Discover the best AI tools curated for professionals.

AIUnpacker

Search everything

Find AI tools, reviews, prompts, and more

Quick links
AI Skills & Learning Updated May 3, 2026 Verified

12 Days of OpenAI

A thoroughly researched, cross-verified breakdown of every OpenAI 12 Days announcement: what launched, who it was for, what it cost, and what changed since.

AIUnpacker

AIUnpacker Editorial

March 6, 2026

9 min read
AIUnpacker

AIUnpacker

Mar 6, 2026 · 9m read

Mar 6, 2026 9 min Updated May 3, 2026

Key Takeaways

A thoroughly researched, cross-verified breakdown of every OpenAI 12 Days announcement: what launched, who it was for, what it cost, and what changed since.

Editorial Disclosure & Affiliate Notice

This content is published for informational and educational purposes only. It is not intended as a substitute for professional, legal, financial, or medical advice. AIUnpacker is reader-supported — when you buy through our links, we may earn a commission at no extra cost to you, and our editorial picks are never influenced by compensation.

  • For educational purposes only. Nothing here should be taken as a guarantee, recommendation, or professional recommendation.
  • AI-assisted editing. Drafts are produced with AI assistance and reviewed by our human editorial team.
  • Opinions are our own. Also, we are not affiliated with most tools we cover unless explicitly stated.
  • Information may be outdated. Verify pricing, features, and policies directly with the vendor.
  • Last reviewed: March 6, 2026.

Read more on our About page, Terms and Editorial Policy.

12 Days of OpenAI

OpenAI’s “12 Days of OpenAI” ran from December 5 through December 20, 2024. The company kicked it off by announcing ChatGPT had crossed 300 million weekly active users, then dropped something new every weekday: models, subscription tiers, a video generator, developer APIs, even a literal phone number you could call.

I cross-checked each announcement against OpenAI’s own blog posts, their official 12 Days hub, The Verge’s live “ship-mas” coverage, VentureBeat’s reporting, Wikipedia, and the actual API documentation. Every benchmark number and launch detail below is verified against at least two independent primary sources. Here is what actually happened.


The Short Version: The Full 12 Days

DayDateAnnouncementStatus
1Dec 5o1 full release + ChatGPT Pro ($200/month)Launched
2Dec 6Reinforcement Fine-Tuning research programNow in production API
3Dec 9Sora video generation (Sora Turbo)Discontinued April 26, 2026
4Dec 10Canvas out of previewAvailable to all users
5Dec 11ChatGPT in Apple IntelligenceiOS 18.2 / macOS Sequoia
6Dec 12Advanced Voice with video + Santa ModeLaunched; Santa was seasonal
7Dec 13Projects in ChatGPTLaunched
8Dec 16ChatGPT Search for all usersFree tier; originally Oct 31 for paid
9Dec 17Developer tools: o1 API, Realtime API, Preference FT, Go/Java SDKsLaunched for tier-5 API users
10Dec 181-800-CHATGPT phone + WhatsAppUS; 15 min/month free
11Dec 19Work with Apps: desktop integrationsmacOS: Notes, Notion, Quip, coding
12Dec 20o3 and o3-mini preview + deliberative alignmentPreview; safety apps closed Jan 10, 2026

Day-by-Day Breakdown

Day 1: o1 Full Release and ChatGPT Pro

“The model that thinks before it speaks finally gets a price tag.”

OpenAI shipped the full o1 model, replacing o1-preview. The headline was o1 pro mode exclusive to the new $200/month ChatGPT Pro tier which uses more compute to “think harder.” Benchmarks: o1 scored 79.2% on AIME 2024 math (up from 42% for o1-preview), 75.7% on GPQA Diamond (PhD-level science), and 48.9% on SWE-bench Verified for coding. O1 pro mode added a “4/4 reliability” metric (getting it right across four independent attempts): 80% on competition math and 74% on PhD science, both solid gains over base o1.

OpenAI also awarded 10 ChatGPT Pro grants to medical researchers, including Catherine Brownstein (Boston Children’s Hospital / Harvard) and Rhoda Au (Boston University).

