Benchmark grids that report a single score per model hide the test-time compute budget spent, making model comparisons misleading.
When GPT-5.5 launched, its benchmark gains over 5.4 looked small, prompting public skepticism. Brown says 5.5 is actually far more efficient per unit of thinking time; 5.4 was simply allowed to think longer. Once compute is held constant, 5.5 is a substantial jump, which matches users' actual day-to-day experience of the model.
benchmark-evaluation
The point where model performance on a benchmark plateaus has moved so far out that it is no longer practical to evaluate to it.
With GPT-3-era models, performance saturated quickly, so labs could just run to the plateau. Modern models, scaffolded well, can keep improving on some benchmarks for weeks or months of test-time compute, so evaluators must instead fix a budget (tokens, cost, or time) and compare performance at that budget rather than waiting for saturation.
test-time-compute
Safety and preparedness frameworks do not specify what inference budget a model should be evaluated at, leaving a real policy gap.
Responsible scaling policies and preparedness frameworks were built in the pre-test-time-compute era and measure a model's raw capability rather than its capability as a function of dollars or tokens spent. Since a model can do far more with a $10 million budget than a $10 budget, dangerous-capability evaluations that ignore budget may understate what a well-resourced bad actor could extract from a released model.
ai-safety-evaluations
Fully evaluating a model's capability may require running it for as long as the task itself would take, which conflicts with a 2-3 month release cadence.
Brown argues the only way to know what a model can do after running for a month is to actually run it for a month, and the same logic extends to six months or a year. Because labs ship new models every couple of months, nobody has actually pushed a model to its true ceiling before the next one replaces it, so the real capability frontier is unknown even to the labs.
test-time-compute
OpenAI's internal model disproved a long-standing open math problem (the Erdos unit distance conjecture) cheaply, and the same result was later reproducible via public GPT-5.5 with the right scaffolding.
The disproof was produced at low cost with minimal effort, mostly out of curiosity after training a new model. After OpenAI announced it, outside users found they could get GPT-5.5 to the same result by asking it to enumerate strategies and then investigate the promising ones iteratively, though a general-purpose scaffold to do this reliably would cost an estimated $1,000 to $100,000.
test-time-compute
The cost of reproducing a frontier research result drops roughly 10-100x with each model release cycle, which creates pressure to wait for the next model rather than do the engineering work now.
Brown references the meme of skipping current engineering effort and waiting a couple of months for the next model to make the same task an order of magnitude cheaper. He says OpenAI deliberately discourages researchers from spending all their time mining current models for one-off wins (like solving open math or physics problems) and instead wants them focused on building more capable and cost-effective models.
recursive-self-improvement
Brown does not believe an overnight, self-reinforcing intelligence explosion is near, because capability gains are bottlenecked by the time it takes to run large-scale test-time compute.
The hard-takeoff hypothesis assumes a model could discover a breakthrough and immediately compound it into further breakthroughs. Brown argues that because unlocking a model's full capability requires long stretches of test-time compute, progress is rate-limited by wall-clock time even as research acceleration inside labs increases, producing a gradual rather than instantaneous takeoff.
recursive-self-improvement
Current models can dramatically optimize and accelerate existing research work but still lack 'research taste' - the ability to originate genuinely novel directions.
Using his own poker-solver PhD work as a test case, Brown found GPT-5.x models could make his algorithms 10-100x faster and build a full river solver largely unassisted, but when asked to synthesize published literature and propose a better algorithm than any human had published, the model could not do it even given a long time to work.
recursive-self-improvement
Benchmark-maxing techniques like running a model multiple times and taking the best (or judge-selected) response inflate apparent capability without a real per-compute improvement.
Brown says this kind of scaffolding trick is easy to apply and produces headline-grabbing benchmark jumps, but is misleading because it does not control for the extra test-time compute spent generating and judging multiple attempts. He says OpenAI tries to avoid optimizing directly for public benchmarks and keeps a held-out private eval set to reduce this risk.
benchmark-evaluation
Model reliability on his poker-bot evals moved from confidently wrong ('gaslighting') answers to accurate zero-shot performance within a few model generations.
Brown describes an earlier model insisting that folding away $100 from a pot only lost $92, and downplaying the discrepancy when challenged. By GPT-5.5, the model could handle equivalent tasks essentially zero-shot with only light steering, a qualitative jump he treats as more informative than benchmark scores.
test-time-compute
Routing and consensus layers built on top of frontier models should be judged against the counterfactual of giving one model the same total compute budget to think longer, not against a raw model run once.
Some vendors position their value as choosing which model or how much inference to allocate per subtask. Brown agrees multi-model consensus can beat any single model's raw output, but argues the real test is whether it beats a single model given equivalent test-time compute, and whether the gains hold up on real-world tasks rather than just on benchmarks that may have been implicitly optimized for.
benchmark-evaluation
Multi-agent systems that accumulate and share knowledge across instances, rather than existing in isolated short context windows, are an underexplored frontier.
