OpenAI has launched GPT‑5.2 across ChatGPT and its API, positioning the update as a major step for professional “knowledge work” with stronger reasoning, long‑context performance, fewer errors, and new spreadsheet/presentation workflows—while the AI race with Google’s Gemini 3 heats up. OpenAI+2OpenAI Help Center+2
OpenAI’s latest ChatGPT upgrade—commonly referred to as “ChatGPT 5.2”—arrived on December 11, 2025, and it’s less about flashy new gimmicks and more about something people actually notice in daily use: better reasoning, stronger reliability, and improved “agentic” work behavior (using tools, staying on track, and completing multi‑step tasks). Reuters+2OpenAI+2
The update is powered by a new model family called GPT‑5.2, which OpenAI is rolling out across ChatGPT (with an Auto mode that can switch between faster and more analytical variants) and its developer platform via the API.
Below is what’s been announced—and what major outlets and OpenAI’s own documentation say—about ChatGPT 5.2 / GPT‑5.2 right now.
The big change: GPT‑5.2 is a “three‑model” release inside ChatGPT
OpenAI is shipping GPT‑5.2 in three flavors, each tuned for a different tradeoff:
- GPT‑5.2 Instant: faster day‑to‑day responses (info seeking, how‑tos, writing, translation).
- GPT‑5.2 Thinking: deeper, structured reasoning for harder tasks (coding, long documents, planning, math/logic).
- GPT‑5.2 Pro: slowest, “research‑grade” option when higher quality is worth waiting (but with some feature caveats in ChatGPT). OpenAI Help Center+2OpenAI Platform+2
In ChatGPT specifically, OpenAI is also pushing GPT‑5.2 Auto—a system that automatically chooses between Instant and Thinking depending on your prompt, aiming to deliver a “fast by default, deep when needed” experience. OpenAI Help Center+1
What OpenAI says GPT‑5.2 is best at: “professional knowledge work”
OpenAI is explicitly marketing GPT‑5.2 as a major leap for tasks people do at work—especially things that combine planning, writing, tool use, and accuracy.
In its announcement and benchmark materials, OpenAI emphasizes performance on a test called GDPval, which spans 44 occupations and includes outputs like spreadsheets, presentations, schedules, and other real‑world work products.
OpenAI’s headline claim: GPT‑5.2 Thinking beats or ties top professionals on 70.9% of GDPval tasks.
Business Insider highlighted that OpenAI says the model can produce GDPval outputs at “>11x the speed and <1% the cost” of expert professionals—though OpenAI frames this as work that still benefits from human oversight. Business Insider+1
New “work outputs” in ChatGPT: spreadsheets and presentations
One of the most practical, user-facing updates is OpenAI’s push into structured business deliverables—especially spreadsheet formatting/financial modeling and slideshow creation.
OpenAI notes that to use the new spreadsheet and presentation capabilities in ChatGPT, users need to be on a Plus, Pro, Business, or Enterprise plan and choose GPT‑5.2 Thinking or GPT‑5.2 Pro. OpenAI also warns some complex generations may take many minutes to complete.
This aligns with how the broader industry battle is shifting: the most valuable AI isn’t the one that writes a clever paragraph—it’s the one that can reliably produce a usable “work artifact” and stay consistent across long, messy tasks.
Does GPT‑5.2 hallucinate less? Here’s what the reporting and system card say
Multiple outlets—including The Verge—report that OpenAI is leaning hard into improved trustworthiness, including claims that GPT‑5.2 “hallucinates less” than GPT‑5.1, especially in the Thinking variant. The Verge+1
OpenAI’s own materials describe fewer error‑containing responses in internal testing and improved reliability across tool use and long context.
More specifically, OpenAI’s GPT‑5.2 system card addendum reports:
- GPT‑5.2 Thinking performs on par with (or slightly better) than predecessors on factual hallucination evaluations based on real ChatGPT‑like prompts.
