What is GPT-5.3-Codex?
GPT-5.3-Codex is OpenAI's coding-tuned reasoning model, built to work through software tasks on its own in Codex and similar agent tools. It reads text and images, writes text, holds 400K tokens of context and runs only on the Responses API, with reasoning effort from low to xhigh. OpenAI says it pairs GPT-5.2-Codex's coding with GPT-5.2's reasoning and runs 25% faster in Codex.
It is now a legacy pick. OpenAI has dropped it from Codex for people who sign in with ChatGPT and points most Codex users to GPT-6 Sol instead. It stays on the API, in Microsoft Foundry and in GitHub Copilot, where GitHub made it the long-term support model for business plans.
What changed from GPT-5.2-Codex?
More reasoning, more speed and more updates while it works, at the same API price.
- Reasoning
- OpenAI says it combines GPT-5.2-Codex's coding with the reasoning and professional knowledge of GPT-5.2 in one model.
- Speed
- It runs 25% faster for Codex users, by OpenAI's measure.
- Working style
- It sends more frequent progress updates and takes steering while it works, which needs the Responses API's phase field.
- Price and status
- The API price is unchanged at $1.75 / $14 per million tokens, and GPT-5.2-Codex has since been shut down on the API.
Who is GPT-5.3-Codex best for?
GPT-5.3-Codex is best for GitHub Copilot Business and Enterprise teams who want a stable default, and for API builders whose coding agents already run on it.
GitHub Copilot Business and Enterprise
GitHub made it the base and long-term support model, commits to keeping it for a year, and falls back to it when no other model is enabled.
Custom coding agents on the API
In the Responses API it can use hosted shell, skills, web search and your own functions, so it can run commands and fetch docs while it works.
Agents that resend a large codebase
Cached input costs $0.175 per million tokens, so an agent that sends the same files every turn pays mostly the cached rate.
Matching effort to the task
Reasoning effort runs from low to xhigh, so quick fixes stay cheap and hard refactors get more thinking.
What's included with GPT-5.3-Codex?
Four reasoning effort levels, hosted shell and skills on the Responses API, progress updates while it works, a Fast mode, and cheap cached input.
- Four effort levels
- Low, medium, high and xhigh; OpenAI suggests medium for everyday interactive coding and high or xhigh for the hardest tasks.
- Hosted shell and skills
- On the Responses API it can run commands in a hosted container, load skills, search the web and call your functions.
- Progress updates while it works
- It sends short updates between tool calls, which the Responses API labels with a phase field so your app can tell them from the final answer.
- Fast mode
- Faster, steadier responses at twice the price per token: $3.50 / $28 per million tokens.
- Image input and structured outputs
- It can read screenshots and diagrams alongside code and return JSON that matches a schema you set.
How much does GPT-5.3-Codex cost?
GPT-5.3-Codex costs $1.75 per million input tokens and $14 per million output tokens on the OpenAI API, or $3.50 / $28 in Fast mode.
- Cached input $0.175
- 400K-token context, up to 272K tokens of input
- Reasoning effort from low to xhigh
- Cached input $0.35
- Faster, steadier responses
GPT-5.3-Codex works only through the Responses API; it has no Batch or Flex discount and no Chat Completions support. Tools such as web search are billed on top at their own rates. Codex signed in with an API key bills it at these API rates; it is no longer offered in Codex for ChatGPT sign-ins. Full pricing on developers.openai.com. We checked these prices on .
How do readers rate GPT-5.3-Codex?
No one has rated GPT-5.3-Codex yet. Be the first.
Sign in to rate. One rating per account.
GPT-5.3-Codex pros and cons
Users found it fast and dependable for backend code, but it can ignore detailed instructions, is weak at UI work, and has less context than OpenAI's newer models.
Pros
- Many Codex users found it much faster than GPT-5.2 at about the same quality; OpenAI itself says it runs 25% faster than GPT-5.2-Codex.
- Developers rate it for backend work, refactors and bug fixing, and several went back to it from GPT-5.4 because it did the task without extra back-and-forth.
- Cheaper input than OpenAI's newer full-size coding models: $1.75 per million tokens, against $2 for GPT-6 Sol and $2.50 for GPT-5.4.
- GitHub keeps it in Copilot as the long-term support model, so Copilot Business and Enterprise teams can build around it without a sudden removal.
Cons
- Several users say it acts before it asks and can ignore detailed instructions, and one keeps an AGENTS.md rule telling it to ask when a task is unclear.
- User reports rate it weak at UI work, with plain interfaces, and some hand design touch-ups to Claude or Gemini models.
