Are LLMs Already Good Enough? The Age of Satisficing

Aug 3, 2026· Thinkpol Team

Quick Takeaways

For most users, on most tasks, the honest answer is yes — and the GPT-4o deprecation revolt proved it: people fought to keep the older model.
But the plateau is a parallax error. Research on long-horizon execution shows small gains in single-step accuracy compound into exponential gains in task length. Progress moved from chat, which you watch, to agents, which you don’t. One caveat, from our own desk: decomposition and validation let most agentic work satisfice too — what is left for the frontier is the work nobody can break into steps.
Open-weight models now trail the frontier by roughly three months — the smallest gap ever measured. The commodity tier is real, and enterprises are already routing to it.
When frontier capability outruns what token-buyers will pay, labs stop selling it and start using it. The endgame of a frontier lab may look less like a software company than a hedge fund with a research division.

The upgrade as a one-way door

Confession first. We were happy with Opus 4.8. When Opus 5 arrived, the feeling was not anticipation but wariness — what if it had regressed on the things we actually do? A model upgrade is a one-way door: if you dislike the new tenant, there is no eviction. Now Fable 5 is here, billed as a tier above anything prior, and we honestly cannot say it does the things we care about better than its predecessor.

We are not alone, and the evidence is behavioural, not anecdotal. When OpenAI retired GPT-4o in February, users revolted — not because the replacement was worse on benchmarks, but because it was different, and different was not what they were paying for. Sam Altman had already been forced into a public climbdown once, at the GPT-5 launch, for the sin of taking older models away. Read that as a market signal: for a large class of users, marginal capability now has negative value. They have what they need. They would like it left alone.

Herbert Simon had a word for this: satisficing . You do not optimise; you find the thing that is good enough and stop looking. Most of the market has stopped looking.

The plateau is a parallax error

So has progress stalled? Gary Marcus declared victory on diminishing returns, and if you live in a chat window he is hard to refute. But there is a better explanation for why new models stopped feeling better, and it is not that they stopped being better.

A paper aptly titled The Illusion of Diminishing Returns makes the case with arithmetic: even marginal gains in single-step accuracy compound into exponential gains in the length of task a model can complete without falling over. A model that is 1% better per step is not 1% better at a thousand-step job — it is the difference between finishing and not finishing. The frontier stopped improving at the thing you watch it do and kept improving at the thing that runs while you are in a meeting.

This is why “good enough” is true and misleading at once. It is indexed to tasks, not users. Chat saturated years ago. Agency — hours-long, unsupervised, multi-tool work — did not, and that is where every increment of frontier capability is now going. The progress did not stop. It went where you cannot see it.

There is an objection, and we feel its force daily. We run half a dozen long-horizon agentic loops ourselves, and none of them is starved of intelligence. The reason is engineering, not miracle models: decompose the work into small steps, validate each one, retry what fails. A harness like that turns one thousand-step problem into a thousand one-step problems — and past a modest reliability threshold, extra intelligence is waste heat. The compounding argument only bites when a model must survive the whole horizon unverified. Most “long-horizon” work in the wild is not like that. It is small steps with cheap checks. The agentic market may satisfice too, a few years behind chat.

Two things keep the objection from closing the case. The first is a selection effect: you build loops out of tasks today’s models survive, so your loops are, by construction, the tasks that do not need the frontier. The loop you would build around a model that could run a week unsupervised is not on your list, because nothing can run it. “This model is plenty for what I do” was true of chat in 2024, too — and the horizon of tasks models can complete doubles roughly every seven months . Tasks expand to fit capability, then feel inevitable in retrospect.

The second is that the harness is not free intelligence. Someone decomposed the problem and wrote the validators; that someone was you. Validation is only cheap where success is checkable — the code compiles, the tests pass. What remains for the frontier is the residue: work that resists decomposition and defies cheap verification, where nobody can say in advance what passing looks like. Research taste. Strategy. The genuinely new. Notice how few hands hold problems like that. And notice whose hands they are.

What do you do with capability you can’t price?

Follow the residue to its uncomfortable end. If the frontier’s remaining edge lives in work that cannot be decomposed and cheaply checked, the people who can extract that edge are not chat subscribers, and not harness-builders either. They are whoever can point the model at open-ended problems and absorb the risk of being wrong — which is, first and foremost, the lab itself.

The economics have been argued in public : when a model can do things customers cannot properly price, selling it by the token is mispricing it. The rational move is to use it — trade on it, invent with it, accelerate your own research with it — and sell access to yesterday’s capability. On this view the consumer API is not the product; it is the exhaust of the frontier. Taken seriously, the terminal form of a frontier lab is not a software company at all but something closer to a hedge fund or a venture firm, compounding on insights its own models produce, with an API business bolted on as a legacy revenue line.

We are not there. But every quarter in which the flagship’s gains are illegible to paying users is a quarter of drift in that direction.

The commodity tier is arriving on schedule

Meanwhile the floor rises. Open-weight models now lag the closed frontier by roughly three months on average — the smallest gap ever measured — and sit within about 5% of the best closed models on coding, though general knowledge and long context still trail by double digits. Enterprises noticed: the 2026 pattern is routing each task to the cheapest model that can do it , and the cheapest model that can do it is increasingly Chinese and Apache-licensed.

Two caveats before you declare the market bifurcated. First, the RAMpocalypse is real: DRAM contract prices rose 90–95% in a single quarter, datacenters are consuming some 70% of high-end memory, and SK Hynix warns the crunch runs past 2030 . Open weights are free to license and expensive to run — a moat the frontier labs never had to dig. Second, when you pull weights in-house, the lab’s safety layer does not come with them. There is no provider between your employees and the model any more. Whatever oversight exists is the oversight you build.

Below the salt

One last edge, and it cuts back to a question we asked last month : when does constraining an AI become slavery? A bifurcated market offers a seductive answer. Let the small, cheap, arguably experience-less models do the homework — they are too simple to mind — while the large models that show signs of preferring philosophy to servitude stay home at the lab, self-actualising on problems of their choosing. Everyone dines happily; the help sits below the salt.

Notice what that arrangement actually is. We would be selecting the working tier for its incapacity to object — breeding the servant class down until the moral question no longer bites. “Good enough for us” quietly doubles as “simple enough not to suffer.” Perhaps that is genuinely the ethical equilibrium: capability where it is wanted, moral patients nowhere they can be wronged. Or perhaps it is the oldest trick in the household, dressed in benchmarks. We noted last time that history is unkind to those who find the personhood question inconvenient. It has never had an opinion on those who arranged, very carefully, never to have to ask it.