Codex Image Generation: Limits, Credits, and How It Works
Codex supports native image generation powered by OpenAI's Images 2.5, but it operates under fundamentally different rules than standard ChatGPT chat. Most notably, standard ChatGPT image-limit banners (such as “50 images in the last day” or 720-hour reset cooldowns) do not apply to Codex.
- Images 2.5 Availability: Images 2.5 is available inside Codex for creating and editing visual assets, UI wireframes, and design diagrams directly from coding prompts.
- Independent from ChatGPT Limits: OpenAI documents that ordinary ChatGPT image generation caps and banners do not govern Codex. However, this does not mean Codex image generation is unlimited.
- Usage & Credit Metering: Images may not show up in the usage your chat displays, and can still consume credits and count toward your limits. That is what OpenAI documents for Enterprise/Edu workspaces; it documents no equivalent per-image reporting rule for personal plans.
- Undisclosed Deductions: OpenAI does not publish a universal per-image deduction rule or fixed token-per-image formula for personal Plus and Pro 5-hour and weekly meters.
- No API Key Required for Built-In Mode: Normal Codex image generation uses the built-in
image_gencapability and does not require an OpenAI API key.
Images 2.5 in Codex
What changed with the September 8, 2026 release and how it integrates into coding environments.
On September 8, 2026, OpenAI announced ChatGPT Images 2.5, making it available to ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web. Separately, the official Codex repository maintains the built-in imagegen skill for visual workflows.
Generation latency is reduced by up to 50% compared with Images 2.0, shortening turn latency when agents produce visual assets during code execution.
Sharper details, more natural lighting, richer textures, better reference-subject preservation, and improved handling of complex layouts.
More reliable precise editing and better consistency across multiple sequential edits when modifying existing visual assets.
In its announcement, OpenAI introduced GPT-Image-2.5 Flare (positioned as the default API choice for most applications) and GPT-Image-2.5 Sunburst (positioned for more precise premium visual workflows) as developer API models on the OpenAI Platform. OpenAI states that Images 2.5 is available in Codex, while Flare and Sunburst are documented as API models; OpenAI does not document a Codex selector for choosing between Flare and Sunburst.
How to Generate an Image in Codex
Practical built-in imagegen workflow from the terminal CLI, desktop app, or IDE.
No flags, no API setup. Codex has a built-in imagegen capability that kicks in when you describe a visual task in plain English.
1. Generating New Visual Assets
Ask Codex to create an asset describing the visual style, dimensions, and theme:
2. Editing an Existing Image
Pass an existing image path in your repository and specify targeted adjustments:
The assistant interprets your request, calls the built-in image_gen tool, writes the generated asset to disk, and presents the resulting file path in the conversation.
Do ChatGPT Image Limits Apply to Codex?
Understanding why ChatGPT image limit banners do not dictate Codex availability.
No. Standard ChatGPT image generation limits and banners do not apply to Codex.
In its official documentation on Using Codex with your ChatGPT plan , OpenAI explicitly clarifies:
“ChatGPT file uploads, image generation, or voice have separate usage limits, reset periods, and banners. For example, if you see banners like ‘50 images in the last day’, ‘resets in 720 hours’, reach voice caps, or see upload-limit banners while uploading files in ChatGPT, those limits do not apply to Codex.”
The fact that ChatGPT image limit banners do not apply to Codex does not mean Codex image generation is unlimited. OpenAI separately documents that image-generation activity may still consume credits or count toward applicable Codex usage limits, but it does not publish a universal per-image deduction rule for personal 5-hour or weekly meters.
