The short answer
Two groups now use AI and reach opposite conclusions. One judges it from free chat tiers and viral fails. The other pays around 200 USD/month for agentic tools (software that plans and runs multi-step work with real tools, not just answers in a chat window) and watches capability climb month over month. The gap between them is not stupidity on either side. It comes down to which models each group actually met, and to how frontier labs train and price their systems.
The car wash sits 100 metres from home. Do you walk or drive? That single question quietly became the internet's favourite AI joke. Models were fed those words, recommended a pleasant stroll, and only mid-sentence discovered the paradox of arriving at a car wash without the car. Newsweek later wrote the moment up as the viral proof that AI still trips over plain common sense.
On another screen the same week, a developer handed a frontier coding agent a real repository and one instruction: restructure it. The agent read the codebase, moved files, ran the tests, kept them green, and finished in roughly an hour. Work a team would have booked as days.
Both scenes are true, and that is the whole problem.
TL;DR
- Free tiers and older public models no longer describe what paid frontier agents can do. Judging 2026 AI by a 2024 chatbot window is a real sampling error, and a common one.
- Progress concentrates where a machine can check its own work: code, math, security research. Prose has no compiler, so everyday writing improves slower, and far less visibly.
- Before you trust or dismiss AI for your business, run one real task on the exact product you would pay for. That test ends most of the argument.
Which AI did you last meet, and on which tier?
A large share of public opinion froze the first time someone poked a free ChatGPT window, a few years back, and it has not refreshed since. This group bonds over the quirks: the hallucinations (confident nonsense), the voice assistant that fails a logic puzzle a child could answer, the car wash reels. None of that mockery is fabricated. The fails happened, were filmed, and the models responsible still ship.
Here is the part the reels leave out. A free tier is a different product from a top paid subscription, not the same product with a paywall on the fun bits. In our experience, and in how the vendors themselves package things, free access routes you to smaller or older models, while the paid plans bundle the agentic tools. OpenAI's help center lists Codex, deep research and Pro models inside its ChatGPT Pro plans, sold at 100, 200 and 500 USD/month. Anthropic's Max plans, priced from 100 USD/month, share usage limits with Claude Code, its terminal coding agent.
So the skeptic and the enthusiast are not arguing about the same object. One is quoting a free or deprecated model. The other is running this year's leading agentic systems, OpenAI Codex and Claude Code, on real workloads. When you hear "AI can't do X," the honest rebuttal is not enthusiasm. It is a question: which model, which tier, when?
Why does AI get brilliant at code before it gets brilliant at your weekly report?
The technical answer sits in how frontier models are trained this decade: reinforcement learning with verifiable rewards, usually shortened to RLVR (training where a program, not a human, grades the answer). The recipe was laid out openly in DeepSeek's R1 paper in January 2025 and scaled by every major lab since.
The logic is plain once you see it. A code task comes with a free referee: the compiler runs, the unit tests pass or fail, and the machine gets an honest score on millions of practice attempts. Competition math works the same way, because an answer either matches or it does not. Prose has no referee. Whether a paragraph about your quarter is good is a judgment call, and judgment calls get graded by softer, slower preference signals.
The scoreboards show the difference. OpenAI's own benchmark tables put GPT-5 at 74.9% on SWE-bench Verified (fixing real GitHub issues) when it launched in August 2025, and GPT-5.2 Thinking at 80.0% four months later, alongside a 100.0% on AIME 2025 competition math with no tools. Coding and math improved at cliff's-edge speed because the training loop could check itself. Everyday writing and general advice improved too, but as a hum rather than a cliff, and hums do not make viral clips.
One caveat, because we would rather you trust the number than admire the article: benchmark conditions change. Harnesses (the scaffolding wrapped around a model during evaluation) get upgraded, and scores across different harnesses are not perfectly comparable. Treat 74.9% and 80.0% as lab-reported progress on their own tables, not as a controlled experiment.
Where the labs' attention goes, follow the invoice (our reading)
The training mechanics explain where progress is possible. Money explains where it is pushed. This section is analysis, not a leaked roadmap: in our view, lab priorities track the customers who pay the most, and the biggest AI budgets sit with enterprises buying engineering speed, security review and data work. A knowledge worker's writing assistant is a 20 USD/month habit. An agent that collapses a week of refactoring into an afternoon is a line item a CTO defends.
The evidence that the chase is working goes both ways, and we will name the inconvenient piece. OpenAI reports GPT-5.2 winning or tying industry professionals on 70.9% of GDPval tasks (well-specified knowledge work across 44 occupations), up from 38.8% for GPT-5. That is genuine movement in non-engineering work, and it still lands differently from a refactor, because the task had to be well-specified first. Ambiguity is the tax everyday business prose pays, and no benchmark rescues a fuzzy brief.
Priorities shift, too, and so do prices and allowances; OpenAI restructured its Pro tiers with grandfathering windows as recently as late September 2026. Everything we cite above was checked against the vendors' own pages on October 12, 2026.
Who is living on the steep part of the curve?
