ARC Prize puts Astra 37 points below the score OpenAI led with

The two figures describe the same model on the same benchmark run through different plumbing. ARC Prize's Standard harness hands every model an identical minimal, provider-neutral interface and makes it decide for itself what to keep in visible notes ^[13]. The Provider Adapter preserves OpenAI's opaque reasoning state between requests and compacts longer conversations, and across the game-and-reasoning pairs both setups solved, adapter runs used 49% fewer tokens and finished about 3.66 times faster ^[14][15]. The distance between the two results is a description of the scaffold as much as of the model.

The cost line inverts the usual assumption that the better score is the expensive one. Astra at maximum effort scored 62.7% on the semi-private set for $26,098, while the high-effort Provider Adapter run that produced 99.9% cost $18,817 ^[16], so the lower score cost about 39% more ^[25]. Persistent reasoning state made the model cheaper and stronger at the same time, an engineering result rather than a sleight of hand. The marketing is in the comparison, not the adapter.

The adapter is how paying customers will actually call the model, so the neutral figure understates what a buyer receives. ARC Prize agrees on direction, calling Astra "a noticeable step-function change in frontier model capabilities" and "a major milestone worth celebrating" ^[17]. The same evaluation says saturating ARC-AGI-3 would not prove AGI, and that the benchmark's mechanics are deterministic and closed-ended and do not represent the complexity of the real world ^[18]. That gap between a capability signal and a procurement input is exactly what a buyer has to bridge on their own.

Pricing is where the buyer's exposure gets concrete. Sol Standard runs $5 per million input tokens and $30 per million output ^[11], so a task consuming a million tokens each way costs $35 on Sol against $60 on Astra, a rise of about 71% ^[27], and Astra must finish the same work with roughly 42% fewer tokens to reach parity ^[26]. OpenAI says the model can do exactly that by spending fewer tokens per task ^[19], a claim only a buyer's own traffic can settle. It is worth crediting that OpenAI published a losing number too: its own table puts Astra at 57.2% on Humanity's Last Exam with tools against Sol's 65.0% ^[10].

The absent benchmark is the one that matches the claim being made. OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" ^[2], and GDPval is its own benchmark for economically valuable work ^[12]. Days before the launch, Altman told the Sources podcast that AGI is "at best a very poorly defined term" and "like an irrelevant marketing term" ^[3]. The Guardian notes the company is pushing toward a stock market listing it hopes will value it above $850bn ^[6], which is the visible incentive behind a launch-day label.

What a buyer decides this quarter is narrow, because the rollout is narrow. Astra went first to a limited set of organizations including OpenAI's Daybreak program, with ChatGPT tiers, the API and Amazon Bedrock following over days, and enterprise access stays off until an administrator turns it on ^[20][21]. Because the model carries OpenAI's critical cybersecurity classification, less restrictive access sits with an initial set of trusted defenders ^[4], and chief scientist Jakub Pachocki said confidence in monitoring may constrain further development ^[5]. So the trade-off is legible: a buyer either pays about 71% more per equal-token task now for a gain the vendor's headline number cannot evidence, or waits a quarter, measures token consumption on its own workloads, and buys the same model on numbers it owns.