Pro $200 is available for new subscriptions with a lower allowance for non-grandfathered accounts; OpenAI has not published its replacement multiplier. Eligible grandfathered subscribers keep the previous allowance through October 29, 2026 while subscribed, then move to the lower allowance. Earlier published Pro 5x and Pro 20x message estimates are historical evidence, not current five-hour caps or proof of current Pro $200 entitlement. Pro $100, $200, and $500 currently have no five-hour usage limit in Work and Codex, but still have included usage allowances. Pro $500 has the highest included Pro usage and Astra Ultrafast, with no published multiplier or model-specific message estimate. OpenAI Pro tiers.
OpenAI deprecated GPT-5.3-Codex-Spark on September 14, 2026. The research preview is no longer available in the ChatGPT desktop app, Codex CLI, or IDE extension. OpenAI advises updating saved configurations, custom agents, and scripts that select gpt-5.3-codex-spark to use a supported recommended model. For faster responses with a supported model, OpenAI points to Fast mode.
OpenAI deprecated GPT-5.3-Codex-Spark on September 14, 2026, removing the model from the ChatGPT desktop app, Codex CLI, and IDE extension. During its research preview, Spark operated under a dedicated, separate rate-limit tier that did not draw from standard 5-hour rolling or weekly subscription quotas. If Spark is missing from your model picker or your CLI commands fail, the research preview has ended. Below is the verified record of how Spark's separate quota worked, what OpenAI documented about its limits, and how to transition your workflows.
Removed from ChatGPT desktop app, Codex CLI, and IDE extension.
Operated independently; never drew from standard 5-hour or weekly allowances.
The Codex app, CLI, and IDE preview was offered to ChatGPT Pro users; OpenAI separately provided limited API access to a small set of design partners.
Powered by Cerebras WSE-3 hardware with dynamic queuing during high demand.
Recent reports in the official openai/codex issue tracker show that some accounts may continue to display Spark quota after the September 14 deprecation. OpenAI's changelog is authoritative on availability: Spark is no longer available in the ChatGPT desktop app, Codex CLI, or IDE extension. A remaining usage meter does not establish that the model is still usable.
Spark vs. Standard Codex Quotas vs. Astra vs. Luna Reserve
These models represent distinct metering tiers. Astra draws from your unified subscription allowance, Luna Reserve provides emergency fallback capacity, and Spark ran as an independent preview tier prior to its deprecation:
| Dimension | Standard Codex Quota | GPT-5.3-Codex-Spark (Deprecated Preview) | GPT-6 Astra | Luna Reserve | Source |
|---|---|---|---|---|---|
| Where it draws from | Shared subscription pool (5-hour rolling + weekly window) | Dedicated, separate preview rate-limit bucket (when active) | Draws directly from existing Work & Codex allowance pool | Separate fallback reserve bucket (activates only when regular hits 0%) | Official |
| Drawdown on Normal Codex Quota | Yes — depletes both 5-hour and weekly allowances | No — operated independently; did not draw from standard limits | Yes — consumes standard 5-hour and weekly limits at an accelerated rate | No — consumes only the dedicated gpt-reserve fallback pool | Official |
| Plan Availability | Free, Plus ($20/mo), Pro ($100, $200, or $500), Business, Enterprise | Formerly ChatGPT 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 Limits | Est. 5–3,000 (Codex Usage summary of OpenAI's model-specific estimates) local msgs (Plus); Pro allowance varies by tier and grandfathering | Not published by OpenAI (dynamic demand-based limits during preview) | Est. 5–45 local msgs (Plus); 25–225 (Pro $100); 100–900 (previous Pro 20x estimate) | Not published by OpenAI (dynamic fallback quota) | Official model estimates / Codex Usage summary / Not published |
| Published Weekly Window | Not published by OpenAI for any subscription tier | Not published by OpenAI | Not published by OpenAI | Not published by OpenAI | Not published |
| Underlying Infrastructure | Not specified by OpenAI | Cerebras Wafer-Scale Engine 3 (WSE-3) | Not specified by OpenAI | Not specified by OpenAI | Official (Spark) / Not specified |
| Behavior at Limit / Peak Load | Prompts pause until 5-hour recovery or weekly reset anchor | Access could be limited or temporarily queued during high demand | Active turn completes subject to fair use; subsequent prompts pause | Returns to paused state until regular quota resets or credits applied | Official |
How Spark's Separate Quota Bucket Worked
During its research preview, OpenAI provisioned Spark on dedicated infrastructure with specific rate-limiting rules:
During the research preview, Spark interactions did not deduct turns, tokens, or messages from your primary rolling 5-hour window or weekly subscription allowance. Running heavy Spark sessions left your standard allowance intact for models like GPT-5.6 Sol or GPT-6 Astra.
