Executive summary
OpenAI's GPT-6 Astra scores 62.7% on ARC-AGI-3 under the harness ARC Prize runs for every model, and 99.9% under a Provider Adapter that OpenAI supplied, per ARC Prize's Astra analysis. The 37-point gap comes from scaffolding that preserves reasoning state between requests and compacts long conversations. The weights are identical across both runs.
Strategic implications:
The number a vendor quotes is not the number an independent evaluator reports. ARC Prize's replay archive records the standard-harness runs it scored itself, and only the adapter configuration OpenAI supplied reaches the high 90s, per ARC Prize's replay archive.
Top scores cost real money and don't scale predictably. The standard-harness run cost $26,098. The higher-scoring adapter run cost $18,817, per ARC Prize's results page; cost and score are not simply correlated. OpenAI is not betting its narrative on this benchmark alone. The company separately told CNBC that Astra crossed its "Critical" cybersecurity threshold, finding unknown flaws without human step-by-step guidance, a claim independent of ARC-AGI-3 entirely.
A vendor citing an ARC-AGI-3 score should be asked which harness produced it, and whether that harness ships in the product being purchased. If the answer is the Provider Adapter, the number describes a research configuration.
For practitioners
ARC Prize published twelve configurations for this model, and the harness split is only the widest of them. Holding the harness fixed and moving reasoning effort alone takes the score from 35.18% to 62.71%, a 27-point swing before scaffolding enters the picture at all ARC Prize's GPT-6 Astra results page. An engineering team quoting "Astra's ARC-AGI-3 score" without naming both variables is quoting a config, not a model.
The harness is the code that manages state, retries, context assembly and tool-calling around the model. Crediting its work to the weights is the category error here, and it is the buyer who inherits it: the adapter number is a realistic ceiling only for an integration that already retains reasoning state and compacts context the same way. Most do not, which is why the question to put to a vendor is whether the harness that produced the number ships in the product being sold.
Price is the second discount. Astra lists at $10 per million input tokens and $50 per million output, against $4 and $20 for GPT-5.6 Sol (OpenAI's API pricing). A buyer is paying 2.5 times its predecessor's rate, and the headline justifying that rate is the one that moves 37 points on harness choice.
The engineering takeaway is procedural: before wiring an eval into a purchasing decision, the buyer has to ask which harness produced the number, whether reasoning effort was set to max, and what the run cost. ARC Prize publishes all three for every model on its results page.
Deep dive
The two harnesses are not measuring the same thing, which is what makes the gap interpretable rather than merely large. ARC Prize's Standard harness lets a model "carry forward notes it chooses to keep with it throughout the environment" — memory the model manages, in its own output channel, at its own cost. The Provider Adapter keeps the reasoning trace alive across requests without ever routing it through that channel (ARC Prize's results page). That is memory the platform manages on the model's behalf. Of the twelve configurations tested, the highest score came from the one least likely to resemble a typical enterprise integration.
ARC Prize says so itself, in the writeup announcing the result: saturating the benchmark "would not represent 'proof of achieving AGI'" (ARC Prize's Astra analysis). The organization that built the test, and that has the most to gain from its result mattering, is the one asking readers to discount it.
Reasoning effort alone produces a swing nearly as large as the harness choice. The state-retention mechanism appears to be doing most of the work; reasoning effort matters far less once that state is preserved. An enterprise evaluating Astra on a pilot workload has at least two free variables, harness and reasoning level, that can move the observed score by 20 to 40 points without touching the model itself.
The cost data complicates any easy story about scaffolding being simply a matter of spend. Whatever the Provider Adapter harness is doing, it is not brute-forcing the benchmark with additional compute. It is a different architecture for holding state, and architectures are the kind of thing an enterprise's own engineering team can attempt to replicate, or fail to, depending on what access the vendor grants outside a benchmark environment.
There is a live counterargument here worth taking seriously: if state retention and compaction are patterns enterprises can build themselves, the adapter score is not fraud, it is a preview. The honest position is that this has not happened yet, and ARC Prize's own framing treats the adapter score as the exception requiring explanation, not the baseline.
There is a precedent for how a headline ARC number ages, and it is two generations old. In December 2024, o3 scored 75.7% on ARC-AGI-1's semi-private set in its high-efficiency configuration, and 87.5% in a low-efficiency configuration costing $4,560 per task (ARC Prize's o3 breakthrough writeup). Those figures were read at the time as a step change in general reasoning. When ARC Prize replaced the test, o3-low scored 4% on ARC-AGI-2 (The Decoder's report on the ARC-AGI-2 results). ARC-AGI-2 was built to require symbol interpretation, multi-step compositional reasoning, and context-dependent rule application. Nothing about o3 changed between those two numbers. The test stopped rewarding what o3 had been good at.
That is the discount to apply to Astra's 99.9%, as reported in ARC Prize's Astra analysis. A score describes a model, a harness, and a test, and only the first of those three carries over to the next benchmark. Astra's ARC-AGI-3 result is evidence about ARC-AGI-3. Anyone extrapolating it to ARC-AGI-4, or to a workload that resembles neither, is repeating the reading that 75.7% got in 2024.
The benchmark itself is young enough that neither reading should be pushed too hard. ARC-AGI-3 is, per ARC Prize's 2026 competition page, "the first interactive reasoning benchmark," built around agents exploring game-like environments with no instructions, carrying an $850,000 total prize pool and a $700,000 grand prize reserved for the first agent to reach 100%. That is a research instrument for measuring exploration and world-modeling in novel settings, not a certification that any given customer support or coding workflow will behave predictably. OpenAI's own capability claims for Astra do not rest on this benchmark alone: the company told CNBC that Astra crossed the "Critical" threshold in its Preparedness Framework for cybersecurity, capable of finding and exploiting unknown flaws without step-by-step human guidance, a claim standing independent of any ARC-AGI-3 number. A model can be dangerous at one task and unreliable at reproducing its own headline score at another.