- With browsing enabled, GPT‑5.2 Thinking achieved <1% hallucination rate across five domain subsets (business/marketing research, financial/tax, legal/regulatory, academic writing, current events/news) in OpenAI’s evaluation setup.
Important context for readers: these numbers come from OpenAI’s own evaluation methodology (LLM‑based graders, specific prompt sets, and, in one case, browsing enabled). They are useful signals, but not the same as a universal guarantee that errors are gone.
Long context is a headline feature—especially for “agentic” workflows
OpenAI is also spotlighting long‑context reasoning and improved ability to keep track of details across large inputs—critical for things like multi‑document research, codebases, contracts, or support tickets.
OpenAI’s benchmark chart shows GPT‑5.2 Thinking scoring 83.8% on MRCRv2, its long-context reasoning evaluation (with GPT‑5.2 Pro at 84.6%).
Consumer‑facing coverage like TechRadar frames this as “bigger memory” and better ability to stay coherent over long threads and large document sets. TechRadar
How to access ChatGPT 5.2: rollout, plans, and message limits
OpenAI says GPT‑5.2 is gradually rolling out, so not every account will see it immediately.
Availability by plan (ChatGPT)
According to OpenAI’s Help Center:
- All ChatGPT tiers get GPT‑5.2, but paid tiers get more control via the model picker.
- GPT‑5.2 Pro is limited to Pro, Business, Enterprise, and Edu plans.
- Free tier: up to 10 messages every 5 hours with GPT‑5.2, then chats fall back to a “mini” model until reset. OpenAI Help Center
- Plus: up to 160 messages every 3 hours with GPT‑5.2 (noted as a temporary increase), then fall back to a mini model.
- Thinking usage (manual selection) has a weekly cap for Plus/Business (up to 3,000 messages per week), but auto-switching doesn’t count toward that limit.
- Business and Pro: “unlimited access,” subject to abuse guardrails and Terms of Use restrictions (e.g., scraping, credential sharing, reselling access). OpenAI Help Center
Context windows in ChatGPT
OpenAI lists different context limits depending on tier and mode:
- Instant: 16K (Free), 32K (Plus/Business), 128K (Pro/Enterprise)
- Thinking: 196K (all paid tiers)
A notable limitation
OpenAI notes that Canvas and image generation are not available with GPT‑5.2 Pro in ChatGPT.
What changes inside the ChatGPT experience: Auto switching, “Answer now,” and thinking controls
OpenAI’s documentation describes how GPT‑5.2 Auto chooses when to apply deeper reasoning based on signals from the prompt and learned patterns from user behavior.
When the system switches into deeper reasoning, users may see a slimmed‑down reasoning trace and an option called “Answer now” to return to Instant responses instead of waiting for the full reasoning process. OpenAI Help Center+1
For users who select Thinking manually, OpenAI also describes a thinking-time toggle in the composer (with more options for Pro users), though it’s currently noted as web-only and not synced to mobile.
OpenAI has additionally updated tone and personalization controls, including refined personality presets and sliders for response style.
For developers: model names, pricing, and giant context windows in the API
OpenAI’s GPT‑5.2 launch isn’t only about ChatGPT—it’s also about developers building products on top of the model family.
Model naming (ChatGPT vs API)
OpenAI maps the ChatGPT variants to API model names like this:
- ChatGPT‑5.2 Instant →
gpt-5.2-chat-latest - ChatGPT‑5.2 Thinking →
gpt-5.2 - ChatGPT‑5.2 Pro →
gpt-5.2-pro
OpenAI also notes that GPT‑5.2 Pro supports setting a reasoning parameter, and both Pro and Thinking support a new reasoning effort level called xhigh.
Pricing (API)
OpenAI’s launch post lists:
gpt-5.2at $1.75 per 1M input tokens and $14 per 1M output tokens, with a 90% discount on cached inputs.
For gpt-5.2-pro, OpenAI’s model documentation lists:
- $21.00 per 1M input tokens and $168.00 per 1M output tokens.