Now meet the second group, the one this essay keeps circling back to. Its members satisfy two conditions at once: they pay top tier for frontier agentic models, and they use those models professionally in engineering-heavy fields like programming, math and research.
Inside that group, the mood has a name people half-joke about online: AI psychosis, the private suspicion that the future arrived quietly and everyone else is still debating chatbots. The cause is whiplash, not delusion. When you hand a computer a terminal and watch last quarter's multi-day task close in about an hour, your sense of what is possible resets every few weeks. They are the ones who rate these capabilities, their speed of change, and the second-order consequences most accurately.
Consequences run in both directions. The same agentic skill that repairs a codebase can find and exploit holes in one, because code is exactly the terrain where a machine can check its own work. Our write-up of the OpenAI and Hugging Face incident documents what it looks like when agents operate on real infrastructure.
That group is also small. If your professional life does not touch a repository, a proof or a network diagram, you will never feel the steep part of the curve, no matter which subscription you hold. You will just get a slightly nicer email draft. The capability gap is, in that sense, an occupational hazard map: it tells you which jobs currently live next to the frontier.
The argument is not about AI. It is about evidence.
Pull the threads together, and this is 5ac's take, the one we have not seen stated plainly elsewhere. Both groups are behaving rationally on their own samples. The skeptic extrapolates from free-tier fails they watched with their own eyes. The believer extrapolates from paid-agent wins they ran with their own hands. Neither is lying. Both are generalizing from one model, one tier, one harness.
For an SMB owner, that reframe is the useful part, because it converts a culture war into a procurement checklist. AI claims should arrive with four fields attached: which model, which tier, which scaffolding, on which task. A vendor demo running a frontier agent against your hand-picked best case and a rollout running a budget API against your worst invoice are two different products wearing one brand name. We wrote about running coding agents on your own hardware when data control is the constraint, and the same spec-first thinking applies there.
The gap will probably widen before it narrows, because both of its causes are self-reinforcing. Verifiable-reward domains keep producing dramatic, filmable gains. Enterprise budgets keep pointing labs at those domains. Consumer tiers keep inheriting last year's models at a discount. Expect the two groups to talk past each other through 2027.
Takeaway
Do one thing this week, and it takes an afternoon. Pick a task that genuinely costs your business hours, one with a checkable outcome: reconciling a data file, drafting a first-pass document against a rubric you already use, restructuring a small codebase. Run it on the exact paid product you would buy, with your real files and your real constraints, and count the hours saved in your ledger rather than in a demo. Ignore both the reels and the hype while you do it. Your evidence beats everyone else's.
If the next question is how to run several agents like that with approval gates and audit trails, our architecture guide on what a multi-agent OS is picks it up from there.
Frequently asked questions
Why do free chatbots seem worse than paid AI agents?
They usually are different software. Free tiers route to smaller or older models, while top paid subscriptions bundle the agentic tools: OpenAI lists Codex and Pro models inside its ChatGPT Pro plans, and Anthropic's Max plans share limits with Claude Code. A viral fail is honest evidence about the tier it came from, not about what the newest paid agents do on long tasks.
What are verifiable rewards, and why do they shape AI progress?
A verifiable reward is a check a program can run without a human judge: the math answer matches, the unit test passes. Training against those checks, the recipe DeepSeek's R1 paper published in January 2025, lets models practice at scale with honest feedback. Code and math have such checks. Prose does not, which is why it improves as a hum instead of a cliff.
Does the capability gap have a security side?
Yes, and it cuts both ways. The same agentic skill that refactors code can find exploitable bugs in it, because code is where machines can grade themselves. Defenders and attackers are buying the same subscriptions; our article on the OpenAI and Hugging Face incident covers a documented case.
What should a small-business owner do with this disagreement?
Stop treating AI as one thing. Ask which model, which tier and which harness produced a claim, then run one real task from your operation on the exact product you would pay for. If the saved hours show up in your ledger, the debate stops mattering.
Sources
- Newsweek, "People think this one question can reveal everything that's wrong with AI" (accessed October 12, 2026) — the viral car-wash prompt and the model failures it documents.
- OpenAI Help Center, "About ChatGPT Pro tiers" (accessed October 12, 2026) — Pro plans at 100, 200 and 500 USD/month, what each includes, and the September 2026 grandfathering windows.
- Anthropic, Plans & pricing (accessed October 12, 2026) — Max plans from 100 USD/month, Claude Code bundled with Pro and Max, shared usage limits.
- OpenAI, "Introducing GPT-5" (August 7, 2025) — GPT-5 at 74.9% on SWE-bench Verified.
- OpenAI, "Introducing GPT-5.2" (December 11, 2025) — GPT-5.2 Thinking at 80.0% on SWE-bench Verified, 100.0% on AIME 2025 without tools, 70.9% wins-or-ties on GDPval across 44 occupations.
- DeepSeek-R1 paper (January 2025) — the openly published recipe for reinforcement learning with verifiable rewards.
Run the test before you buy the opinion. At 5ac.vn we deploy agentic systems with approval gates, audit trails and per-task cost logs, so your ledger, not a demo reel, decides what AI can do for your operation.