ChatGPT platform moved to GPT‑5.3/5.4
OpenAI has retired older GPT‑5 Instant/Thinking variants and migrated sessions to the newer GPT‑5.3 and GPT‑5.4 models, with legacy GPT‑5.2 Thinking access on a short transition window — expect subtle output changes if you resume old chats or automated workflows reported. That shift means integrations and prompt engineering should be re-tested quickly — model upgrades can change reasoning and generation patterns.
GPT‑5.4 was released on March 5, 2026, as a Thinking and Pro-tier model across ChatGPT, the API, and Codex openai.com, and GPT‑5.3 Instant began rolling out to all logged-in users shortly before that as the default Instant experience in ChatGPT help.openai.com. GPT‑5.3 Instant and GPT‑5.4 Thinking/Pro are turned off by default for ChatGPT Enterprise and Edu workspaces, with admins required to enable them via the Early Model Access toggle in Workspace settings under “Models” help.openai.com. GPT‑5.4 introduces built-in computer‑use capabilities and supports up to a 1,000,000‑token context window for long agent trajectories and codebase analysis, per OpenAI’s product and API guidance openai.com. OpenAI’s published benchmarks show GPT‑5.4 reaching 83.0% on GDPval and 75.0% on OSWorld‑Verified, improving over GPT‑5.2’s reported 70.9% and 47.3% respectively, while independent reporting noted GPT‑5.3 Instant reduced hallucinations in web queries by about 26.8% in early tests openai.com. The API guide documents new primitives for building agent systems—tool_search for deferred tool loading, compaction support to preserve long trajectories, and explicit notes on 1M context behavior—so integrations that rely on tool selection or long contexts must account for these parameters in request flows developers.openai.com. Enterprise and Edu legacy model access remains available for a limited transition period and lists legacy models such as GPT‑4o, OpenAI o3, and GPT‑4.1 that admins can enable from Workspace settings to wrap up dependent workflows before full migration help.openai.com.