Because Spark ran on Cerebras Wafer-Scale Engine 3 hardware, OpenAI documented that during periods of high demand, Spark access could be limited or temporarily queued, and its separate preview rate limit could vary with demand.
Spark operated with a 128k token context window and was text-only. OpenAI designed Spark for real-time coding, rapid iteration, and targeted edits, delivering generation speeds exceeding 1,000 tokens per second.
OpenAI oriented Spark toward real-time coding and interactive loops, focusing on quick turnarounds during iterative development rather than long-running background tasks.
What OpenAI Published (and Did Not Disclose) About Spark Limits
Official documentation established clear boundaries on what was and was not documented during the preview:
What to Do Now: Migration, Recommended Models & Fast Mode
If you previously used Spark for fast coding loops, adjust your configuration and quota expectations as follows:
Remove references to gpt-5.3-codex-spark across your local workflows. If you have custom agents, CI scripts, or configuration files specifying --model gpt-5.3-codex-spark, update them to use a supported recommended model. Because Spark is no longer available in the Codex CLI, commands specifying it will fail.
Select one of OpenAI's supported recommended models (see our guide to Codex recommended models). Unlike Spark's dedicated preview tier, supported models draw directly from your standard 5-hour rolling and weekly subscription allowances. See the Codex usage limits table for current plan and model allowances.
If you relied on Spark for low-latency coding loops, OpenAI directs users to Fast mode with a supported model. Fast mode provides faster responses on supported models in exchange for higher credit consumption. With ChatGPT sign-in, Fast mode is a ChatGPT credit feature and consumes credits at the applicable higher rate. With an API key, API token pricing applies instead and ChatGPT credit multipliers do not apply.
OpenAI documents Fast mode as a higher-credit-consumption speed setting on supported models, not as a separate Spark-style research-preview rate-limit tier. See our Fast mode guide for details on configuration and consumption.
Frequently Asked Questions
Why is GPT-5.3-Codex-Spark missing from my Codex model selector?
OpenAI deprecated GPT-5.3-Codex-Spark on September 14, 2026, and removed it from the ChatGPT desktop app, Codex CLI, and IDE extension. If it no longer appears in your model picker, the research preview has ended.
What were the usage limits for GPT-5.3-Codex-Spark?
OpenAI never published a fixed numeric message or token cap for Spark. Instead, it operated under a dedicated rate-limit bucket during its research preview, with access that could be limited or temporarily queued during periods of high demand.
Did using Spark count against my regular 5-hour or weekly Codex limits?
No. During the research preview, Spark ran on its own rate-limit tier. Prompts sent to Spark did not deplete your primary 5-hour rolling window or your weekly subscription quota.
Why did OpenAI deprecate GPT-5.3-Codex-Spark?
OpenAI did not state a reason in the September 14 deprecation notice. It confirmed that Spark was deprecated and instructed users to update configurations that select it, choose a supported recommended model, and use Fast mode with a supported model when speed is the goal.
What should I replace Spark with for fast interactive coding?
OpenAI recommends switching to a supported recommended model and enabling Fast mode if you require lower response latency. Be aware that supported models draw against your standard 5-hour and weekly allowances, and Fast mode applies a higher credit consumption multiplier.
Does Fast mode use a separate rate-limit bucket like Spark did?
No. OpenAI documents Fast mode as a higher-credit-consumption speed setting on supported models, not as a separate Spark-style research-preview rate-limit tier. With ChatGPT sign-in, Fast mode is a ChatGPT credit feature and consumes credits at the applicable higher rate. With an API key, API token pricing applies instead and ChatGPT credit multipliers do not apply.
Is GPT-5.3-Codex-Spark still available via the API?
OpenAI's September 14 deprecation notice specifically says Spark is no longer available in the ChatGPT desktop app, Codex CLI, or IDE extension. The notice does not state the current status of the limited API design-partner access associated with the original research preview. Check current OpenAI API documentation before relying on Spark in an API workflow.
- Codex usage limits table — Free, Plus, Pro, Business, and Enterprise allowances.
- ChatGPT Pro Codex limits — current Pro allowances and the legacy Pro 20x estimates.
- GPT-6 Astra Codex limits — how Astra draws from your unified Work & Codex allowance pool.
- Codex recommended models — understanding recommended model presets in client interfaces.
- Fast mode guide — low-latency response setting and credit multiplier consumption.
- When your Codex usage resets — understanding 5-hour and weekly recovery windows.
- Codex limit change ledger — dated history of Codex quota and policy updates, including the September 14 Spark deprecation.