Skip to content
Bud's Wiki

GPT-5.4 Thinking and GPT-5.4 Pro rolling out in ChatGPT (product announcement)

TL;DR

Section titled “TL;DR”

A March 5, 2026 OpenAI tweet announcing the rollout of GPT-5.4 Thinking and GPT-5.4 Pro in ChatGPT, with GPT-5.4 also made available in the API and Codex. The tweet positions the GPT-5.4 family as bringing OpenAI’s advances in reasoning, coding, and agentic workflows into a single frontier model. The post is a pure product announcement with no benchmark numbers, system-card link, technical detail, or methodology disclosed in the tweet text — only the existence and availability of the two ChatGPT variants (Thinking, Pro) and the API/Codex availability of the base model. Filed here as a tweet-as-primary artifact since no companion blog post or technical report was linked.

Key claims

Section titled “Key claims”
  • GPT-5.4 Thinking and GPT-5.4 Pro are rolling out in ChatGPT, dated 6:10 PM PT March 5, 2026 [tweet body, timestamp].
  • GPT-5.4 is simultaneously available in the OpenAI API and in Codex [tweet body].
  • The GPT-5.4 family is framed as consolidating reasoning, coding, and agentic-workflow capability into one frontier model [tweet body].

Method

Section titled “Method”

Not disclosed. The tweet links no system card, blog post, or technical report; the only attached media is a single image (not analyzed here). The earlier GPT-5.3-Codex announcement (Introducing GPT-5.3-Codex) accompanied a long product-blog post with a benchmark table and a system card link — the GPT-5.4 launch as captured by this tweet does not.

Results

Section titled “Results”

No benchmark numbers in the tweet. The only result-shaped claims are the consolidation framing (“reasoning, coding, and agentic workflows into one frontier model”) and the surface-area expansion (Thinking + Pro in ChatGPT, base model in API + Codex on the same day).

Why it’s interesting

Section titled “Why it’s interesting”

This is the headline closed-frontier datapoint that sits one step above GPT-5.3 — the version of OpenAI’s flagship that the open-agentic-SWE cluster (Agentic Software Engineering) and the reasoning-RL cluster (Reasoning RL) will now be measured against. Reading it alongside the much more detailed Introducing GPT-5.3-Codex post (which set SWE-Bench Pro and Terminal-Bench 2.0 as the consensus arenas) shows OpenAI continuing to ship via a single unified frontier model rather than a coding-specialized split; whether the Thinking / Pro split inside ChatGPT corresponds to distinct training endpoints or a single base with different inference-time scaffolds is not stated. Worth re-filing once a blog post, system card, or independent benchmark table appears — the tweet itself is a tracker, not a source.

See also

Section titled “See also”