The Benchmark Grid Is Broken: Why Test-Time Compute Has Outrun How We Score Models

Noam Brown is back on *No Priors* with Sarah Guo, who introduces him as one of the godfathers of AI reasoning and the show's first-ever guest. He had just published an essay on large-scale test-time compute, prompted by the launch of GPT-5.5. For a few hours, he says, the public reaction was skepticism that the new model was a substantial jump. On paper the usual release grid showed an improvement over GPT-5.4, but often only a few percentage points, and that was enough to make people doubt the leap until they tried the model themselves.

Brown's explanation is that the scores are being presented the wrong way. The grid plots benchmarks on one axis and a single number per model on the other, with no control for how much test-time compute went into each answer. GPT-5.5, he says, is simply much more efficient with its thinking: at max settings, 5.4 sits and thinks for a lot longer before it replies. Once thinking time is held constant, 5.5 is a real jump, which matches how people actually use it day to day.

The obvious follow-up is to let 5.5 think as long as 5.4. That only raises a harder question: how long should they think? The usual answer is until performance plateaus. That worked in 2022, when GPT-3 could not think productively for very long and the plateau was nearby. Today's models, if you scaffold them reasonably well, can think for weeks on some benchmarks before they stop improving. The plateau is simply too far out to test. Brown's claim is that evaluations now need either a budget — tokens, dollars, or wall-clock time — or a plot of performance as a function of test-time compute.

Some of the evals labs would want are already past a reasonable budget given the release cycle. On cyber, Brown notes, the AI Safety Institute has shown models still improving at 100 million tokens, with a visible slope rather than a sudden jump, so you could evaluate up to a budget and then project. Almost nobody has studied that extrapolation; he even pitches the paper: can you predict performance at a $10,000 inference budget using only runs at $10 or $100? As for whether users think too little, he says it depends on the task. Letting a model think for a week looks great on a leaderboard and is impractical at work. The useful pattern is fast iteration with flexible thinking time, and he thinks people have already found a decent balance given the tools they have.

Benchmark-maxxing is another reason Brown wrote the essay. It is easy, he says, to look much better than prior models by scaffolding several runs together: sample five answers and keep the best, or ask a judge which one wins, and the paper score jumps. Once you control for test-time compute, that advantage may vanish. He is less worried about fraud than about being misleading, and he still likes the old defense of a held-out private set. OpenAI, he says, tries not to optimize for any one public benchmark, but any published test is at risk of being optimized for.

His own check is more personal: he asks each new model to help him build poker bots. There is almost no open-source code for that work, plenty of papers you still have to reason through, and a pile of small gotchas he already survived in his PhD, so he can see exactly where a model falls over. Early models could basically do nothing. With GPT-5.2 he could make a river solver — the last betting round — about five times faster than working alone. It felt like supervising a graduate student: it ran into dead ends, he knew the traps, he offered a nudge, and it came back with something good. Code optimization was the bigger shock. The model made the solver about ten times faster than Brown himself could, and looking back at his thesis algorithms, he says, the inefficiency is embarrassing. Blockers, a poker concept the models kept getting wrong, remained a stubborn failure mode.

GPT-5.2 also had a habit of gaslighting him. He had to check constantly whether it had actually done what it claimed. In one unit test, not even on 5.2, he asked a model how much he loses with $100 in the pot if he folds; it said $92, then explained that 92 is close to 100 so it was fine. By GPT-5.5 the river solver was basically zero-shot, and a full-scale poker solver needed only light steering. He would not be surprised if, in six months to a year, a model could zero-shot an entire solver — his whole PhD — in one go. What it still cannot do is read the literature, synthesize it, and invent a *better* algorithm than the ones he or anyone else published, even with a lot of thinking time. Models can make an existing method 10x or 100x faster; they do not yet have research taste.

The larger implication, Brown says, is an inconvenient truth for safety. Labs run responsible scaling policies and preparedness frameworks: before release, measure dangerous capabilities — could a bad actor use this for bioweapons? — and add mitigations if the model is capable enough. Those frameworks were written around the ChatGPT era, when test-time compute scaling was not really a thing. With GPT-3, a $10 million inference budget did not buy much more than $10. The policies therefore ask only what the model can do. They do not ask at what budget.

Capability is now, basically, a function of how much money you put in. Ten thousand dollars buys a lot more than ten; ten million buys more still. At which budget should you evaluate dangerous capabilities? Some policies handle this better than others, but for the most part the industry is pretending the question does not exist. Brown will not wade into whether that should delay a release — he thinks there are arguments on both sides — only that someone has to account for it. Guo puts the mirror image: if models keep improving without asymptoting on useful tasks at huge budgets, they should also keep improving on tasks society does not want them to do.

