Unable to set "gpt-5.3-codex" via /model. | NO API

Same here; I cant ever hardcode it into config.

json

 {
  "error": {
    "message": "The requested model 'gpt-5.3-codex' does not exist.",
    "type": "invalid_request_error",
    "param": "model",
    "code": "model_not_found"
  }
}

@milosgajdos,

Same here; I cant ever hardcode it into config. json `` { "error": { "message": "The requested model 'gpt-5.3-codex' does not exist.", "type": "invalid_request_error", "param": "model", "code": "model_not_found" } } ``

If this is for Codex using API key access, currently API does not have access to gpt-5.3-codex.

Yes, gpt-5.3-codex is not yet available for API Key usage. You need to use a ChatGPT subscription to access it.

@etraut-openai via this comment.

If this is for Codex using API key access, currently API does not have access to gpt-5.3-codex.

It is via API, yes; oof, that's very unfortunate and kinda sucky 😮‍💨

Yes, if you're using the API, this is expected behavior, not a bug. We typically roll out new models first for ChatGPT accounts, and it takes us a little while (typically less than a week) to make them available through the API. Stay tuned.

@etraut-openai, I am NOT accessing via API.

Sorry, didn't ready closely enough. Reopening.

All good, that last thread was a bit of a dumpster fire of many issues. Thanks for the reopen.

Same issue here for me, I think at bottom maybe uninstall reinstall is what fixed it. Regardless here's what I did and now I am able to access 5.3.

1. Log out & uninstall codex
2. Re-install codex
3. Had to fight some issues with my account logging back in with my chatgpt account. My email had changed so after I plugged in the new email I was good to go.

codex --model=gpt-5.3-codex is available

╭───────────────────────────────────────────────────╮
│ >_ OpenAI Codex (v0.98.0)                         │
│                                                   │
│ model:     gpt-5.3-codex xhigh   /model to change │
│ directory: C:\Codex                               │
╰───────────────────────────────────────────────────╯

  Tip: New Build faster with the Codex App. Run 'codex app' or visit https://chatgpt.com/codex

I didn't have any problems, but I've heard from some people that they had similar experiences.

Edit: I also had it on a Windows 11 computer with Codex CLI 0.98. When I started Codex for the third time and jumped to /model, it was finally displayed.

All I can say is that you can force model=gpt-5.3-codex in the API access, and that works, but I don't know if it's actually working or if it's a bug.

I had the same issue, just logout and login again you should get the last model.

Also have the issue

Yes, if you're using the API, this is expected behavior, not a bug. We typically roll out new models first for ChatGPT accounts, and it takes us a little while (typically less than a week) to make them available through the API. Stay tuned.

@etraut-openai could you please clarify this part ⬆️

I am trying the following using the latest codex SDK

import { Codex } from "@openai/codex-sdk";

async function main() {
  const model = process.env.CODEX_MODEL || "gpt-5.3-codex-spark";
  const prompt =
    process.argv[2] ||
    "Write a haiku about TypeScript. Output ONLY the haiku, nothing else.";

  const codex = new Codex({});

  console.log(`Model: ${model}`);
  console.log(`Prompt: ${prompt}`);
  console.log(`---`);

  const thread = codex.startThread({
    model,
    sandboxMode: "read-only",
    skipGitRepoCheck: true,
  });

  const startTime = Date.now();

  const { events } = await thread.runStreamed(prompt);
  for await (const event of events) {
    switch (event.type) {
      case "thread.started":
        console.log(`Thread: ${event.thread_id}`);
        break;
      case "item.completed":
        if (event.item.type === "agent_message") {
          console.log(`\nResponse:\n${event.item.text}`);
        }
        break;
      case "turn.completed":
        console.log(`\nUsage: ${JSON.stringify(event.usage)}`);
        break;
      case "turn.failed":
        console.error(
          `\nFAILED: ${event.error?.message || JSON.stringify(event)}`,
        );
        process.exit(1);
        break;
    }
  }

  const elapsed = ((Date.now() - startTime) / 1000).toFixed(1);
  console.log(`\nDone in ${elapsed}s`);
}

main().catch((err) => {
  console.error(err);
  process.exit(1);
});

which produces

Model: gpt-5.3-codex-spark
Prompt: Write a haiku about TypeScript. Output ONLY the haiku, nothing else.
---
Thread: 019c70dd-a234-7ae1-ac38-924b616a58fa

FAILED: {
  "error": {
    "message": "The requested model 'gpt-5.3-codex-spark' does not exist.",
    "type": "invalid_request_error",
    "param": "model",
    "code": "model_not_found"
  }
}

On the same exact system, using

codex --model gpt-5.3-codex-spark

works just fine. If codex SDK is a wrapper leveraging codex binaries and piggy backing on my auth session, I am expecting the API snippet to work as well. What am I missing?

Now it's possible to set it in the config.toml, but I'm experiencing a strange beavior: the model returns very quickly without any error, doing only the first step of the plan. I cannot get it to work longer. gpt-5.2-codex in the exact same configuration works fine. Strange...
I'm using it via a litellm gateway hosted on prem and linked to an azure subscription where gpt-5.3-codex is hosted (same setup as for gpt-5.2-codex which works fine).

Now it's possible to set it in the config.toml, but I'm experiencing a strange beavior: the model returns very quickly without any error, doing only the first step of the plan. I cannot get it to work longer. gpt-5.2-codex in the exact same configuration works fine. Strange... I'm using it via a litellm gateway hosted on prem and linked to an azure subscription where gpt-5.3-codex is hosted (same setup as for gpt-5.2-codex which works fine).

Thats "Fast" inference by cerebras for you, I had the similar experience with them while directly working using APIs and Kilocode and Qwen 3, the model would seem totally uninterested, in delving deeper, its only priority would always be just providing a response as fast as it could and be done with it :D kinda found it funny, its a shame they haven't fixed it yet.