GPT-5.2

A General Professional Model Before the GPT-5.4 and GPT-6 Generations

GPT-5.2 was released as a flagship general-purpose model for both ordinary professional work and agentic tasks. OpenAI highlighted improvements over GPT-5.1 in general intelligence, instruction following, accuracy, token efficiency, multimodality, code generation, tool calling, context management, and spreadsheet understanding.

That breadth is what separates GPT-5.2 from more specialized variants. GPT-5.2 Codex focused on interactive coding products, while GPT-5.2 Pro traded more time and compute for harder problems. The base GPT-5.2 occupied the middle: a broad model for complex reasoning, knowledge-heavy tasks, code-heavy work, and multi-step agents.

The model also introduced several API-era improvements that are useful to understand historically and operationally: an xhigh reasoning level, concise reasoning summaries, client-side compaction, Apply Patch support, local shell support, custom tools, and allowed-tool restrictions.

OpenAI now recommends GPT-6 Astra as the current flagship for the most demanding API workloads. GPT-5.2 itself, however, remains documented as an available previous flagship rather than a deprecated model.

• Previous flagship model for complex professional work.
• Broad fit across reasoning, knowledge, coding, spreadsheets, and multimodal tasks.
• Supports reasoning from none through xhigh.
• Includes explicit context-management and tool-control features.
• Current OpenAI guidance recommends GPT-6 Astra for newer flagship deployments.

Cost Depends on Reasoning Depth and Workflow Shape

A GPT-5.2 request can be inexpensive when the prompt is small and reasoning is disabled, or materially more expensive when a task requires long context, extensive reasoning, tool feedback, and a large final answer.

Reasoning tokens are billed as output tokens, so increasing reasoning effort changes more than latency. It can directly raise the cost of a request even when the visible final response stays relatively short.

Cached input is priced at one tenth of the standard input rate. That can be useful for workflows that repeatedly reuse stable instructions, schemas, repository guidance, or other prefix content.

For multi-step agents, cost should be measured per completed workflow. A task may involve several model turns, shell output, tool calls, context compaction, and retries before producing one user-visible result.

• $1.75 per 1M input tokens.
• $0.175 per 1M cached input tokens.
• $14.00 per 1M output tokens.
• Reasoning tokens count toward output usage.
• Tool-specific services can introduce separate charges.

Start Cheap, Increase Deliberation Only When Evaluation Justifies It

The none default lets GPT-5.2 handle lower-latency interactions without forcing every request through an extended reasoning phase.

For more demanding work, developers can progressively increase reasoning effort. low and medium are useful evaluation points for structured professional tasks, while high and xhigh are better candidates for harder planning, synthesis, debugging, and multi-step agent execution.

GPT-5.2 was the generation that added the xhigh level to this model line. OpenAI also introduced concise reasoning summaries, giving applications a more compact way to expose useful information about the reasoning process without relying on hidden reasoning tokens.

The right setting is workload-specific. An application that always uses the maximum level may waste time and output-token budget on tasks that already succeed at none or low.