distinguishing how OpenAI provisions preview capacity versus how standard subscription rate limits are calculated:

1. Independent Rate-Limit Tier

During the research preview, your Spark interactions do not deduct turns, tokens, or messages from your primary rolling 5-hour window or weekly subscription allowance. Running heavy Spark sessions leaves your standard allowance intact for frontier reasoning models like GPT-5.6 Sol or GPT-6 Astra.

2. Dynamic Capacity & Queuing

Because Codex-Spark runs on Cerebras Wafer-Scale Engine 3 hardware, OpenAI documents that access can be queued or dynamically adjusted during periods of high global demand.

3. 128k Token Context Window

Spark operates with a 128k token context window and is strictly text-only. While generation latency is near-instantaneous (1,000+ tokens/sec), feeding large repository maps, memory markdown files, or multi-megabyte log dumps can chew through active context rapidly.

4. Interactive Pairing Working Style

Unlike autonomous agents designed for long multi-step background refactors, Spark is tuned for rapid, localized edits. By default, it does not auto-execute build test suites unless explicitly directed, preserving inference bandwidth for fast back-and-forth loops.

What OpenAI Has Not Published About Spark

In accordance with our factual accuracy standards, we explicitly separate verified OpenAI product statements from unconfirmed assumptions:

1. No Fixed Numeric Message Cap:OpenAI publishes estimated 5-hour ranges for standard models (e.g. 5–2,000 local messages on Plus, scaled 5x or 20x on Pro). OpenAI has not published any numeric message or token ceiling for Spark.
2. No Guaranteed Weekly Allocation:OpenAI does not document a fixed weekly volume or reset anchor specific to Spark. Its availability is framed as dynamic and subject to infrastructure demand.
3. Research Preview Terms Are Temporary:OpenAI explicitly states that separate rate limits apply “during the research preview.” OpenAI has not stated whether Spark will remain on a separate limit, be folded into the unified Work & Codex pool, or carry separate billing once the preview concludes.
4. Meter Presence Does Not Guarantee Immediate Availability:Seeing a Spark status line in your CLI does not guarantee instantaneous execution during traffic spikes; server-side capacity queuing can still occur.

How Spark Appears in Codex Telemetry

In the Codex CLI, running /status displays your account telemetry directly from OpenAI endpoints:

codex status
OpenAI Codex Quota Status
▸ 5-Hour Window: 68% remaining (resets in 2h 14m)
▸ Weekly Window: 41% remaining (resets in 3d 11h)
▸ GPT-5.3-Codex-Spark: Active (Separate Preview Tier)
Account allowance status. Spark runs independently of primary and secondary reset windows.

Because Spark is metered on an isolated hardware pool, your primary 5-hour and secondary weekly percentages will not decrement when you send prompts to Spark.

Troubleshooting Spark Quota & Client Inconsistencies

Heavy users frequently encounter interface friction where client applications handle multiple quota meters inconsistently.

Reported Client Behaviors vs. Official Product Policy

The troubleshooting scenarios below describe observed software bugs and telemetry quirks reported by developers in the official openai/codex repository and community forums. They reflect client-side edge cases rather than intended OpenAI policy.

1. Desktop App Blocks Spark When Standard Quota Hits 0%

Reported Symptom: When a user exhausts their regular 5-hour rolling limit or weekly Codex quota, the Codex Desktop app composer input field disables completely or throws a global rate-limit modal. This occurs even when the model selector is switched to Spark and Spark preview quota remains 100% available.

Diagnosis & Workaround: This is an interface-level issue reported in the desktop client where the composer checks primary account quota before evaluating model-specific allowances. There are reported cases where CLI access still worked while Desktop blocked Spark, but this is not guaranteed and should be treated as a reported client bug, not expected behavior:

codex -m gpt-5.3-codex-spark
2. Spark Missing or Grayed Out in Model Selector

Diagnosis: Spark is strictly restricted to active ChatGPT Pro subscriptions. If you downgraded, hold a ChatGPT Plus seat, or are working within a ChatGPT Business workspace, Spark will not appear in the model selector. If you are on Pro and it disappears temporarily, this may indicate temporary capacity limits during periods of high demand.

