On this pageJump to a section
The first thing to know about OpenAI's o1-preview is how few messages you get.
At launch, ChatGPT Plus and Team users were limited to 30 messages a week for o1-preview and 50 for o1-mini. On September 17, OpenAI raised that to 50 queries a week for o1-preview and 50 a day for o1-mini.1 That tells you something about what kind of tool this is. It is not a chatty assistant for every task. It is a slower instrument for a narrower set of jobs.
Running my business solo, I care a lot about which tool does which job. Here is how I would think about reasoning models as an operator, as of late September.
What is o1-preview, and how is it different?
It thinks first and talks second.
OpenAI describes the o1 series as models designed to spend more time thinking before they respond, able to reason through complex tasks and solve harder problems than previous models in science, coding and math.1 Through training, the company says, the models learn to refine their thinking process, try different strategies and recognize their mistakes.1
A faster model for contrast
Anthropic's Claude 3.5 Sonnet, released in June, is a different kind of tool. Anthropic says it operates at twice the speed of Claude 3 Opus and is priced at $3 per million input tokens and $15 per million output tokens, with a 200K token context window.2 It is the one I would reach for when I want a draft now.
What does a reasoning model do well, according to OpenAI?
Hard, structured problems with a checkable answer.
OpenAI reported that in a qualifying exam for the International Mathematics Olympiad, GPT-4o solved 13% of problems while the reasoning model scored 83%, and that its coding ability reached the 89th percentile in Codeforces competitions.1
These are the company's own test results, so treat them as a signal about direction, not a promise about your spreadsheet.
What that means in plain language
If a task looks like a puzzle, with steps, constraints and a right answer you can verify, a model that works through it carefully has an edge. If a task looks like writing, with taste, tone and no single right answer, the extra thinking buys you much less.
What are its limits right now?
OpenAI is unusually direct about this.
For many common cases GPT-4o will be more capable in the near term.
The announcement says that as an early model, o1-preview does not yet have many of the features that make ChatGPT useful, like browsing the web and uploading files and images.1 In the API, access at launch was limited to developers in usage tier 5, with a rate limit of 20 requests per minute, and the API did not include function calling, streaming or system messages.1 Treat all of this as a snapshot. It will change.
The practical costs
Speed and quota. A model that thinks first answers more slowly. A tool you can only message a limited number of times has to be saved for questions that deserve it.
AI Explained, a channel that follows the research closely, described the shift in a video posted on September 18, 2024: "[o1] for me at least represents Paradigm 3 we want to reward answers that are objectively correct."4 That helps explain the limits above. A model trained to reward checkable answers does best where answers can be checked, and the same video notes a regression in personal writing.
Where does a reasoning model earn its keep for an operator?
On decisions where being wrong is expensive and the logic has several moving parts.
- Pricing and margin analysis. Working through scenarios with constraints: costs, capacity, discounts, payment plans.
- Break-even and cash-flow modeling. Checking the logic of a model you built, or building one from described assumptions.
- Planning. Sequencing a launch with dependencies, then asking what breaks if one piece slips.
- Reviewing a process or a contract-like document for gaps and contradictions. As a first pass for you to check, not a legal opinion.
- Debugging logic. Spreadsheet formulas, automation rules, conditional logic in a funnel.
How I would use it
Give it the full set of facts, ask it to show its reasoning and then check the arithmetic and the assumptions yourself. A reasoning model is still a model. It can be confidently wrong, and a longer chain of reasoning is not proof.
Where is a fast model still the better tool?
Almost everywhere else.
- Drafting and rewriting. Emails, posts, scripts, first drafts of SOPs.
- Summaries. Calls, long documents, meeting notes.
- Classification and extraction. Sorting leads, tagging feedback, pulling fields out of text.
- High-volume, low-stakes tasks. Anything you would hand to a junior assistant.
- Anything needing speed or files. When you need an answer in seconds, or need to upload a document, the fast model fits better today.
| Reach for a reasoning model | Reach for a fast model |
|---|---|
| Multi-step analysis with constraints | Drafting and rewriting copy |
| Checking logic or finding contradictions | Summarizing calls and documents |
| Planning with dependencies | Sorting, tagging and extracting |
| A wrong answer is costly | Output gets reviewed anyway |
| You can wait and have quota to spare | You need it now or in volume |
How do you choose in thirty seconds?
- Is there a right answer I can check? If yes, a reasoning model may help. If it is taste, use the fast one.
- Does it have several steps with constraints? More steps favor the reasoning model.
- What does a mistake cost? If it is cheap, go fast. If it is expensive, slow down and verify.
- Do I need files or the web? Today, that points to the fast model.
- Run the same prompt in both on a real task once, and keep the notes.
What about the claims you make about AI?
Be careful. The FTC just made that louder.
On September 25, 2024, the FTC announced a set of law enforcement actions called Operation AI Comply, against companies using AI hype or selling AI tools in deceptive ways, including a scheme that took at least $25 million from consumers by promising AI-powered passive income. Chair Lina Khan said there is no AI exemption from the laws on the books.3 If you sell with AI in the story, do not promise results you cannot back up, which is the same logic as the FTC's new fake review rule. And keep the copy yours: buyers can already smell generic AI marketing.
There is no AI exemption from the laws on the books.
Reasoning models are interesting and they will improve quickly. For now, treat them as a specialist, not a replacement. Use them for the hard thinking, use the fast models for the daily work and put your own judgment on top. If you want to work out where AI fits in your business, let us talk.
Keep building.Ekai
Questions people ask
What is OpenAI o1-preview?
o1-preview is the first of OpenAI's o1 series of reasoning models, released September 12, 2024. OpenAI says the models are trained to spend more time thinking before they respond, and are aimed at harder problems in science, coding and math.
Is o1 better than GPT-4o or Claude 3.5 Sonnet?
Not for everything. OpenAI says that for many common cases GPT-4o will be more capable in the near term, because o1-preview lacks features like browsing and file uploads. For complex multi-step reasoning, o1 may help. Test both on your own tasks.
How many o1 messages can I send?
At launch ChatGPT Plus and Team users had weekly limits of 30 messages for o1-preview and 50 for o1-mini. On September 17, OpenAI raised them to 50 queries a week for o1-preview and 50 a day for o1-mini.
Should a small business use a reasoning model?
For analysis, planning and checking multi-step logic, possibly. For routine drafting, summarizing and formatting, a faster model is usually the better fit. Choose by task, and keep a human checking anything that spends money or reaches a customer.
Sources
- 1
- 2
- 3
- 4
