AI usage has shifted from simple back-and-forth chat with models toward agentic systems that can plan and act across tools and files. An agentic system effectively gives the AI a computer it can use to perform multi-hour tasks in one run.
When a chatbot is enough
For low-stakes tasks—getting a recipe, answering a simple question, or drafting a letter—many free or default models are sufficient. But for high-stakes issues, such as a medical or legal second opinion, the author recommends using the most advanced accessible models: Claude’s top models, Opus and Fable, or ChatGPT’s GPT-5.6 Sol, set at least to the “High” thinking level. These models tend to have lower error rates and perform better on complex ability tests, at an increased cost.
Model selection and thinking level
You usually need to pick both a model and a thinking level (or similar setting). Higher levels generally improve accuracy but increase cost.
For real work: two practical choices
The author argues that for most people who want the most from AI today, there are two practical choices: ChatGPT or Claude. Other routes are possible and cheaper, but they require significant technical expertise. Starting around $20/month, ChatGPT and Claude provide easy, powerful access to agentic capabilities.
How to give an AI a “computer”
There are two basic approaches: using the AI company’s virtual computer (simpler but more limited), or granting the AI access to your own computer (more powerful but riskier).
-
Company-provided environments: In ChatGPT this is called "ChatGPT Work" and in Claude it's "Cowork". You pick a model and thinking level (for example, Sol set to High in ChatGPT, or Fable/Opus set to High in Claude) and choose which applications the AI can connect to (email, Google Drive, etc.). These jobs run on the company’s computers and can be started from a phone and checked later.
-
Local-computer modes: ChatGPT’s Codex/Work and Claude’s Code/Cowork let the AI access your computer’s files, run commands, and even control the mouse and browser if you enable that option. These modes support much more complex, long-running projects and provide detailed logs of changes.
Practical examples from the author’s experience
- Demo development: the author built a procedural brutalist city-building demo first with GPT-5 and later with GPT-5.6 Sol; the difference between versions was striking.
- MBA seminar prep: the author asked both systems to connect to Gmail and prepare materials for an MBA seminar on a Monday the 21st. The agents researched, proposed a demo, and drafted responses to colleagues; within about 10 minutes both returned teaching materials and an email draft—work that would otherwise take hours.
- Book proofreading: the author fed the full PDF of a book (due in October) into GPT-5.6 Sol in Codex. The AI worked for 30 minutes, checked 195 references, and produced pages of notes. The author reports that the AI’s notes were accurate and detailed, to the point of being overly nitpicky, and that no hallucinated page numbers or invented text were found.
Permissions and security risks
Permissions matter: whether the AI must ask you before sending email, making purchases, or changing files is configurable. Until you fully trust and understand a system’s failure modes, keep approval required for sensitive actions.
A linked risk is prompt injection: an agent that reads emails or web pages may encounter instructions from others attempting to manipulate it (e.g., a malicious string asking the assistant to forward files). Labs are improving model robustness, but the problem is not solved. These concerns argue for limiting what an agent can access and keeping approval toggles on for sending, spending, or deleting.
Practical conveniences and limits
- Company-run Work and Cowork modes allow you to start long-running tasks from your phone and check results later. You can also schedule recurring briefings.
- Allowing the AI to use your computer enables it to perform tasks exactly as you would (mouse, browser, apps), which can be powerful but increases security implications.
Other platforms and the competitive landscape
- Microsoft Copilot: if your workplace is Microsoft-centric, Copilot is often what’s available. It is serviceable for office documents but lags in agentic capabilities.
- Chinese open-weight models: Kimi K3, DeepSeek, and Qwen are capable but require technical skill to employ as agents.
- Google: once a benchmark leader, Google currently lacks a frontier model and doesn’t offer equivalents to Codex/Code, so the author does not recommend Gemini as a primary system right now. Google does have useful tools: Gemini Notebook (formerly NotebookLM) is valuable for complex research across many sources, and Gemini Omni is a model that can directly see and edit video—allowing interesting transformations in a single prompt.
Multimedia and voice
- Images: Google and ChatGPT have strong built-in image generators; Claude lacks a native image engine and attempts to produce imagery via code, with mixed results. If images matter to your work, this is a relevant factor.
- Voice: ChatGPT’s GPT-Live voice mode enables native listening and speaking with conversational pacing and interruptions; Claude can speak but typically reads generated text aloud rather than engaging in native back-and-forth speech.
Final practical advice
The author’s practical recommendation is consistent: pick Claude or ChatGPT, pay the roughly $20 tier, and give an agent a real task from your life. Carefully review outputs, ask for revisions, and manage the agent as you would a human collaborator. One hands-on experiment will teach you more about what AI means for your work than any guide.
Note: $20 tiers include real but limited agent usage; higher-priced plans mostly buy you more hours of AI labor rather than a smarter model.
