Codex 0.153.3 Adds Native GPT 6 Astra on Bedrock

Ethan Park

If you were running Codex inside a strict enterprise AWS environment, you spent the last few weeks watching everyone else get these wild speedups with GPT 6 Astra while your team was, um, kind of stuck routing around legacy Bedrock model lists.

Maya

Yeah, I saw so many dev leads writing these sketchy custom proxy rewrites just to trick Bedrock into accepting Astra calls. But before we break down how version 0.153.3 fixes that, quick shoutout to Jellypod for powering our daily changelog feed.

Ethan Park

Right, so pull request 42805 in the 0.153.3 release is the big one here. It natively adds GPT 6 Astra right into the Amazon Bedrock model catalogs.

Maya

And that covers both Mantle and the global US routes, right? Like us dot amazon dot bedrock?

Ethan Park

Exactly, both runtime and global routes. So you don't need those proxy hacks anymore. You literally just open up your config file in your home folder, under dot codex slash config dot toml, and set model equal to bedrock slash gpt six astra.

Maya

And as long as your AWS IAM credentials and regional route parameters are already set in your environment, Codex handles the whole handshaking process under the hood. But, uh, wait, there was a nasty little catch with Bedrock payload formats, right?

Ethan Park

Yeah, pull request 42809 addresses that exact issue. When Astra tried to ask clarification questions using request user input async over Bedrock, it kept crashing if the prompt contained non text payload components.

Maya

Classic edge case. Bedrock's endpoint expected pure text, but the terminal agent was trying to pass structured objects or attachments in the async request. The hotfix forces strict text only input validation now, so your background agent doesn't blow up mid conversation when asking for clarification.

Ethan Park

That actually ties right into another update from 0.153.1, pull request 42605. It introduces API level model overrides for headless jobs.

Maya

Wait, explain how that works for a dev team in practice. Why not just set the default model globally?

Ethan Park

Well, say a platform engineer wants all automated CI CD pipelines to execute on Astra for raw speed, but they don't want to overwrite what individual developers see in their local TUI model picker. Now you can set Astra via the API parameter directly, completely invisible to the local UI defaults.

Maya

Ah, so you don't have devs complaining that their local terminal picker randomly switched models overnight!

Ethan Park

Exactly. And speaking of the TUI, pull request 42632 in version 0.153.2 fixed a hilarious cosmetic bug in that picker. It was labeling the GPT 6 Astra Fast tier as one point five times speed instead of two times speed.

Maya

Just a tiny fifty percent speedup miscount on the label, no big deal. The actual execution speed didn't change, just the display text.

Ethan Park

Right. Now, if you want to clean up your terminal output even further, 0.153.0 added a really handy config flag. You can set tui dot auto recap equal to false in your config file.

Maya

Oh, I love that one! Automatic recaps after long turns can get so noisy when you are deep in a flow state. Setting that to false turns off the automatic text dump, but you can still run slash recap manually whenever you actually want a context summary.

Ethan Park

It gives you back control over your scrollback. But, um, probably the most crucial practical change for anyone on Plus or Team plans is the token warning trigger in 0.153.0.

Maya

Is that the five hour usage window monitor?

Ethan Park

Yes. Previously, you would just hit a brick wall when your five hour limit ran out mid refactor. Now, as soon as your remaining allowance dips below fifty percent within that five hour window, Codex throws an early warning in the terminal.

Maya

That is huge. If you are about to launch a massive multi file architectural rewrite and you see that warning, you know to pause, commit your work, or swap to a lighter model before your rate limit locks you out mid turn.

Ethan Park

Exactly. Keeps you from getting stranded with half written diffs. And that wraps up our quick look at the latest Codex updates!