Day 2: Reinforcement Fine-Tuning

Reinforcement Fine-Tuning (RFT) uses a programmable grader that scores model outputs, then reinforces reasoning chains that produced high scores unlike supervised fine-tuning, which trains on fixed “correct” answers. The key insight: instead of telling the model “this is the right answer,” you define what “right” looks like and let the training loop discover the best reasoning path. OpenAI said it works best on verifiable tasks in math, science, legal, healthcare, and financial services domains where you can programmatically or expert-review correctness. At launch this was an alpha research program with an application form; it is now a production API feature supporting o-series reasoning models, with documented use cases for legal analysis, medical diagnosis, and security compliance.

Day 3: Sora Video Generation

Sora launched at Sora.com on December 9, powered by Sora Turbo (much faster than the February 2024 preview). Plus ($20/month) gave 50 priority videos at 480p/5 seconds. Pro ($200/month) unlocked 500 videos at 1080p/20 seconds, watermark-free downloads, and five simultaneous generations. It included a storyboard tool, C2PA metadata, and visible watermarks. OpenAI blocked CSAM and sexual deepfakes, and limited people-uploads while refining mitigations.

Critical update: Sora was discontinued April 26, 2026 (web/app), with the API following September 24, 2026. If you built workflows on it, this is your reminder that platform products are not permanent.

Day 4: Canvas for Everyone

Canvas a side-by-side editing window for writing and code exited beta. Writing shortcuts include suggest edits, adjust length, change reading level (Kindergarten to Graduate School), add final polish, and emoji insertion. Coding shortcuts cover review, debugging logs, commenting, bug fixing, and language porting (JS, TS, Python, Java, C++, PHP). The release added Python code execution and made Canvas available inside custom GPTs.

Day 5: ChatGPT in Apple Intelligence

Launched with iOS 18.2 and macOS Sequoia. Siri hands off queries to ChatGPT and delivers responses inline. On macOS, ChatGPT can analyze documents. On iPhone 16, the Camera Control button launches ChatGPT’s vision feature. No account required for basic use Apple Intelligence handles privacy, ChatGPT handles frontier capability.

Day 6: Advanced Voice with Video and Santa Mode

Advanced Voice Mode gained video the model sees you through your camera while you talk. Practical applications: tutoring (show me the math problem), troubleshooting (look at this error screen), and hands-free coaching. Santa Mode was a seasonal voice persona activated by a snowflake icon available through the end of 2024. It was a lighthearted feature, but the underlying video-capable voice infrastructure was the real signal.

Day 7: Projects in ChatGPT

Folder-like organization inside ChatGPT: group related chats, upload files, and set custom instructions that apply across a project. Solved the “which chat was that in?” problem for anyone with more than a few active threads. It turned ChatGPT from a stream of disconnected prompts into something closer to a lightweight workspace particularly useful for consultants tracking multiple clients, researchers managing literature reviews, and teams collaborating on shared context.

Day 8: ChatGPT Search for Everyone

ChatGPT Search rolled out to all logged-in free users on December 16 (originally October 31 for paid subscribers), extended to logged-out users on February 5, 2026. The search model is a fine-tuned GPT-4o distilled from o1-preview, pulling from third-party search providers and publisher partners including Associated Press, Axel Springer, Financial Times, Le Monde, News Corp, Reuters, and Time. The update also brought search to Advanced Voice Mode.

Day 9: Developer Tools Blitz

The densest day. OpenAI shipped:

  • o1 in the API with function calling, Structured Outputs, vision, developer messages, and a reasoning_effort parameter. Used 60% fewer reasoning tokens than o1-preview.
  • Realtime API with WebRTC support, 60% price reduction on audio tokens ($40/1M input, $80/1M output), and GPT-4o mini Realtime support.
  • Preference Fine-Tuning using Direct Preference Optimization (DPO) train by comparing preferred vs. non-preferred responses. Rogo AI reported accuracy improvements from 75% to 80%+.
  • New Go and Java SDKs in beta, joining Python, Node.js, and .NET.

Day 10: 1-800-CHATGPT

Call 1-800-CHATGPT (1-800-242-8478) for up to 15 minutes of free voice interaction per month in the US. The same number works globally on WhatsApp. OpenAI CPO Kevin Weil said it was spun up in just a few weeks an experiment in lowering the access barrier to zero. No app, no account, not even a smartphone required if you called the voice line. It was a clever move to reach people who would never install ChatGPT but might dial a toll-free number.

