Unique Research · Industry insight · September 13, 2024
The much-watched “Strawberry” model is here. OpenAI says it can handle more complex queries — at a higher cost.
OpenAI surprises with reasoning model o1
In the early hours of September 13, OpenAI unexpectedly launched a new model named o1 — the first in a planned series of “reasoning” models trained to answer more complex questions faster than humans.
Alongside it came o1-mini, a smaller, lower-cost version. Anyone tracking OpenAI closely will recognize o1 as the long-anticipated Strawberry model.
For OpenAI, o1 is not only a step toward human-like artificial intelligence; it also surpasses prior models on coding and multi-step problem solving.
Compared with GPT-4o, however, o1 costs more and responds a bit more slowly. OpenAI therefore framed this release as a “preview” — underscoring that it is still early.
From today, ChatGPT Plus and Team users can use o1-preview and o1-mini; Enterprise and Edu users get access early next week.
OpenAI plans to offer o1-mini to all free ChatGPT users eventually, though no date is set. For developers, o1 is relatively expensive: in the API, o1-preview is $15 per million input tokens and $60 per million output tokens. By comparison, GPT-4o is $5 / $15 per million input/output tokens.
OpenAI research lead Jerry Tworek said o1’s training differs fundamentally from prior products, while keeping exact details opaque. o1, he said, was trained with “a brand-new optimization algorithm and a specially curated training dataset.”
Earlier GPT models were trained to imitate patterns in training data. For o1, the company used reinforcement learning — rewards and penalties (in ML/AI, “penalties” can also mean regularization terms that discourage overfitting or unwanted behavior, pushing the model toward simpler, more generalizable patterns) — so the model learns to solve problems on its own. o1 also uses a chain-of-thought approach, closer to how humans break problems into steps.
That training approach raised OpenAI’s hopes for accuracy. “We’ve noticed this model hallucinates less,” Tworek said — while adding that the problem is not fully solved: “We can’t say we’ve completely solved hallucinations.”
Per OpenAI, the main gap versus GPT-4o is that o1 is stronger on complex problems such as coding and math, and can explain its reasoning.
“This model is definitely better than I am at AP math tests — and I minored in math in college,” said Bob McGrew, OpenAI’s chief research officer. On the qualifying exam for the International Mathematical Olympiad, o1 solved 83% correctly versus 13% for GPT-4o.
On contests hosted on Codeforces, the new model ranked in the top 89% of all participants. OpenAI also claims the next update will reach PhD-student-level performance on complex benchmark tasks in physics, chemistry, and biology.
In many other respects o1 is not as strong as GPT-4o. It is weaker on world-fact knowledge, and cannot browse the web or handle files and images. Still, OpenAI believes o1 represents a new class of capability. The name o1 is meant to “reset the counter to 1” — a new beginning.
“Frankly, I don’t think we’ve traditionally been great at naming,” McGrew admitted. “So I hope this is a first step toward fresher, more sensible names that better tell the world what we’re doing.”
Tworek stressed that OpenAI does not equate how AI models “think” with human thought — but the interface is designed to show the model spending more time and going deeper. “In some ways it feels closer to humans than previous models.”
“You’ll find it alien in many ways, and surprisingly human in others,” McGrew said. Because processing time is limited, it may say something like: “Oh, I’m running out of time — let me give a quick answer.” Early in its chain of thought it may even brainstorm: “I could do this or that — which should I pick?”
Building toward agents
Today’s large language models are not truly intelligent. They mainly predict word sequences from patterns learned in vast data. ChatGPT, for example, has sometimes wrongly claimed that “strawberry” has only two letter R’s — because it did not parse the word’s structure correctly. Notably, the new o1 model gets that right.
Multiple reports say OpenAI is planning to raise more capital at a valuation as high as $150 billion; momentum depends on continued research breakthroughs. OpenAI is folding reasoning into LLMs because it already looks toward a future of systems — or agents — that can decide and act for users.
For AI researchers, reasoning is a key step toward human-like intelligence. If a model can go beyond pattern matching, it may unlock major advances in medicine, engineering, and more. Even so, o1’s reasoning is still relatively slow, not yet agent-efficient, and costly for developers.