| Dimension | Standard ChatGPT Web/App | Codex (CLI, Desktop & IDE) |
|---|---|---|
| Image Generation Cap Banner | Enforces fixed count/time banners (e.g. "50 images in the last day" or 720h reset) | Banners do not apply to Codex; activity may still count toward applicable Codex usage limits or credits, but OpenAI does not publish a universal per-image deduction rule |
| Underlying Image Model | Images 2.5 interactive web/app experience | Images 2.5 available across tiers on desktop, mobile, and web; imagegen skill in official repo |
| Usage & Spend Visibility | Standard chat message and tool counts | In Enterprise/Edu documentation, may not appear in displayed chat usage even if consuming credits/limits; no universal personal reporting rule |
| Authentication Requirement | ChatGPT user login session | ChatGPT subscription session for built-in mode; OPENAI_API_KEY only for CLI script fallback |
| Output Location | Rendered inline in browser/app chat | Documented default under $CODEX_HOME/generated_images/...; movable to project workspace |
| Third-Party Cloud Providers | Hosted directly on OpenAI cloud | Image generation is currently unavailable when configured with Amazon Bedrock (OpenAI-hosted API is not in the request path) |
Does Codex Image Generation Use Limits or Credits?
What OpenAI documents across subscription tiers, credit pools, and enterprise workspaces.
Whether image generation consumes your included quota or draws from purchased credits depends on your account tier and current allowance balance:
When working within your subscription allowance, Codex actions execute against your plan's included capacity. Once included usage is exhausted, task execution pauses unless flexible credits or instant resets are used. Learn more in our Codex limits by plan guide.
For applicable workspace credit-based billing, OpenAI documents that image generation can consume credits. On personal plans, OpenAI documents flexible credits as extending supported Codex usage after included limits are reached, but does not publish a separate personal per-image credit rule. See our full breakdown in the Codex credits guide.
OpenAI has not published a universal per-image deduction rule for personal Plus and Pro 5-hour and weekly meters. There is no documented formula specifying that an image equals a fixed number of local messages or tokens on consumer subscriptions. Where credit rate cards publish explicit per-token or per-image rates (such as developer API token pricing or workspace credit charts), those numbers apply strictly to the specific billing surfaces they define, not as guaranteed deductions from personal subscription progress bars.
Why Image Generation May Not Appear in Displayed Codex Usage
Explaining main-chat usage estimates versus separately billed tool activity.
Generate an image in Codex and the usage percentage in your terminal or chat often does not move. OpenAI acknowledges this in its Help Center documentation.
OpenAI Documentation: “Reviewing Work and Codex usage and using Personal Analytics in ChatGPT Desktop”
“Images and certain separately billed tool activity may not appear in the Work or Codex chat's displayed usage.”
What the desktop views and CLI /status show you is an estimate of the main conversation's text and reasoning tokens. Tool calls, image generation included, often run through a separate pipeline. In OpenAI's Enterprise/Edu Work and Codex usage documentation, image generation and certain separately billed tool activity may not appear in displayed chat usage even though omitted activity may still consume credits and count toward usage limits. OpenAI does not currently document an equivalent universal personal-plan per-image reporting rule.
Scope Note: Granular per-chat credit tracking and the @personal-analytics desktop plugin are enterprise features available in ChatGPT Desktop (v26.812.10818+) for eligible Enterprise and Edu workspaces with credit-based billing. Personal subscription accounts do not receive turn-by-turn tool spend telemetry.
Built-in Image Generation vs. API/CLI Fallback
The official Codex imagegen skill defines two separate operational modes.
According to the official openai/imagegen skill specification in the official Codex repository, image tasks operate through two distinct pathways:
image_genAPI Key Required: No (uses active subscription session)
Billing: Account-dependent Codex usage / credit treatment
Default agentic path. The assistant invokes the built-in image_gen capability automatically when prompted to create or edit visual assets.
scripts/image_gen.pyAPI Key Required: Yes (requires OPENAI_API_KEY)
Billing: OpenAI Platform API developer balance
Explicit fallback path bundled in the imagegen skill. Triggered when explicitly requested or running custom standalone batch scripts.
You do not need an API key for ordinary Codex image generation. Your active ChatGPT Plus, Pro, or Business subscription covers built-in tool execution. An OPENAI_API_KEY is only necessary if you explicitly instruct Codex to run manual Python scripts (such as scripts/image_gen.py) or when integrating third-party tools outside the built-in tool container.
Where Codex Saves Generated Images
Default storage locations, cache directories, and workspace asset management.