Day 11: Work with Apps

ChatGPT’s macOS desktop app gained integrations with Apple Notes, Notion, Quip, and coding environments. The demo showed ChatGPT reading from multiple apps and performing tasks across them a step toward “agentic” workflows where the AI operates inside your existing tools.

Day 12: o3 Preview and Deliberative Alignment

OpenAI saved the biggest research announcement for last. o3 and o3-mini were previewed skipping “o2” to avoid trademark conflict with the British telecom O2. Benchmarks: 96.7% on AIME 2024 math, 87.7% on GPQA Diamond, Codeforces rating of 2727 (beating OpenAI’s Chief Scientist), and 25.2% on EpochAI Frontier Math (no other model exceeded 2%). On ARC-AGI, o3 tripled o1’s score past 85%.

These were NOT released to the public. OpenAI opened safety researcher applications (closed January 10, 2026). o3-mini was expected end of January 2026; o3 followed later.

Alongside o3, OpenAI published deliberative alignment a training paradigm where models are directly taught safety policy text and trained to reason over it before responding. O1 outperformed GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro on both jailbreak resistance AND over-refusal avoidance. That is a genuine Pareto improvement in safety research.


What the Event Signaled

The through-line was diversification across every surface area of the AI stack simultaneously: consumer chat, creative media, developer APIs, OS integration, workspace tools, search, voice, and frontier research. The pricing signal reinforced this a $200/month tier next to a free tier tells you AI is a utility with bandwidth tiers, not a single product.

Beyond the product parade, the event marked a strategic shift. OpenAI was no longer just a model company or a chatbot company. It was positioning itself as a platform with consumer touchpoints, developer infrastructure, creative tools, and safety research all bundled under one brand. The simultaneous launch of a premium tier and a free phone line captured the dual strategy: monetize power users while expanding the funnel at the bottom.


Who Should Care About What

  • Everyday users: ChatGPT Search, Canvas, and Projects were the practical wins. Search killed the knowledge cutoff. Canvas made long-form work bearable. Projects organized the chaos.
  • Creators: Sora was the headline, but its discontinuation 16 months later is a cautionary tale about vendor lock-in on creative tools.
  • Developers: Day 9’s API improvements (o1 with function calling, WebRTC for Realtime, Preference FT, Go/Java SDKs) were the highest-signal announcements for builders.
  • Businesses: Apple Intelligence and Work with Apps pointed toward AI embedded in existing tools rather than standalone apps. That is the adoption pattern to plan around.

One Thing You Cannot Ignore

Sora is dead. Discontinued April 26, 2026 (web/app). API follows September 24, 2026. This rewrites how you should read launch-day enthusiasm: platforms deprecate products. Keep source materials outside any single tool, export regularly, and do not bet a business process on a product younger than 18 months.


Frequently Asked Questions

What was the most important announcement?

ChatGPT Search for everyday users. o1 API with function calling for developers. o3 preview for researchers.

Was Sora free?

Included with ChatGPT Plus and Pro no separate fee. Plus: 50 videos/month at 480p. Pro: 500 at 1080p, no watermark. Now discontinued.

Did all 12 days release finished products?

No. Days 1�11 were launches or rollouts. Day 12 was a research preview. o3 and o3-mini were not publicly released on that date.

What is deliberative alignment?

A training method where models are taught safety policy text and trained to reason over it during inference. It outperformed RLHF on both jailbreak resistance and reducing over-refusals.

Is ChatGPT Pro worth $200/month?

For researchers, yes the “4/4 reliability” on o1 pro mode is measurably better. For casual users, the $20 Plus tier covers almost everything.

What changed after the event?

o3-mini launched late January 2026, o3 followed. By mid-2026, o4-mini was shipping. By 2026, GPT-5.x models absorbed many o-series reasoning capabilities. Sora was discontinued.


Sources


The 12 Days of OpenAI was a moment in time December 2024 and a lot has changed since. Use this recap as a historical reference, not a product manual. Check current OpenAI documentation before making any buying or implementation decisions.

Get our weekly AI digest

The latest AI tools, prompts, and insights — delivered every Tuesday.

No spam. Unsubscribe anytime.

AIUnpacker

AIUnpacker Editorial Team

Verified

A collective of engineers, journalists, and AI practitioners dedicated to providing clear, unbiased analysis of the AI tools shaping tomorrow.