“We’ve spent months on reasoning because we believe that’s where the real breakthrough is,” McGrew said. “Fundamentally it’s a new mode of model development — one that can solve truly hard problems, which you need on the path to human-level intelligence.”
ChatGPT milestone timeline
OpenAI announces o1, a model that can fact-check itself
OpenAI released a preview of OpenAI o1 — also known as “Strawberry.” The series is available in ChatGPT and via the API as o1-preview and o1-mini. The company says o1 reasons more effectively through math and science and fact-checks itself by spending more time considering every part of a prompt. Unlike ChatGPT generally, o1 cannot yet browse or analyze files; it is rate-limited and costlier than other models. OpenAI plans free-user access to o1-mini, date TBD.
OpenAI enterprise paid users hit 1 million
OpenAI said paid users of its enterprise ChatGPT offerings — ChatGPT Team, ChatGPT Enterprise, and ChatGPT Edu — have surpassed 1 million. Nearly half of enterprise users are in the United States.
Volkswagen rolls out ChatGPT assistant in the U.S. market
Volkswagen is extending its ChatGPT voice-assistant experiment to cars sold in the U.S. The Plus Speech assistant, built on Cerence’s Chat Pro and OpenAI’s LLM, starts on September 6 for 2025 Jetta and Jetta GLI models.
OpenAI signs content deal with Condé Nast
Under the new agreement, OpenAI will integrate and surface Condé Nast brands — including The New Yorker, Vogue, Vanity Fair, Bon Appétit, and Wired — inside ChatGPT and SearchGPT. CEO Roger Lynch called it a multi-year deal with payment to Condé Nast in some form. A spokesperson told TechCrunch that OpenAI may use Condé Nast content in training.
ChatGPT mobile app records its highest-revenue month
After the GPT-4o “omni” launch, the ChatGPT mobile app posted its highest revenue month since launch. Appfigures data: $28 million net in July from the Apple App Store and Google Play combined.
OpenAI unveils SearchGPT prototype
OpenAI is testing SearchGPT, an AI search engine aimed at challenging Google — “instant answers” from the web plus follow-up questions. The prototype is limited to a small set of users and publishing partners including The Atlantic.
Report: OpenAI could lose up to $5 billion this year
A The Information report, citing undisclosed financials, said high operating costs could leave OpenAI with losses of up to $5 billion this year, and that 2024 ChatGPT training and operations may require as much as $7 billion.
OpenAI releases GPT-4o mini
GPT-4o mini, OpenAI says, beats existing models on cost-efficiency and speed, and tops small-model standards on text and vision reasoning. It replaces GPT-3.5 Turbo as OpenAI’s smallest offering.
ChatGPT launches Mac app
An official macOS app lets users wake ChatGPT with Option + Space, upload files and images, chat on the desktop, and find past conversations.
Apple integrates ChatGPT, including in Siri
At WWDC 2024 Apple said ChatGPT (powered by GPT-4o) will land in Siri and first-party apps later with iOS 18, iPadOS 18, and macOS Sequoia. Free to use without a separate ChatGPT/OpenAI account; Apple device users will also get some features designed for paid ChatGPT users.
OpenAI partners with PwC at enterprise scale
PwC becomes OpenAI’s largest customer to date — serving 100,000 employees — and the first reseller partner for OpenAI’s enterprise products to other companies.
OpenAI training GPT-4’s next flagship
In an announcement for its new Safety and Security Committee, OpenAI said it has begun training the next flagship after GPT-4. The committee will support safety decisions across the product line.
Former board member: board learned of ChatGPT from Twitter
On the TED AI Show podcast, former board member Helen Toner said directors only learned of ChatGPT when it launched in November 2022. She also said Sam Altman gave the board inaccurate information on safety processes and did not disclose his involvement with the OpenAI Startup Fund.
ChatGPT mobile revenue surges after GPT-4o
Even with GPT-4o free on the web, the launch drove record mobile revenue. Mobile users wanting the latest are steered to ChatGPT Plus at $19.99/month.
ChatGPT adds Google Drive and Microsoft OneDrive file intake
Users can upload from Drive and OneDrive, interact with tables and charts, and export custom charts for decks — features expected into GPT-4o over coming weeks.