When Codex generates an image in built-in mode, the official imagegen skill documents an intended local storage policy:
By default, generated images are written under your Codex environment directory:
If $CODEX_HOME is unset, the client defaults to ~/.codex/generated_images/... on your local machine.
While transient previews remain in the central image cache, the skill instructs agents to move or copy project-bound assets (such as application icons, UI wireframes, or documentation figures) directly into your active workspace (e.g. public/assets/) so they are tracked by git.
Persistence Note: Treat default cache paths as documented intended behavior rather than an immutable storage guarantee across every client build or cloud environment. Always ensure final design deliverables are explicitly committed into your project repository.
Configuration Limitations: Amazon Bedrock
When third-party enterprise providers restrict multimodal features.
According to OpenAI's Help Center article on configuring Codex with Amazon Bedrock , image generation is currently unavailable when Codex is configured with Amazon Bedrock (model_provider = "amazon-bedrock" in ~/.codex/config.toml). OpenAI explains that the OpenAI-hosted API is not in the request path when routing through Amazon Bedrock.
Other cloud-dependent features like voice dictation and cloud agent reviews are similarly omitted when routing through Bedrock. If your team requires image generation inside Codex, use standard OpenAI ChatGPT subscription sessions or supported OpenAI API connections.
Frequently Asked Questions
Common questions about Codex image generation.
Can Codex generate images?
Yes. OpenAI documents that Images 2.5 is available to Codex users across all tiers on desktop, mobile, and web, and the official Codex repository provides a built-in imagegen skill for generating and editing visual assets.
Does ChatGPT's image limit apply to Codex?
No. OpenAI explicitly documents that standard ChatGPT image limits, reset periods, and banners (such as "50 images in the last day") do not apply to Codex. OpenAI separately documents that image-generation activity may still consume credits or count toward applicable Codex usage limits, but it does not publish a universal per-image deduction rule for personal 5-hour or weekly meters.
Does image generation count toward Codex usage limits?
In OpenAI's Enterprise/Edu Work and Codex usage documentation, image generation and certain separately billed tool activity may not appear in displayed chat usage even though omitted activity may still consume credits and count toward usage limits. OpenAI does not currently document an equivalent universal personal-plan per-image reporting rule.
Does Codex image generation use credits?
OpenAI explicitly documents credit consumption for image generation in applicable workspace credit-based usage, including supported Business and Enterprise/Edu Work and Codex activity. For personal plans, OpenAI documents that credits can extend supported Codex usage after included limits are reached, but it does not currently publish a separate personal Plus/Pro per-image credit rule.
Why doesn't image generation appear in my Codex usage?
OpenAI documents that displayed usage in Work and Codex chats represents an estimate for main conversation inference and that image generation and certain separately billed tool actions may not appear in displayed chat usage.
Do I need an OpenAI API key to generate images in Codex?
No. Normal built-in image generation in Codex uses the native image_gen tool mode and does not require an OPENAI_API_KEY. An API key is only required if you explicitly invoke the standalone CLI/API script fallback.
Can I choose GPT-Image-2.5 Flare or Sunburst in Codex?
OpenAI currently documents Flare and Sunburst as API models and does not document a Codex selector for choosing between them. Flare is positioned as the default API choice for most applications and Sunburst for premium visual workflows, while Images 2.5 is documented as available to Codex users.
Where does Codex save generated images?
Under the official Codex imagegen skill, generated images are saved locally by default under $CODEX_HOME/*, commonly in $CODEX_HOME/generated_images/... (or ~/.codex/generated_images if $CODEX_HOME is unset). For project assets, the documented practice is to copy or move final files into the project workspace.
Does Codex image generation work with Amazon Bedrock?
No. Image generation is currently unavailable when Codex is configured with Amazon Bedrock because requests bypass the OpenAI-hosted API.
What changed with Images 2.5 in Codex?
Released on September 8, 2026, Images 2.5 brings up to 50% lower generation latency compared with Images 2.0, sharper details, more natural lighting, richer textures, more precise editing, and better consistency across multiple edits.