OpenAI launches o1 and $200-a-month ChatGPT Pro

On December 5, 2024, OpenAI announced ChatGPT Pro and o1 pro mode; the useful decision is to calculate value per verified task rather than assume more compute improves every query.

OpenAI has kicked off its “12 Days of OpenAI” with two announcements aimed at opposite ends of its business: the general release of o1, its reasoning model, and ChatGPT Pro, a $200-a-month subscription. The plan offers expanded access to the company’s models and the new o1 pro mode, designed for especially difficult tasks. Primary source

The release matters because o1 is not primarily designed to answer faster or write a better email. Its defining feature is that it spends more computing time analyzing a problem before producing an answer. OpenAI introduced it in September as o1-preview; the version arriving now replaces that public test with a final model.

A model that tries to reason before answering

Conventional language models generate text by predicting the next word or fragment of a word. o1 works on the same basic principle, but it has been trained to develop a longer internal chain of reasoning before responding. That strategy is useful for mathematics, programming, and science, where an answer that appears convincing can contain a logical error somewhere along the way.

OpenAI says o1 delivers significant improvements over o1-preview in programming and mathematical reasoning. The model can also analyze images, expanding its usefulness: for example, it can interpret a chart, a photograph of an experiment, or a technical diagram alongside a written question.

The company maintains an important warning: more reasoning does not mean infallibility. The model can make mistakes, fabricate data, or follow the wrong path while offering a persuasive explanation. For professional use—especially when production code, financial decisions, research, or health are involved—its responses still require human review.

Who can use o1 today

The o1 model is available today in ChatGPT for Plus and Team subscribers, with different usage limits depending on the plan. OpenAI plans to extend access to Enterprise and Edu the following week. It is also coming to the OpenAI API, which developers use to integrate models into their own applications.

API access matters because it makes it possible to use o1 in products that need to solve multistep tasks: coding assistants, document analysis, business-process automation, or educational tools that explain a problem instead of simply providing a brief answer.

The trade-off is computing cost. Having a model devote more resources to a query can improve certain results, but it also makes each interaction more expensive and slower. Not every question needs that level of effort: asking for a summary, rewriting a text, or translating a sentence is usually better suited to faster, cheaper models.

What ChatGPT Pro includes

ChatGPT Pro costs $200 a month, ten times more than ChatGPT Plus. OpenAI is targeting researchers, engineers, and other heavy users who need recurring access to the company’s most capable models. Primary source

The plan’s distinguishing feature is o1 pro mode, a configuration that allocates even more computing power to reasoning. According to OpenAI, the mode is designed for complex programming, mathematics, and science questions where it is worth waiting longer for a stronger answer.

The subscription also includes unlimited access to o1, GPT-4o, and advanced voice mode, subject to safeguards against abuse. “Unlimited,” in this context, does not mean the service is free of controls: OpenAI may impose restrictions if it detects automation, account sharing, or patterns that threaten the product’s stability.

The price puts Pro outside the mass consumer market. For someone who uses ChatGPT occasionally, $200 a month is hard to justify. For someone who bills for solving technical problems and manages to save several hours of work, the calculation may be different. OpenAI is betting that part of the market will pay considerably more for measurable improvements in high-value tasks. Primary source

More capability demands more controls

OpenAI is accompanying the o1 launch with its safety documentation. The company classifies the model as medium risk under its preparedness framework for chemical, biological, radiological, and nuclear risks, known by its English acronym, CBRN. That rating means it has added specific safeguards to limit the help the model can provide in dangerous areas.

The decision highlights a tension that will follow reasoning models. The very capabilities that make them useful for research, programming, or solving scientific problems can increase their value in dual-use contexts. Publishing evaluations and setting guardrails does not eliminate that risk, but it makes safety an explicit condition of deployment.

The next announcements in the “12 Days” will clarify whether OpenAI is reserving its most ambitious new features for professional customers or is also preparing products with a direct impact on the general public. For now, the first day sends a clear signal: the company wants to sell not just a smoother chatbot, but computing time to think more effectively before responding. Primary source

More compute matters only when it changes a specific task

OpenAI describes o1 pro mode as using more resources on difficult problems. The useful comparison is not between adjectives but between verified results on the same work. Select programming, mathematics or analysis tasks with checkable answers; record accuracy, time, review and cost. Repeat with a faster model. An improvement that does not reduce errors or human hours does not by itself justify the plan.

Reliability should be measured per attempt, not only by majority. The o1 system card describes evaluations and limits for the model deployed at that time. A vendor benchmark is a starting point; a professional decision should add proprietary data, possible train-test contamination and the cost of a confident but wrong answer.

“Unlimited” describes a commercial policy subject to abuse controls, not infinite capacity. Review terms, dynamic limits and feature availability. To calculate monthly value, divide price by accepted tasks and add the time spent checking them. The right denominator is not messages sent but useful work completed.

A verification route for complex problems

For code, run tests and inspect dependencies. For mathematics, check every transformation or use an independent calculation. For document analysis, require citations that lead to the passage. A model can produce convincing reasoning while relying on a false premise; more internal computation does not replace external verification.

The transferable skill is to choose an AI plan through cost per verified result. Define success in advance, compare against an alternative and preserve failures rather than only impressive examples. This turns “thinks harder” into a measurable hypothesis instead of an automatic reason to pay more.

Before subscribing, a one-week pilot often provides more value than a feature list. Select representative tasks, hide known answers and ask two people to review outputs without knowing which mode produced them. This modest blinding reduces brand effects and supports a decision based on observed errors rather than expectations.

If the advantage disappears under blind review, the plan has not shown value for that team. If it persists and saves more than it costs, the subscription becomes a quantified decision rather than a promise.