- OpenAI rates it High for cybersecurity capability, so API traffic flagged as suspicious can pause access for your whole organization unless you send a per-user safety_identifier; at launch, Codex users also complained of silent rerouting to GPT-5.2.
- Narrower than OpenAI's current models: 400K of context against 1.05M for GPT-6 Sol, knowledge that stops at August 2025, and no Batch, Flex or Chat Completions.
What are the best GPT-5.3-Codex alternatives?
Five tools on Recaply do GPT-5.3-Codex's job. GLM-4.7 and Laguna S 2.1 are open source, so you can run them on your own server.
GPT-6 AstraPaidChoose it when the job is the hardest end-to-end work and quality matters more than cost; it is OpenAI's flagship, with a 1.05M-token context, at $10 / $50 per million tokens.
GPT-5.4PaidChoose it when you need a 1.05M-token context, Chat Completions or the Batch API, none of which GPT-5.3-Codex offers; it costs $2.50 / $15 per million tokens.
Kimi K2.7 CodeFreeChoose it when you want open weights you can host yourself or a cheaper API at $0.95 / $4 per million tokens, and can work within a 256K context.
GLM-4.7Open SourceChoose it when coding agents need open weights, or a free Flash tier on Z.ai's API, instead of OpenAI's Codex model.
Laguna S 2.1Open SourceChoose it when you want an Apache-licensed coding model with 8B active parameters and a 1M context to self-host.
Questions about GPT-5.3-Codex
What is GPT-5.3-Codex?
GPT-5.3-Codex is OpenAI's coding-tuned reasoning model for agent tools like Codex, still sold on the API at $1.75 / $14 per million tokens but no longer OpenAI's pick for coding.
Who is GPT-5.3-Codex for?
GPT-5.3-Codex is best for GitHub Copilot Business and Enterprise teams who want a stable default, and for API builders whose coding agents already run on it.
Is GPT-5.3-Codex free?
No. GPT-5.3-Codex is a paid tool, from $1.75 / $14 per million input / output tokens.
How much does GPT-5.3-Codex cost?
GPT-5.3-Codex's paid plans are Standard at $1.75 / $14 per million input / output tokens and Fast mode at $3.50 / $28 per million input / output tokens.
Is there an open source alternative to GPT-5.3-Codex?
Yes. GLM-4.7 and Laguna S 2.1 are open source tools that do the same job, so you can run your own.
What are the best GPT-5.3-Codex alternatives?
The closest alternatives on Recaply are GPT-6 Astra, GPT-5.4, Kimi K2.7 Code, GLM-4.7 and Laguna S 2.1.
Where can I try GPT-5.3-Codex?
On its official website, developers.openai.com.
Can I use GPT-5.3-Codex with a ChatGPT plan?
Not any more. OpenAI has deprecated GPT-5.3-Codex in Codex for people who sign in with ChatGPT and recommends its GPT-6 models there. You can still use it through the OpenAI API, including Codex signed in with an API key, billed per token.
GPT-5.3-Codex vs GPT-6 Sol: which should I use?
For new work, GPT-6 Sol. OpenAI built it for complex coding and agentic work, and it has a 1.05M-token context against 400K, newer training data and cheaper output at $2 / $10 per million tokens. GPT-5.3-Codex only wins on input price, at $1.75.
Is GPT-5.3-Codex available in GitHub Copilot?
Yes. GitHub lists it as generally available and made it the base and long-term support model for Copilot Business and Enterprise, which GitHub commits to keep for a year. Auto model selection also falls back to it when no other model is available.
Why does GPT-5.3-Codex need the phase parameter?
It marks each assistant message as commentary or final_answer so progress updates are not mistaken for the end of the task. OpenAI's Codex prompting guide says your app must pass those items back with phase intact, or performance can drop sharply.
Our verdict
GPT-5.3-Codex is still a capable, low-cost coding model, but new projects should start on GPT-6 Sol.
Users found it fast and dependable for backend work, and its input price undercuts GPT-6 Sol, yet GPT-6 Sol has over twice the context, newer knowledge and cheaper output.
- Skip it for
- New builds, UI-heavy work, or anyone who wants it inside Codex with a ChatGPT plan, where it is no longer offered.
- Check first
- If your agent touches security work, send a per-user safety_identifier, or one flagged request can pause API access for your whole organization.
OpenAI Privacy Filter
GLM-5.3
Rnj-1
KAT-Coder-Pro V2
Composer 2.5
MiMo-V2.5-Pro