OPENAI ROLLS OUT GPT-5.4 MINI IN CHATGPT AND SUNSETS LEGACY DEEP RESEARCH
OpenAI added GPT-5.4 mini to ChatGPT as a fallback for reasoning and is removing the legacy deep research mode. OpenAI is rolling out GPT-5.4 mini in ChatGPT. ...
OpenAI added GPT-5.4 mini to ChatGPT as a fallback for reasoning and is removing the legacy deep research mode.
OpenAI is rolling out GPT-5.4 mini in ChatGPT. It’s available to Free and Go users via the Thinking option, and used as a fallback for GPT-5.4 Thinking during rate limits for paid tiers. Enterprise admins can also default Auto routing to 5.4 mini; it won’t appear as a selectable model and the older GPT-5 Thinking mini will be retired in 30 days release notes.
OpenAI will also remove the legacy deep research mode on March 26, 2026. The current deep research experience remains, and historical conversations and results stay accessible (release notes; see also the digest on Releasebot).
Model routing will change during peak load: GPT-5.4 mini may answer instead of 5.4 Thinking, which impacts quality, latency, and reproducibility.
Teams using legacy deep research workflows must verify continuity with the current deep research experience before March 26.
-
terminal
A/B prompts your team relies on (SQL authoring, pipeline debugging, runbook drafting) across GPT-5.4 Thinking vs fallback mini; track latency, token use, and acceptance rate.
-
terminal
Enable Auto-switch to Thinking and stress-test during peak hours to measure fallback frequency and output consistency; confirm transcripts/logs capture model identifiers.
Legacy codebase integration strategies...
- 01.
If you need consistent outputs for SOPs or parser-dependent prompts, consider disabling auto-switch or pinning model choice where possible in ChatGPT Enterprise.
- 02.
Update evaluations and playbooks to account for GPT-5.4 mini responses and the removal of legacy deep research; re-baseline quality checks.
Fresh architecture paradigms...
- 01.
Design workflows that tolerate small style differences and let Auto routing use mini for throughput while reserving Thinking for harder tasks.
- 02.
Instrument conversations with model id, latency, and outcome metrics to refine routing and prompt patterns over time.