Brown compares today's AI agents unfavorably to human civilization, where accumulated and shared knowledge across billions of people over millennia, not individual intelligence gains, drove capability growth. Individual AI agents are 'born' into a short context window and then disappear; early multi-agent coordination experiments hint at where this could go but remain early and somewhat overhyped.
multi-agent-coordination
Media referenced
Noam Brown's essay on benchmark evaluation and test-time compute - article - Motivated by skepticism over GPT-5.5's benchmark gains; argues benchmarks must be plotted against a test-time compute budget (tokens, cost, or time) rather than reported as a single number.
Companies
OpenAI - Noam Brown's employer; source of the GPT-5.x model series and the internal model used to disprove the Erdos unit distance conjecture.
AISI - UK AI Safety Institute; its evaluations showed model performance on cyber tasks still improving after 100 million tokens of test-time compute.
Techniques and frameworks
Preparedness frameworks / responsible scaling policies - Lab safety protocols that gate model release on dangerous-capability evaluations; Brown argues they were designed before test-time compute scaling and do not specify what inference budget to evaluate at.
Benchmark grid reporting - The standard practice of publishing a single performance number per model per benchmark; Brown argues this hides the test-time compute budget used and should be replaced with performance plotted against an x-axis of tokens, cost, or time.
Recursive self-improvement (RSI) - The idea that AI systems accelerate the research that builds the next AI system; Brown frames current progress as a gradual, bottlenecked version of this rather than an overnight intelligence explosion.
Model routing / consensus layers - Vendor-layer products that route tasks to different models or run multiple models for consensus; Brown argues these should be benchmarked against the counterfactual of one model given the same total compute budget.
Summary
Sarah Guo talks with OpenAI research scientist Noam Brown, a pioneer of test-time compute scaling, about an essay he wrote arguing that the industry is evaluating frontier models the wrong way. The trigger was the release of GPT-5.5, which looked like only a modest improvement over GPT-5.4 on the standard published benchmark grid, drawing public skepticism that dissolved within hours once people actually used the model. Brown's diagnosis: benchmark grids report a single score per model without controlling for how much test-time compute (tokens, dollars, or time) went into producing it. Once you hold compute constant, 5.5 is dramatically better per unit of thinking time than 5.4, which better matches what users actually experienced.
The deeper problem, Brown argues, is that the point where a model's benchmark performance plateaus has moved so far out - sometimes weeks of scaffolded thinking - that waiting for saturation is no longer a practical evaluation strategy. This creates a genuine gap in AI safety policy: preparedness frameworks and responsible scaling policies were designed in the pre-test-time-compute era and measure "capability" without specifying an inference budget, even though a model's dangerous capability, like its useful capability, scales with how much money is thrown at it. He extends the same logic to release cycles generally: since fully evaluating what a model can do might require running it for as long as the task itself takes (a month, six months, a year), and labs ship new models every couple of months, nobody, including the labs, actually knows the true capability ceiling of any given model before it is superseded.
Brown illustrates the stakes with OpenAI's internal disproof of the Erdos unit distance conjecture, produced cheaply and mostly out of curiosity, which outside users later reproduced using public GPT-5.5 with the right iterative scaffolding. He notes the cost of reproducing such results drops roughly 10-100x with each model generation, fueling a "just wait for the next model" mentality that OpenAI deliberately discourages internally, preferring researchers focus on building more capable, cost-effective models rather than mining the current one for one-off wins. Using his own poker-solver PhD work as a running test case, he traces model progress from an earlier generation that would confidently give wrong answers ("gaslighting") to GPT-5.5 handling equivalent tasks nearly zero-shot, while noting that models still lack "research taste": they can optimize and accelerate existing work by 10-100x but cannot yet originate genuinely novel research directions.
On recursive self-improvement, Brown pushes back on the idea of an imminent, overnight intelligence explosion. Because unlocking a model's full capability requires large amounts of test-time compute, progress is rate-limited by wall-clock time even as labs accelerate internal research, which he expects to produce a gradual rather than instantaneous takeoff. He flags benchmark-maxing tricks (running a model multiple times and taking the best or judge-selected answer) as misleading for the same compute-accounting reason, and extends the critique to vendor routing and consensus layers: the real test of a routing product is whether it beats a single model given the same total compute budget, not whether it beats a single un-optimized run. He closes on multi-agent coordination as an underexplored frontier, comparing today's isolated, short-context AI agents unfavorably to human civilization's ability to accumulate and share knowledge across generations, and flags early multi-agent coordination experiments as an overhyped but real signal of where the field is headed.
Notable Quotes
"The problem is we're in a world now where the capability of the model is a function of how much money you put into it, basically." - Noam Brown
"Nobody actually knows what the ceiling of capabilities are for these models because nobody's actually run them for long enough to really tell." - Noam Brown
"I don't think we're headed to that world largely because of the fact that the models rely so much on large-scale test time compute in order to achieve their greatest intelligence." - Noam Brown
"You kind of end up in this bad equilibrium where everybody kind of knows that it's a bad equilibrium, but nobody wants to break out." - Noam Brown