Computer use is not working with the gpt-5.4-pro

[ramkumarkoppu]

Tried with the python sdk. used the same tool calling in both the gpt-5.4 and gpt-5.4-pro but pro returning error.

[sps]

Hi @ramkumarkoppu

Thanks for reporting this issue. I can confirm that enabling tools=[{"type": "computer"}] with gpt-5.4.pro gives:

Error code: 400 - {'error': {'message': "Tool 'computer' is not supported with gpt-5.4-pro.", 'type': 'invalid_request_error', 'param': 'tools', 'code': None}}

I have sent this to the OpenAI team and will follow up once I have updates.

[ramkumarkoppu]

The docs confirm {"type": "computer"} is the correct format, and only mention gpt-5.4 — not gpt-5.4-pro. Despite the model page listing “Computer use: Supported”, the API actively rejects it. This is likely a docs-vs-reality discrepancy.

Result: gpt-5.4-pro still does NOT work for computer use.

[dinag]

Hi @ramkumarkoppu we applied a change for you and this now should be resolved, thank you for reporting it!

[vb]

Follow-up observation by @ramkumarkoppu :

@ramkumarkoppu

Thanks for looking into this. The 400 rejection is gone, but now the API hangs indefinitely — 0 tokens returned even after 300s. Tested with tools=[{“type”: “computer”}] and both high and xhigh reasoning effort. gpt-5.4 works fine with the same request.

[dinag]

Thanks @ramkumarkoppu I just tried now and it worked fine, maybe it was a temporary thing? Can you try again please?

[ramkumarkoppu]

Hi @Dinag, thanks for looking into this. I just retested — the 400 rejection is indeed gone, but the API still hangs indefinitely with zero output tokens. I ran a side-by-side comparison with gpt-5.4 as a control:

Setup:

• OpenAI Python SDK: 2.29.0 (latest)

• Tool: [{"type": "computer"}]

• Reasoning effort: high

• Input: simple prompt (“Click the Start button”) + JPEG screenshot

• Tested: March 24, 2026

gpt-5.4-pro (stream=True):

1.7s  → response.created
1.7s  → response.in_progress
31.8s → keepalive
61.6s → keepalive
91.6s → keepalive
(no output — hangs indefinitely, only keepalive pings every ~30s)

Response ID: resp_08b77db23fe9622d0069c2c85df720819487c5d010f5c5a589

gpt-5.4-pro (stream=False): Timed out at 90s, no response.

gpt-5.4 control (stream=True) — same exact request:

1.6s → response.created
1.6s → response.in_progress
1.9s → response.output_item.added (reasoning)
3.8s → response.output_item.done
4.2s → response.output_item.done (computer_call)
4.3s → response.completed ✓

Response ID: resp_06901a17dd17d1d60069c2c913bb40819caa10e8aca8d67c81
Usage: 2,392 input / 106 output tokens (89 reasoning) — completed in 4.3s.

Summary: gpt-5.4-pro accepts the request (no more 400), creates a response, goes to in_progress, but never produces any output items — just keepalive pings until timeout. gpt-5.4 returns a valid computer_call in ~4s with the identical payload.

[dinag]

Thank you so much @ramkumarkoppu ! I checked your response id and it was actually successful and finished in 1 minute 46 seconds, can you increase your timeout? That should do it for you.