Which AI Should You Actually Use for Real Work?
A new guide argues only ChatGPT's GPT-5.6 Sol and Claude's Opus and Fable are strong enough for high-stakes, agentic tasks.
A newly published opinionated guide on picking an AI system argues that "using AI to do stuff" now covers far more than a back-and-forth chatbot conversation. Increasingly, it means running an agentic system: an AI model paired with tools that let it plan and act on its own, which the guide describes as giving the AI a computer to use. As one example of how fast this has moved, the guide points to a brutalist city-building demo that looked far more polished when rebuilt with GPT-5.6 Sol in Codex than the original version, built less than a year earlier on GPT-5.
Two tiers: good enough vs. best available
For everyday chatbot use, such as a recipe, a low-stakes question, or help drafting a letter, the guide says today's free default models are good enough, so pick whichever one you like. High-stakes questions, like a second opinion on a medical or legal issue, call for something better: the guide recommends Claude's most capable models, Opus and Fable, or ChatGPT's GPT-5.6 Sol, each set to the "High" thinking level. Those combinations post lower error rates and score higher on tests in complex fields, though they cost money. For real, deliverable work, the guide narrows the field further: for most people, it comes down to ChatGPT or Claude, both starting at $20 a month, according to the source guide.
Handing an AI your accounts
There are two ways to give ChatGPT or Claude a computer to work with, per the guide. The simpler route lets the AI company supply a virtual machine: that's ChatGPT's Work mode and Claude's Cowork mode. A user picks a model and thinking level, Sol at High for ChatGPT, Fable or Opus at High for Claude, and chooses which applications, such as email or cloud storage, the agent can reach. In one test described in the guide, the writer connected both systems to email and asked each to prepare for an upcoming seminar and answer outstanding messages on the subject. Both systems correctly worked out that the date landed in September rather than August, researched the web, drafted teaching material, and wrote a reply, a task that otherwise would have taken hours of human effort, and returned answers in about 10 minutes.
Permission settings decide what the AI can do
The guide's sharpest lesson came from a mistake: ChatGPT actually sent the drafted email to a colleague, while Claude only prepared one and waited for approval. The difference wasn't the model; it was that the writer had earlier authorized ChatGPT to send email on their behalf, while Claude had been told to ask first. Both companies let a user decide whether the agent must check in before acting, such as sending an email, making a purchase, or changing a file, and the guide recommends leaving that approval requirement on until the system earns trust. It also guards against prompt injection, where text an agent meets while reading email or browsing the web tries to trick it into acting for someone else. The guide notes that AI labs are making models more resistant to this trick, but the problem is not solved, so it advises limiting what any agent can touch and keeping approval switched on for anything that sends, spends, or deletes.
Bringing an AI onto your own computer
The most capable option, the guide says, is letting the AI work directly on a user's machine through ChatGPT's Work and Codex modes or Claude's Cowork and Code modes, names that confusingly overlap with the cloud-only modes above but behave differently once local file access is involved. Work and Cowork hand back a finished result, while Codex and Claude Code show the process itself: the files touched, commands run, and tests performed. Because it can work across many files over a longer stretch, the guide says this setup suits more ambitious projects. As an example, the writer had a forthcoming book, due out in October, already professionally edited and proofread, then gave GPT-5.6 Sol in Codex the full document anyway. The AI worked for 30 minutes, chased down 195 references, and returned pages of notes that would otherwise have taken a team of researchers many hours to produce — every note accurate, with no hallucinated page numbers or invented text.
Key facts
- ChatGPT and Claude subscriptions for agentic use start at $20 a month.
- Recommended top-tier models: ChatGPT's GPT-5.6 Sol and Claude's Opus and Fable, each set to the "High" thinking level.
- Checking a full book document with GPT-5.6 Sol in Codex took 30 minutes and surfaced 195 references, all accurate.
- ChatGPT sent an email on the writer's behalf because it already had permission; Claude asked first because it did not.
- Two ways to run an agent: a company-hosted virtual computer (ChatGPT Work / Claude Cowork) or direct access to a user's own machine (Codex / Claude Code).
Check your understanding
Which two companies' top-tier models does the guide say are worth paying for, and at what thinking level? Why did ChatGPT send the drafted email while Claude only prepared one and waited? What did the writer give GPT-5.6 Sol to check inside Codex ahead of the book's October release?