In a year-end roundup published December 28, 2025, OpenAI's developer team (Vaibhav Srivastav and Katia Gil Guzman) framed 2025 not around a single model launch but around AI becoming easier to run in production. As models improved at planning, tool use, and longer-horizon tasks, teams shifted from prompting step-by-step to delegating work to agents. The headline themes: reasoning became a core tunable dial, multimodality became a first-class API citizen, and agent building blocks — the Responses API, the Agents SDK, and AgentKit — made multi-step workflows easier to ship and operate.

On reasoning, the post traces an arc from separate models to a unified line. After the reasoning paradigm was introduced at the end of 2024, early 2025 was the era of distinct reasoning families: o1, o3, and o4-mini demonstrated that spending extra compute to think before answering could dramatically improve reliability on complex, multi-step work, while o3-mini signaled that reasoning could be delivered in cost-efficient, developer-friendly form factors. By mid-to-late 2025 the trend was convergence — reasoning depth, tool use, and conversational quality increasingly lived inside the same flagship line, absorbed into the GPT-5.x family so that picking a model became a cost/latency/quality tradeoff rather than a choice between fundamentally different families.

Multimodality stopped meaning merely accepting an image input and started meaning building an end-to-end product across modalities. Next-generation audio models improved speech-to-text accuracy and added more controllable text-to-speech, and the Realtime API went GA with low-latency bidirectional audio streaming for production voice agents. On images, GPT Image 1 introduced a new generation of image models with strong instruction following and high input fidelity for preserving faces and logos during edits; GPT Image 1 mini made native generation more cost efficient; and GPT Image 1.6 marked a step change in quality and edit consistency, also usable as a tool inside the Responses API.

Video became a first-class modality. Sora 2 and Sora 2 Pro introduced higher-fidelity generation with stronger temporal coherence and remixing support, exposed through a Video API at the v1/videos endpoint alongside text, images, and audio. PDF inputs — including PDF-by-URL, which references documents without an upload step — enabled document-heavy workflows directly in the API.

Codex was the year's other big story, moving beyond a coding model into what OpenAI calls a software-engineer teammate connecting models, local tooling, and cloud. Early reasoning models' gains on multi-file edits, debugging, and planning were consolidated into the GPT-5 family, with GPT-5.2-Codex becoming the default for code generation, review, and repo-scale reasoning. The open-source Codex CLI brought agent-style coding into local environments, with sandboxing and approval modes keeping humans in the loop, support for AGENTS.md and MCP for adapting to a repo, and Codex Autofix for automating parts of the workflow in CI.

The most important platform change was the move toward agent-native APIs. The Responses API added support for multiple inputs and outputs across modalities, reasoning controls and summaries, and better tool calling during reasoning. On top of it, the open-source Agents SDK (for Python and TypeScript, and provider-agnostic) and AgentKit — spanning Agent Builder, ChatKit, the Connector Registry, and evaluation loops — made agents easier to build and orchestrate. A Conversations API and durable conversation state made threads replayable, while connectors and MCP servers brought external context and actions through trusted tool surfaces.

OpenAI also standardized a set of built-in tools: web search for up-to-date retrieval with citations, file search over hosted vector stores as a default RAG primitive, Code Interpreter for running Python in sandboxed containers, and computer use for click/type/scroll automation loops. For running at scale, prompt caching cut latency and input costs on shared prefixes, Background mode enabled long-running responses without holding a client connection open, and webhooks turned polling into event-driven systems for batch, background, and fine-tuning completion.

The post also covered open-weight models — gpt-oss 120b and 20b reasoning models for self-hosting and on-prem, plus gpt-oss-safeguard 120b and 20b safety and policy models — and a maturing evaluation stack including the Evals API, reinforcement fine-tuning with programmable graders, supervised fine-tuning and distillation, and a prompt optimizer. Its end-of-2025 default recommendations: GPT-5.2 for general-purpose and multimodal work, GPT-5.2 Pro for deeper reasoning, GPT-5.2-Codex for coding, GPT Image 1.6 for image generation, and gpt-realtime for low-latency voice.