3. High Demand Queuing vs. Account Quota Limits

Diagnosis: OpenAI documents that temporary queueing or limited access can occur during periods of high demand, separate from the published account quota rules. A queue delay or high-demand notice reflects dynamic demand management rather than personal 5-hour or weekly allowance exhaustion.

4. Stale Telemetry Decrements in /status

Reported Symptom: Users on GitHub (e.g. Issue #33216) have reported instances where the local CLI status telemetry lags behind actual consumption, occasionally showing 100% capacity even after heavy sessions, or suddenly catching up in large steps. If your prompt pauses unexpectedly, verify your connectivity and session authentication rather than relying solely on cached status output.

Frequently Asked Questions

What is the GPT-5.3-Codex-Spark usage limit?

OpenAI has not published a fixed universal numeric prompt or token cap for GPT-5.3-Codex-Spark. Instead, OpenAI documents that Spark operates under a separate rate limit during its research preview, and that availability and limits may dynamically adjust based on live infrastructure demand. Check the /status command in the Codex CLI to view your active account telemetry.

Does Codex Spark have its own separate usage quota?

Yes. OpenAI officially documents that during the research preview, GPT-5.3-Codex-Spark operates on its own dedicated rate-limit tier. It has an independent usage bucket rather than sharing the general token or message allocation used by standard models like GPT-5.6 Sol or Terra.

Does using Spark count against my regular 5-hour or weekly Codex limit?

No. According to official OpenAI documentation, Spark usage does not count against your standard Codex 5-hour rolling limits or weekly subscription quotas during the research preview. You can run Spark without depleting the primary allowance reserved for flagship frontier models.

Can I use GPT-5.3-Codex-Spark on ChatGPT Plus or ChatGPT Business?

No. During the research preview phase, OpenAI documents that GPT-5.3-Codex-Spark is available exclusively to ChatGPT Pro subscribers across the Codex app, CLI, and VS Code extension (alongside select API design partners). ChatGPT Plus, Business Standard, Business Premium, and Enterprise accounts do not currently have access.

Why does the Codex Desktop app block me when I still have Spark quota?

This is a reported client-side software bug documented in developer communities and the openai/codex repository. In the Desktop client, exhausting your standard 5-hour or weekly quota can trigger an interface-level composer freeze, blocking prompt entry even if your separate Spark meter remains at 100%. There are reported cases where CLI access still worked while Desktop blocked Spark, but this is not guaranteed and should be treated as a reported client bug, not expected behavior.

Why does /status show separate Spark meters?

The Codex CLI /status command pulls account telemetry directly from OpenAI endpoints. Because Spark is governed by an independent rate-limit bucket rather than your unified plan allowance, the client reports a distinct Spark status line alongside your primary (5-hour) and secondary (weekly) meters.

What happens when demand for Codex Spark is high?

OpenAI documents that access can be queued or temporarily limited during periods of high demand, separate from the published account quota rules.

Will Codex Spark remain separate from regular limits permanently?

OpenAI has only committed to separate rate limits 'during the research preview.' OpenAI has not published pricing, quota allocations, or pooling policies for when Spark graduates from preview into general availability.

Comparing all tier allowances? Review the complete Codex usage limits table across Free, Plus, Pro, Business, and Enterprise.

Subscribed to ChatGPT Pro? See our detailed guide to ChatGPT Pro Codex limits covering the 5x ($100/mo) and 20x ($200/mo) multipliers on standard models.

Using GPT-6 Astra? Read how GPT-6 Astra Codex limits draw directly from your unified Work & Codex allowance pool.

Exhausted regular quota and seeing fallback capacity? Learn how Codex Luna Reserve provides automated GPT-5.6 Luna fallback access.

Confused about recovery timing? Read when your Codex usage resets.

Hit your limit? Compare Codex credits, paid instant resets, banked resets, and API billing.

Troubleshooting throttled sessions or 429 errors? See why am I rate limited in Codex.