OpenAI partners with Reddit for training data
OpenAI gains access to Reddit’s “real-time, structured, and unique content.” Reddit content will appear in ChatGPT; the companies will also build AI features for Reddit users and moderators.
OpenAI launches GPT-4o “omni” for ChatGPT
At its spring update, OpenAI showed GPT-4o with a black-hole-like interface and voice/vision interaction reminiscent of the film Her. Rollout to developer and consumer products over following weeks.
Media Manager: opt-out tools for creators
OpenAI said it is building “Media Manager” so creators can control how content is used to train generative models, with an opt-out path — expected ready before 2025.
Content license with the Financial Times
OpenAI will pay for FT content access. ChatGPT users will see attributed summaries, quotes, and rich links to FT reporting on relevant queries.
Tokyo office and Japan-optimized GPT-4
OpenAI opened a Tokyo office and plans a GPT-4 model optimized for Japanese — underscoring localization needs as it expands.
Sam Altman pitches ChatGPT Enterprise to Fortune 500
Reuters reported that in April Altman hosted hundreds of Fortune 500 executives from multiple cities to pitch the enterprise AI service.
“More direct, more concise” GPT-4 Turbo
Paid ChatGPT users (Plus, Team, Enterprise) get an updated GPT-4 Turbo with claimed gains in writing, math, reasoning, and coding, plus a fresher knowledge cutoff.
ChatGPT without an account — with limits
Users can try ChatGPT without signing up, but cannot save or share chats, use custom instructions, or other account features. OpenAI says this version has a “slightly stricter content policy”; details to TechCrunch were unclear.
Report: ChatGPT uses over 500,000 kWh per day
A New Yorker piece said ChatGPT’s daily electricity use to handle nearly 200 million requests is about 17,000× that of a typical U.S. household.
Read-aloud for ChatGPT
Web, iOS, and Android gain read-aloud: replies in one of five voices, across up to 37 languages, on both GPT-4 and GPT-3.5.
Memory controls
OpenAI is testing memory controls for some free and paid users: remember specific facts, review memory, or turn memory off. Deleting a chat does not clear memory — users must delete memories directly.
Temporary Chat
Initially for a small free and paid cohort: chat with no history. Temporary Chat does not use prior memory (custom instructions still apply if enabled). OpenAI may keep copies up to 30 days for safety.
@-mention GPTs in a conversation
Paid users can type “@” to bring other GPTs into a thread; the selected GPT sees full context.
Akutagawa Prize winner: ~5% of winning novel from ChatGPT
Author Rie Kudan said about 5% of Tokyo-to Dojo-to (Sympathy Tower Tokyo) directly quotes ChatGPT — set in a future saturated with AI, fueling debate on AI in literature.
Altman: video coming; GPT-5 preview
On Bill Gates’s Unconfuse Me podcast, Altman confirmed upcoming GPT-5 would support full multimodal interaction — voice, image, code, and video — expected sometime in 2024.
Public input into model governance
OpenAI is forming a team of researchers and engineers to “encode” public feedback on model behavior into products — part of a public program funding experiments toward a “democratic process” for AI rules.
ChatGPT Team for small teams
ChatGPT Team gives workspaces for up to 149 people, admin tools, access to GPT-4, GPT-4 with Vision, and DALL·E 3, plus shared custom GPTs.
GPT Store goes live
After months of prep, the GPT Store opens in a new ChatGPT web tab — GPTs from OpenAI partners and the broader developer community. Access requires Plus, Enterprise, or Team.
FAQs
1. What is ChatGPT and how does it work?
ChatGPT is a versatile chatbot from OpenAI that generates text from user prompts, based on the GPT-4 large language model and deep learning to produce near-human text.
2. When did ChatGPT launch publicly?
November 30, 2022.
3. What does “GPT” stand for?
Generative Pre-trained Transformer.
4. How is ChatGPT different from traditional chatbots?
A chatbot can be any conversational software — including rule-based systems with canned answers. ChatGPT is AI-driven: an LLM generates text from prompts.
5. How does ChatGPT handle data privacy?
OpenAI says individuals in “certain jurisdictions” (e.g. the EU) can object to processing of personal information via a form, including requests to remove AI-generated references about them. Not all requests may be granted; OpenAI says it must balance privacy with free expression under applicable law. The form is titled “OpenAI Personal Data Removal Request.”