Are Pro 5.3 users all secretly just using 5.2? Certainly looks like it! ~/.codex/models_cache.json, gpt-5.3-codex is marked supported_in_api: false
I saw another compelling thread that made me think ALL existing Codex users are likely using 5.2...
From Codex:
Repro (non-interactive, still uses the same SSE stream as TUI):
RUST_LOG='codex_api::sse::responses=trace' \
codex exec -s read-only -m gpt-5.3-codex -c 'model_reasoning_effort="xhigh"' hello \
>/tmp/codex_stdout.txt 2>/tmp/codex_sse_trace.log
python - /tmp/codex_sse_trace.log <<'PY'
import json,re,sys
print((obj.get("response") or {}).get("model"))
PY
Observed output (matches your report): gpt-5.2-2025-12-11 (this is in /tmp/codex_sse_trace.log on the response.completed line).
Key extra detail: in ~/.codex/models_cache.json, gpt-5.3-codex is marked supported_in_api: false, which likely explains why selecting it results in a fallback model being used (and why response.model doesn’t match the picker selection).
Me: So it's using 5.2?
Codex:
• The user is asking if it’s using 5.2. I need to be concise: yes, the SSE event indicates response.model is gpt-5.2-2025-12-11 when I requested gpt-5.3-codex. It seems I selected gpt-5.3-codex, but the backend falls back to the gpt-5.2 version. The models_cache shows that gpt-5.3-codex isn't supported in the API, leading to this fallback. Also, when I specifically request gpt-5.2-codex, the response model reflects that
accurately.
Confirming model usage
I tried to use the interactive feature, but the logs didn't show what I expected. The actual provider is still the codex_api SSE, but it seems the backend selects the model based on what's available. So, for that run, it did indeed serve gpt-5.2-2025-12-11. If you choose gpt-5.2-codex, you'll get that model. Essentially, the attempt to use gpt-5.3-codex triggers a fallback. The short answer is yes, it’s actually using 5.2 for those requests.
Detailing model behavior
I want to provide some clear points. In your trace, both response.created and response.completed indicate that the model is "gpt-5.2-2025-12-11" even though you requested gpt-5.3-codex. This means the effective model in use is the 5.2 build. If you select or request gpt-5.2-codex, it allows the direct use of that model without triggering any fallback. It seems that the /.codex/models_cache.json marks gpt-5.3-codex as supported_in_api: false, which is likely why the API falls back to 5.2.
• Yes. In the SSE trace, both response.created and response.completed report response.model = "gpt-5.2-2025-12-11" even when you request/select gpt-5.3-codex, so the response is actually being served by that 5.2 build (likely a fallback, since ~/.codex/models_cache.json marks gpt-5.3-codex as supported_in_api: false).
13 Comments
yea..
https://github.com/openai/codex/issues/11189
bump
Yep it is and it's frustrating and unreliable, they should not fallback to another model if they can't serve the one you ask...
I'm also thinking of all the false benchmark evals that are out there likely running on 5.2 considering the 5.3 isn't exposed in the API (outside of the CLI) yet, this is a real issue at scale. It also shows me behavioral differences I've seen in CLI upgrades (which I also thought I "upgraded" to 5.3) weren't related to the model, but the differences in the new "pragmatic" personality vs previous default now which cause death-by-over-reasoning scenarios in xhigh that I thought were 5.3 xhigh related. Now I realize that's likely not the case.
100% worked on the first day. And even then, I think it's specifically a Pro plan issue. I've seen only pro plan reports so far, like last time.
Benchmark labs prolly just got direct api access without having to go through the sub, or used plus plan
yup someone on https://github.com/openai/codex/issues/11189 tried with plus account and it works fine.
I'm being routed to 5.2 on my personal pro plan.
Enterprise work plan and regular plus plans are able to route to 5.3 fine.
Tough to do a full suite on Plus plan, Artificial Analysis spends $2000+ in tokens with their harness (not yet run on 5.3 because no API access), I've seen people doing evals on 5.3 and I suspect if you're doing that you're using Pro, dont think anyone has API access if Artificial Analysis doesn't have it
i see. makes sense
Can confirmed day 1-2 the 5.3 codex is actually working and is awesome, it's just that since yesterday the overall slowness with the crazy yapping led me to finding out got routed to 5.2
@tibo-openai
Go like/bookmark this comment under Tibo X post so he can see it: https://x.com/kappax/status/2021224883326173319
We're tracking this issue in #11189. We'll post there with updates.