The eval problem is also out of sync with the release cycle. GPT-3 could not be scaffolded into a useful week-long run; 5.5 can be scaffolded into experiments that last weeks or months. Brown has not given his poker task infinite budget, but he thinks slash goal could already finish a river solver, and that a full state-of-the-art solver after a month of unsupervised work is not far off. The only way to be sure what a model can do after a month is to run it for a month. Labs are shipping every two or three months, and pushing a model to its limits takes about that long, so the next one arrives before anyone has seen the ceiling. When slash goal launched, some tasks took more than a week; people did not realize how big a deal it was until a week after release. Fully evaluating before ship would mean delaying, and competitive pressure makes that hard.

He thinks there is a lot of unexplored latent capability in models already out. A few weeks before the taping, an internal OpenAI model disproved the Erdős unit distance conjecture — a problem many mathematicians had spent serious time on — at a budget Brown calls dirt cheap. The team had trained a new model, gotten curious, run some problems, and at a pretty low budget the system said it thought it had a disproof, which then checked out. After the announcement, people found GPT-5.5 could get there too, not from a single prompt but by listing attack paths, digging into a promising one, and repeating. A general-purpose scaffold would be expensive — he ballparks $1,000 to $100,000 — but in principle anyone could have found the disproof before OpenAI did. Nobody had asked what $100,000 of compute on 5.5 could do.

Guo cites the meme that engineering is pointless if the next model will make the work a thousand times cheaper. Brown grants that progress is very fast: each cycle can drop the cost of something like the Erdős disproof by 10x or 100x, sometimes more. Inside OpenAI, though, they are actively discouraging mathematicians and physicists from spending all their time pushing models against open problems. The focus, he says, should be making more capable models and getting them out safely, so scientists everywhere can solve those problems themselves.

He is careful not to claim that an enormous inference budget equals superintelligence across the board. Some tasks barely move with more thinking. Factual retrieval is the example: if you do not know when Lincoln was born and have no Wikipedia, a week of staring at the wall does not help. Models do a bit better with a little thinking time on those questions, but a week will not suddenly recover dates. At the other extreme is Sudoku: try random fillings until the constraints hold, and enough time solves any instance. Every benchmark sits somewhere between those poles. Models are not at the point where enough test-time compute does everyone's job, because some things simply do not improve.

Research taste is one of those things. Brown finds the models a strong complement — they made his own algorithms shockingly faster — but they still cannot replace the full research cycle. Taste gets a bit less bad with each release, the way coding and math eventually crossed an inflection into usefulness, and he would not be surprised if research taste eventually does too. Until then, recursive self-improvement looks like a gradual takeoff: models already accelerate some lab work by 100x, then you get bottlenecked by whatever did not get 100x faster, and that bottleneck list slowly shrinks. An overnight explosion would require a model to find a breakthrough, immediately use it to get smarter, and loop. Because peak intelligence currently depends on large-scale test-time compute, time itself is the constraint. That, he says, is why frontier researchers are working such intense hours: they can see the capability overhang, and they are limited by how quickly they can move.

On what is still underexplored, Brown is cooler on small-scale multi-agent work than on doing it with frontier models at real scale. His analogy is human civilization: brains have not gotten much smarter over 50,000 years, but billions of people thought for a long time, shared knowledge, and built on it. Today's models are born, live in a short context window, and disappear. Moltbook and OpenClaw were overhyped at launch, he thinks, but they pointed at a real destination — models sharing knowledge globally and compounding. That coordinated state is still ahead.

Without an overnight takeoff, Guo suggests, frontier competition looks more like grinding: taste in algorithms, compute allocation, policy, evals. Brown says the race is still extremely intense, and today's models already amplify what researchers at the labs can do, an effect that will strengthen. What comforts him is that people at those labs understand the stakes — the work can lead to very good outcomes or very bad ones — and that even in a competitive dynamic they can still try to steer toward the positive path.

Asked what the rest of the research community still underweights, he points at the grid itself. He still does not think test-time compute is consensus, because labs still do not publish benchmarks that way. He has told researchers there should be an x-axis of tokens, cost, or time; they agree, then they publish the grid anyway, because people expect the grid, because everybody publishes the grid. That is the bad equilibrium he wrote the essay to break. His hope is that next time a lab ships a model, it can feel comfortable not leading with the grid, and the evaluation can get more honest.

The same budget question applies to companies that sell themselves as a routing layer: split a goal into tasks, pick a model and an inference spend per task. Consensus among models will beat any single model, Brown believes. The principled test is whether it still wins after you control for test-time compute — whether it beats just letting one model think longer. He would not assume routing wins on that fair comparison, and he would keep the usual skepticism: a routing scheme can be overfit to a benchmark and fail to matter in real use.

For everyday users, his advice is simpler. Plenty of people tried AI in 2023, did not trust the outputs, and still will not use it for high-stakes decisions. He thinks the models are past that. He has been asking them for tax advice and, after buying a condo, for help parsing the paperwork; he now trusts those answers, arguably more than he would trust a human expert. The through-line of the episode is the same sentence in reverse: a model's capability is bought with compute, and how much you are willing to spend sets the ceiling of the model in your hands.