ChatGPT 5.2 arrives

ChatGPT 5.2 is out. Here’s a quick rundown of what changed and what to expect.

Highlights

• Faster: average response latency is roughly 30% lower than 5.1 in my tests.
• Sharper reasoning: better multi-step math and code generation, fewer hallucinated APIs.
• More controllable: improved system message adherence and JSON formatting reliability.
• Longer focus: modest context boost with better retention across longer chats.

What feels different

• Instructions stick: nuanced tone and formatting prompts are respected more consistently.
• Code answers: 5.2 returns smaller, runnable snippets and includes lint-friendly formatting.
• Tool use: the model follows function/tool schemas with fewer edge-case failures.
• Search-style answers: citations and inline references are cleaner when browsing is enabled.

New capabilities

• JSON mode is stricter, reducing trailing prose after structured replies.
• Inline planning: you can ask for a plan plus execution; it now separates steps more cleanly.
• Tabular summaries: improved Markdown tables and alignment.
• Vision: better OCR for documents and more accurate chart/table reads.

Safety and guardrails

• Reduced content drift in long sessions.
• Better refusal clarity when policies are triggered.

Availability and pricing

• Rolling out to ChatGPT and API with the 5.2 checkpoint; availability may take a few hours to propagate.
• Model ID exposed via the API as gpt-5.2 (name subject to change; check the docs).

Tips to get the most out of it

• Be explicit about output shape: “Return valid JSON with keys X, Y.”
• For code, ask for brief rationale plus the minimal snippet; it keeps responses compact.
• When chaining tools, provide short, numbered steps; 5.2 follows enumerations well.

I’ll update this post as I learn more from real-world use. Let me know if you spot quirks or regressions.