OpenAI's 3-model family (Instant, Thinking, Pro) plus a Codex variant, released Dec 11, 2025 as a response to Google Gemini 3 Pro. Better at spreadsheets, financial modeling, presentations, and multi-step projects.
Context window
Nieujawnione publicznie w komunikacie OpenAI dla GPT-5.2 (rodzina GPT-5 standardowo operuje w zakresie 400K tokenów wejścia)
tokens
Parameters
Nieujawnione publicznie
parameters
Release date
11 December 2025
Access:APIHostedDeployment:☁ Cloud
Overview
Applications
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
📏 Context: Nieujawnione publicznie w komunikacie OpenAI dla GPT-5.2 (rodzina GPT-5 standardowo operuje w zakresie 400K tokenów wejścia)
🧩 Parameters: Nieujawnione publicznie
✓ Tools
📥 Input: text, image, documents
Technical specification
Context window
Nieujawnione publicznie w komunikacie OpenAI dla GPT-5.2 (rodzina GPT-5 standardowo operuje w zakresie 400K tokenów wejścia)
tokens
Parameters
Nieujawnione publicznie
parameters
License
Proprietary (OpenAI Terms of Use; ChatGPT / API commercial terms)
Hardware requirements
Not applicable — the model is available only as hosted on the OpenAI API and in the ChatGPT products (Instant / Thinking / Pro / Codex). No on-device or local version.
Features:✓ Tool use
Modalities
⬇ Input
textimagedocuments
⬆ Output
textcodestructured_data
Capabilities and applications
Native model capabilities
Advanced reasoning
The ability to perform multi-step, structured reasoning: analysing problems, planning steps, and drawing conclusions from hypotheses. Reasoning-first models (e.g. GPT-5.1 Thinking) dedicate a portion of inference to chains of thought before responding.
Category: reasoning
Adaptive reasoning effort
The model decides how much 'thinking' to allocate to a given query: simple questions are answered quickly, complex problems receive more inference cycles. A GPT-5.1 feature (both Instant and Thinking) that shortens time on easy tasks and extends it for hard ones.
Category: reasoning
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Agentic coding
Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).
Category: coding
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Long context
Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Natural conversation
Conducting a conversation with a tone close to human: a warmer voice, empathy in emotional responses, humour, and avoiding the stiff 'AI assistant jargon'. Introduced as a deliberate improvement in GPT-5.1 Instant.
Category: language
Personality customization
The ability to tailor the model's response style and tone to the user's preferences via personality presets (e.g. Default, Friendly, Professional, Candid, Quirky, Efficient, Nerdy, Cynical) or fine-grained parameters (length, tone, emojis).
Category: language
Prompt caching
Cost-performance optimisation: repeated prompt fragments (e.g. system prompt, long documentation) are cached server-side and cheaper in subsequent calls. Significantly reduces cost for applications with long contexts.
Category: other
Shell and apply_patch tool use
The model's ability to invoke a system shell and the apply_patch tool to modify code files in agentic workflows. Characteristic of GPT-5.1 Codex variants (Mini, Max) and Claude Code / OpenCode-style pipelines.
Category: coding
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Multimodal understanding
Category: multimodal
Spreadsheet generation
Generating complete spreadsheets (Excel, Google Sheets, CSV) with formulas, conditional formatting, pivot tables, and multi-sheet data structures based on a natural-language description.
Category: structured_generation
Financial modeling
Building financial models: DCF, company valuations, scenario analysis, budget forecasts, P&L sheets, cap table models. Requires precise numerical reasoning and knowledge of accounting conventions.
Category: reasoning
Presentation creation
Generating presentations (PowerPoint, Google Slides, Keynote): slide structure, text content, graphic suggestions, and a talking-points outline. The model picks a layout and narrative appropriate to the presentation's purpose.
Category: structured_generation
Multi-step project execution
The ability to autonomously drive multi-hour, multi-step projects: decomposing the task, planning the sequence of actions, iteratively delivering results, and adjusting based on feedback. Key for enterprise and knowledge-work agents.
Category: planning
Application domains
ChatbotsCodingAI coding assistantAutonomous agent workflowsResearch assistanceWriting assistanceSearch assistanceWorkflow automationQ&A / Question answeringEmail draftingBrainstormingCreative writingKnowledge workTutoring / EducationMultimodal document understandingDocument analysisTranslationSummarizationContent generationDocument generationPresentation generationData analysisMeeting / Note assistanceSpreadsheet analysisSpreadsheet creationFinancial modeling
Benchmark results
1 benchmark
Box internal knowledge work benchmark
score delta · real-world knowledge-work tests, Box internal evaluation
+7 punktów vs GPT-5.1points
📅 11 Dec 2025📄 Aaron Levie (CEO Box) via VentureBeat, 2025-12-11
A result reported publicly by Box's CEO; the absolute value and methodology were not disclosed. Suggests measurable progress of GPT-5.2 over GPT-5.1 on knowledge-work tasks.
Pricing
Technical architecture
Core Architecture
Training Techniques
Deployment and security
🔒 Security / Enterprise
✓ Verified enterprise information
GPT-5.2 inherits OpenAI's security stack: SOC 2, GDPR, business terms in ChatGPT Enterprise/Edu/Business. The GPT-5.2-Codex variant explicitly addresses enterprise coding safety — per VentureBeat it 'weaves security into large-scale software refactors', introducing built-in security patterns into agentic large-scale code changes.
Safety context: OpenAI publishes separate system cards for GPT-5.x iterations. Note the controversy reported by The Guardian (2026-01-24) about the use of Grokipedia as a source — a training-data / retrieval quality issue, not the enterprise security stack itself.
Updated: 13 Jul 2026↗ Security documentation
Sources and related pages
10 sources
BlogOpenAI — Introducing GPT-5.2 (official post)WebWikipedia — GPT-5.2BlogThe Verge — GPT-5.2 is OpenAI's latest move in the agentic AI battle (2025-12-11)BlogTechCrunch — OpenAI fires back at Google with GPT-5.2 after 'code red' memo (2025-12-11)BlogReuters — OpenAI launches GPT-5.2 after 'code red' push to counter Google's Gemini 3 (2025-12-11)BlogCNBC — Sam Altman expects OpenAI to exit 'code red' by January after GPT-5.2 launch (2025-12-11)BlogWired — OpenAI Launches GPT-5.2 as It Navigates 'Code Red' (Fidji Simo interview)BlogVentureBeat — GPT-5.2 first impressions: powerful for business tasksBlogVentureBeat — Enterprise AI coding grows teeth: GPT-5.2-Codex (2025-12-18)BlogThe Guardian — Latest ChatGPT model uses Elon Musk's Grokipedia as source (2026-01-24)