Codex is now searching the web way more frequently and at unnecessary times

Resolved 💬 8 comments Opened May 4, 2026 by Terrydaktal Closed May 4, 2026
💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

0.128

What subscription do you have?

Plus

Which model were you using?

gpt-5.3-codex high

What platform is your computer?

linux

What terminal emulator and version are you using (if applicable)?

xfce4-terminal

What issue are you seeing?

A few days ago codex started searching the web way more frequently and at unnecessary and bizarre times where it obviously wouldn't find the answer. It now searches the web 2-10 times frequently for even basic tasks when before it would search 0 times and it would only use the web when logical. It's searching the web for proprietary details and things clearly only invented by me.

What steps can reproduce the bug?

• Explored
└ Read LocalBoardApi.kt
Search robust_loc_scale|sigmoid|solve_shift_for_target_sum|xorshift128plus|generate_flat\(|
kMaxWords|g_dictionary_words|thread_word_indices|list_words in main.cpp
Search dictionary|word_to_id|g_dictionary|vector<string>|all_words|list_words in main.cpp

• Searching the web

• Searched

• Searching the web

• Searched

• Explored
└ Read main.cpp, boardgen_jni.cpp

• Searching the web

• Searched

• I’ve confirmed the sampled EWMA is failing for a concrete reason in your device log. I’m now
checking the native generator code for an existing on-device EWMA routine we can call
directly (instead of desktop path execution).

• Explored
└ Search ewma|total_rating|board-analysis|bot-speed|benchmark_summary|rating in main.cpp

• Searching the web

• Searched

• Searching the web

• Searched

What is the expected behavior?

_No response_

Additional information

_No response_

View original on GitHub ↗

8 Comments

Mahkhmood9 · 2 months ago

I also have issues with gpt-5.3, and I did noticed a lot of web search,

gpt5.4 and gpt-5.4-mini(which I like the most) seems to be the sweet point for me

is that in this specific problem or a wide problem ?
the jump between 5.3 is huge, to 5.4

I recomend trying 5.4 mini - its actually really good , and it knows when to stop better

new issue you had? you been using 5.3 in the past, and now its new, you have issues? has the codebase increased?

Terrydaktal · 2 months ago

@Mahkhmood9 I think it's only happening with 5.3 codex. I don't recall it happening with 5.5. I'll check next time I need to do a hard task. I know 5.4 mini is lighter on limits but I've never used it because i use gemini-cli flash for the basic tasks but it's been having a lot of problems lately so I'm using codex more of the time and I'm hitting limits faster. I'm going to start using it now for basic tasks to see what it can handle

etraut-openai contributor · 2 months ago

Thanks for the bug report. This appears to be model behavior, not a Codex harness issue. We'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the /feedback command. That will provide us with additional details including logs and conversation text.

kovalensue · 2 months ago

I tested it on same task using Codex CLI with gpt-5.3-codex model and Cursor with Codex 5.3 model and I noticed web searches only in CLI. In cursor I walked through all collapsible "Explored" parts and not a single web search.

Are you sure it is a model behavior?

Mahkhmood9 · 2 months ago
I tested it on same task using Codex CLI with gpt-5.3-codex model and Cursor with Codex 5.3 model and I noticed web searches only in CLI. In cursor I walked through all collapsible "Explored" parts and not a single web search. Also in Codex CLI i used other models and searches where also there even though I told the model to not do web search at all. So are you sure it is model behavior?

the decision to do a web search is model based

Is that something new that happening?
It models in my opnion, because

5.3 is much weaker than 5.4 so note that

Personally I can recomend VSCODE and installing Codex extention in it ,, it has a very smooth exprience

Can you say the same for 5.4 /5.4mini or 5.5

in my expirence Codex is smarter on VSCODE / CLI

see https://github.com/openai/codex/blob/main/codex-rs/core/src/state/turn.rs

but honestly I dont know https://docs.rs/codex-protocol/latest/src/codex_protocol/models.rs.html#231-253

and codex_protocol::models
Enum WebSearchAction

which is client side? Unless the model uses the client side logs , and hallucinate?

keep in mind you mention 5.3,
second cursor , can use another model for search?

You can't swap the OpenAI search tool itself, but you can add your own search tools alongside?
and the model-web-search is server side?

so when the model says "do web search" , you cant enact it locally

so they use
In ~/.codex/config.toml:
tomlweb_search = "disabled"

and some type of mcp behvior (wheter if to use )
https://developers.openai.com/codex/mcp

how does it know to respect web search , it self, its llm model , it init the next token right? so something wierd is happening

~/.codex/config.toml
[tools]
web_search = false

far as I concern , it could also be something related to the (premission level) like when the init a web search or not
but there is external_web_access also?

turtle261 · 2 months ago

I've done some investigation. I think the models aim is to determine "Will I have to consider backwards compatibility or not" or something, it is checking against what is public, vs what is local, to determine something. It's determining what the public surface already is.
That's my speculation.
Still ridiculous, because planning out a +100 -150 LOC change, should not take thirty-five web searches. Why do more web searches than uses of grep?

I strongly believe you need to improve the CLI UX now to accomadate the new model behavior @etraut-openai -- maybe collapse web searches, since it is extremely inclined to make 15 in a row. Collapse it to "made 15 web searches".

What a rush to closure.

Mahkhmood9 · 2 months ago
I've done some investigation. I think the models aim is to determine "Will I have to consider backwards compatibility or not" or something, it is checking against what is public, vs what is local, to determine something. It's determining what the public surface already is. That's my speculation. Still ridiculous, because planning out a +100 -150 LOC change, should not take thirty-five web searches. Why do more web searches than uses of grep? I strongly believe you need to improve the CLI UX now to accomadate the new model behavior @etraut-openai -- maybe collapse web searches, since it is extremely inclined to make 15 in a row. Collapse it to "made 15 web searches". What a rush to closure.

Please use the /feedback if you notice it , (they actually read it i think)

turtle261 · 2 months ago

thread ID 019e1510-5645-79c1-946e-dbb5cf1c9c16