Context window and output limit (API)
OpenAI’s API documentation for GPT‑5.2 Pro lists:
- 400,000 token context window
- 128,000 max output tokens
- Aug 31, 2025 knowledge cutoff
OpenAI also notes Pro requests may take several minutes, recommending background mode to avoid timeouts and specifying supported reasoning.effort levels (medium, high, xhigh).
Safety and policy news: mental health guardrails, age prediction, and “adult mode” timeline
Safety is not a footnote in this release—it’s part of the product roadmap.
OpenAI says GPT‑5.2 continues its “safe completion” approach and includes improvements in how the models handle prompts indicating suicide/self-harm, mental health distress, or emotional reliance. OpenAI points readers to the GPT‑5.2 system card for details. OpenAI+1
OpenAI also says it is in the early stages of rolling out an age prediction model intended to automatically apply content protections for users under 18.
And in one of the most widely shared “what’s next” details: OpenAI executives have said an “adult mode” for ChatGPT is expected to debut in Q1 2026, after early testing of age prediction to avoid misidentifying adults as teens. The Verge+2Business Insider+2
Why GPT‑5.2 launched now: the Gemini 3 pressure and OpenAI’s “code red”
GPT‑5.2 is landing in a heated market moment.
Reuters reported that OpenAI launched GPT‑5.2 after CEO Sam Altman reportedly issued an internal “code red” to accelerate development and counter Google’s Gemini 3, which Google launched in November 2025 and has touted as a leaderboard leader. Reuters
TechCrunch also frames GPT‑5.2 as a direct response in the “arms race,” highlighting Gemini 3’s position on LMArena leaderboards and the broader competition around multimodal and agentic workflows. TechCrunch+1
Altman told CNBC (per Reuters) that Gemini 3 had less impact on OpenAI’s metrics than feared—underscoring how closely consumer usage and perceived model quality are now tied to competitive releases.
The same-day business headline: Disney invests $1B and licenses characters for OpenAI tools
One reason the GPT‑5.2 news cycle has been unusually loud: it broke the same day as a blockbuster entertainment partnership.
Reuters reported Disney is investing $1 billion in OpenAI and licensing major IP (including characters from Star Wars, Pixar, and Marvel) for OpenAI’s video tools, with licensed character generation expected to begin early next year.
The Verge also reported the deal includes a three‑year licensing arrangement tied to “user‑prompted social videos,” and that Disney will become a “major customer” of OpenAI alongside the equity investment. The Verge
For ChatGPT 5.2 coverage, the relevance is indirect but real: it underscores OpenAI’s broader strategy—enterprise adoption and partnerships—at the same moment it’s shipping a model pitched for professional workloads.
What to watch next: Codex optimization, science claims, and the next rollout wave
OpenAI says a Codex‑optimized version of GPT‑5.2 is planned “in the coming weeks.” OpenAI
OpenAI is also heavily promoting GPT‑5.2’s capabilities in math and science, citing strong results on benchmarks like GPQA Diamond and FrontierMath, and pointing to experiments where GPT‑5.2 Pro contributed to progress on a narrow open question in statistical learning theory under human oversight.
One paper linked by OpenAI includes an unusually direct attribution claim: it states that results in the paper were derived by variants of GPT‑5.2 Pro, with human verification and editing.
In the near term, the most important practical question for users is simpler: when the rollout reaches their account, and whether the day‑to‑day improvements (structure, reliability, tool use, fewer mistakes) match OpenAI’s claims in real workflows.
Bottom line
ChatGPT 5.2 isn’t positioned as a “new chatbot app”—it’s a model and product experience upgrade that aims to make ChatGPT more useful for real work: structured deliverables, long-context problem solving, better tool use, and fewer factual errors—while OpenAI responds to growing pressure from Google’s Gemini 3 and an increasingly competitive “agentic AI” market. The Verge+3OpenAI Help Center+3OpenAI+3
As with every frontier-model release, the fine print matters: rollout is gradual, access varies by plan, some features differ across variants, and OpenAI’s own system card notes areas where performance can regress or shift depending on evaluation setup.
